git.iko.soy/grenade
on master help sign in
swap
afb9f9b123 ahead, 0 behind
.gitignore+3−1
since splitsnapshots
unifiedsplit
@@ -1,4 +1,6 @@
1+cabal.project.local
1.cabal-sandbox/
2.cabal-sandbox/
2cabal.sandbox.config
3cabal.sandbox.config
3dist/
4dist/
4−
5+dist-newstyle/
6+.stack-work/
.gitmodulesdeleted−15
@@ -1,15 +0,0 @@
1−[submodule "lib/disorder"]
2− path = lib/disorder
3− url = https://github.com/ambiata/disorder.hs
4−[submodule "lib/x"]
5− path = lib/x
6− url = https://github.com/ambiata/x.git
7−[submodule "lib/p"]
8− path = lib/p
9− url = https://github.com/ambiata/p
10−[submodule "lib/disorder.hs"]
11− path = lib/disorder.hs
12− url = https://github.com/ambiata/disorder.hs.git
13−[submodule "lib/hmatrix"]
14− path = lib/hmatrix
15− url = https://github.com/albertoruiz/hmatrix.git
.travis.ymlnew file+31
@@ -0,0 +1,31 @@
1+# NB: don't set `language: haskell` here
2+
3+# The following enables several GHC versions to be tested; often it's enough to test only against the last release in a major GHC version. Feel free to omit lines listings versions you don't need/want testing for.
4+env:
5+ - GHCVER=8.0.2
6+ - GHCVER=8.2.2
7+ - GHCVER=8.4.4
8+ - GHCVER=8.6.5
9+ - GHCVER=8.8.3
10+ - GHCVER=8.10.1
11+
12+# Note: the distinction between `before_install` and `install` is not important.
13+before_install:
14+ - travis_retry sudo add-apt-repository -y ppa:hvr/ghc
15+ - travis_retry sudo apt-get update
16+ - travis_retry sudo apt-get install cabal-install-3.2 ghc-$GHCVER libblas-dev liblapack-dev
17+ - export PATH=/opt/cabal/bin:/opt/ghc/$GHCVER/bin:$PATH
18+
19+install:
20+ - echo "$(ghc --version) [$(ghc --print-project-git-commit-id 2> /dev/null || echo '?')]"
21+ - travis_retry cabal-3.2 update
22+
23+notifications:
24+ email: false
25+
26+# Here starts the actual work to be performed for the package under test; any command which exits with a non-zero exit code causes the build to fail.
27+script:
28+ - cabal-3.2 configure --enable-tests
29+ - cabal-3.2 build all
30+ - cabal-3.2 test --test-show-details=direct -j1
31+
LICENSE+1−1
@@ -1,4 +1,4 @@
1−Copyright (c) 2016, Huw Campbell
1+Copyright (c) 2016-2017, Huw Campbell
2All rights reserved.
2All rights reserved.
3
3
4Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
4Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
README.md+72−10
@@ -1,6 +1,11 @@
1Grenade
1Grenade
2=======
2=======
3
3
4+[![Build Status](https://api.travis-ci.org/HuwCampbell/grenade.svg?branch=master)](https://travis-ci.org/HuwCampbell/grenade)
5+[![Hackage page (downloads and API reference)][hackage-png]][hackage]
6+[![Hackage-Deps][hackage-deps-png]][hackage-deps]
7+
8+
4```
9```
5First shalt thou take out the Holy Pin, then shalt thou count to three, no more, no less.
10First shalt thou take out the Holy Pin, then shalt thou count to three, no more, no less.
6Three shall be the number thou shalt count, and the number of the counting shall be three.
11Three shall be the number thou shalt count, and the number of the counting shall be three.
@@ -10,14 +15,20 @@ Five is right out.
10
15
11💣 Machine learning which might blow up in your face 💣
16💣 Machine learning which might blow up in your face 💣
12
17
13−Grenade is a dependently typed, practical, and fast recurrent neural network library
18+Grenade is a composable, dependently typed, practical, and fast recurrent neural network library
14for concise and precise specifications of complex networks in Haskell.
19for concise and precise specifications of complex networks in Haskell.
15
20
16As an example, a network which can achieve ~1.5% error on MNIST can be
21As an example, a network which can achieve ~1.5% error on MNIST can be
17specified and initialised with random weights in a few lines of code with
22specified and initialised with random weights in a few lines of code with
18```haskell
23```haskell
24+type MNIST
25+ = Network
26+ '[ Convolution 1 10 5 5 1 1, Pooling 2 2 2 2, Relu
27+ , Convolution 10 16 5 5 1 1, Pooling 2 2 2 2, Reshape, Relu
19−type MNIST = Network '[ Convolution 1 10 5 5 1 1, Pooling 2 2 2 2, Relu, Convolution 10 16 5 5 1 1, Pooling 2 2 2 2, FlattenLayer, Relu, FullyConnected 256 80, Logit, FullyConnected 80 10, Logit]
28+ , FullyConnected 256 80, Logit, FullyConnected 80 10, Logit]
20− '[ 'D2 28 28, 'D3 24 24 10, 'D3 12 12 10, 'D3 12 12 10, 'D3 8 8 16, 'D3 4 4 16, 'D1 256, 'D1 256, 'D1 80, 'D1 80, 'D1 10, 'D1 10]
29+ '[ 'D2 28 28, 'D3 24 24 10, 'D3 12 12 10, 'D3 12 12 10
30+ , 'D3 8 8 16, 'D3 4 4 16, 'D1 256, 'D1 256
31+ , 'D1 80, 'D1 80, 'D1 10, 'D1 10]
21
32
22randomMnist :: MonadRandom m => m MNIST
33randomMnist :: MonadRandom m => m MNIST
23randomMnist = randomNetwork
34randomMnist = randomNetwork
@@ -31,8 +42,10 @@ If recurrent neural networks are more your style, you can try defining something
31["unreasonably effective"](http://karpathy.github.io/2015/05/21/rnn-effectiveness/)
42["unreasonably effective"](http://karpathy.github.io/2015/05/21/rnn-effectiveness/)
32with
43with
33```haskell
44```haskell
45+type Shakespeare
46+ = RecurrentNetwork
34−type Shakespeare = RecurrentNetwork '[ R (LSTM 40 80), R (LSTM 80 40), F (FullyConnected 40 40), F Logit]
47+ '[ R (LSTM 40 80), R (LSTM 80 40), F (FullyConnected 40 40), F Logit]
35− '[ 'D1 40, 'D1 80, 'D1 40, 'D1 40, 'D1 40 ]
48+ '[ 'D1 40, 'D1 80, 'D1 40, 'D1 40, 'D1 40 ]
36```
49```
37
50
38Design
51Design
@@ -45,8 +58,13 @@ data that are passed between the layers.
45The definition of a network is surprisingly simple:
58The definition of a network is surprisingly simple:
46```haskell
59```haskell
47data Network :: [*] -> [Shape] -> * where
60data Network :: [*] -> [Shape] -> * where
61+ NNil :: SingI i
48− O :: (SingI i, SingI o, Layer x i o) => !x -> Network '[x] '[i, o]
62+ => Network '[] '[i]
63+
49− (:~>) :: (SingI i, SingI h, Layer x i h) => !x -> !(Network xs (h ': hs)) -> Network (x ': xs) (i ': h ': hs)
64+ (:~>) :: (SingI i, SingI h, Layer x i h)
65+ => !x
66+ -> !(Network xs (h ': hs))
67+ -> Network (x ': xs) (i ': h ': hs)
50```
68```
51
69
52The `Layer x i o` constraint ensures that the layer `x` can sensibly perform a
70The `Layer x i o` constraint ensures that the layer `x` can sensibly perform a
@@ -71,7 +89,7 @@ backPropagate :: forall shapes layers.
71```
89```
72which takes a network, appropriate input and target data, and returns the
90which takes a network, appropriate input and target data, and returns the
73back propagated gradients for the network. The shapes of the gradients are
91back propagated gradients for the network. The shapes of the gradients are
74−appropriate for each layer, and may be trivial for layers like `Rulu` which
92+appropriate for each layer, and may be trivial for layers like `Relu` which
75have no learnable parameters.
93have no learnable parameters.
76
94
77The gradients however can always be applied, yielding a new (hopefully better)
95The gradients however can always be applied, yielding a new (hopefully better)
@@ -85,9 +103,48 @@ easy in downstream code. If the shapes of a network are not specified correctly
85and a layer can not sensibly perform the operation between two shapes, then
103and a layer can not sensibly perform the operation between two shapes, then
86it will result in a compile time error.
104it will result in a compile time error.
87
105
106+Composition
107+-----------
108+
109+Networks and Layers in Grenade are easily composed at the type level. As a `Network`
110+is an instance of `Layer`, one can use a trained Network as a small component in a
111+larger network easily. Furthermore, we provide 2 layers which are designed to run
112+layers in parallel and merge their output (either by concatenating them across one
113+dimension or summing by pointwise adding their activations). This allows one to
114+write any Network which can be expressed as a
115+[series parallel graph](https://en.wikipedia.org/wiki/Series-parallel_graph).
116+
117+A residual network layer specification for instance could be written as
118+```haskell
119+type Residual net = Merge Trivial net
120+```
121+If the type `net` is an instance of `Layer`, then `Residual net` will be too. It will
122+run the network, while retaining its input by passing it through the `Trivial` layer,
123+and merge the original image with the output.
124+
125+See the [MNIST](https://github.com/HuwCampbell/grenade/blob/master/examples/main/mnist.hs)
126+example, which has been overengineered to contain both residual style learning as well
127+as inception style convolutions.
128+
129+Generative Adversarial Networks
130+-------------------------------
131+
132+As Grenade is purely functional, one can compose its training functions in flexible
133+ways. [GAN-MNIST](https://github.com/HuwCampbell/grenade/blob/master/examples/main/gan-mnist.hs)
134+example displays an interesting, type safe way of writing a generative adversarial
135+training function in 10 lines of code.
136+
137+Layer Zoo
138+---------
139+
140+Grenade layers are normal haskell data types which are an instance of `Layer`, so
141+it's easy to build one's own downstream code. We do however provide a decent set
142+of layers, including convolution, deconvolution, pooling, pad, crop, logit, relu,
143+elu, tanh, and fully connected.
144+
88Build Instructions
145Build Instructions
89------------------
146------------------
90−Grenade currently only builds with the [mafia](https://github.com/ambiata/mafia)
147+Grenade is most easily built with the [mafia](https://github.com/ambiata/mafia)
91script that is located in the repository. You will also need the `lapack` and
148script that is located in the repository. You will also need the `lapack` and
92`blas` libraries and development tools. Once you have all that, Grenade can be
149`blas` libraries and development tools. Once you have all that, Grenade can be
93build using:
150build using:
@@ -102,7 +159,7 @@ and the tests run using:
102./mafia test
159./mafia test
103```
160```
104
161
105−Grenade is currently known to build with ghc 7.10 and 8.0.
162+Grenade builds with ghc 7.10, 8.0, 8.2 and 8.4.
106
163
107Thanks
164Thanks
108------
165------
@@ -128,3 +185,8 @@ core took around 12 minutes, achieving 1.5% error rate on a 1000 sample holdout
128Contributing
185Contributing
129------------
186------------
130Contributions are welcome.
187Contributions are welcome.
188+
189+ [hackage]: http://hackage.haskell.org/package/grenade
190+ [hackage-png]: http://img.shields.io/hackage/v/grenade.svg
191+ [hackage-deps]: http://packdeps.haskellers.com/reverse/grenade
192+ [hackage-deps-png]: https://img.shields.io/hackage-deps/v/grenade.svg
_config.ymlnew file+1
@@ -0,0 +1 @@
1+theme: jekyll-theme-minimal
bench/bench-lstm.hs+6−48
@@ -5,68 +5,26 @@ import Criterion.Main
5
5
6import Grenade
6import Grenade
7import Grenade.Recurrent
7import Grenade.Recurrent
8−import Grenade.Layers.Internal.Update
9−
10−import qualified Numeric.LinearAlgebra.Static as H
11
8
12main :: IO ()
9main :: IO ()
13main = do
10main = do
14− layer60 :: LSTM 40 60 <- createRandom
15− layer512 :: LSTM 40 512 <- createRandom
16 input40 :: S ('D1 40) <- randomOfShape
11 input40 :: S ('D1 40) <- randomOfShape
17− rec60 :: S ('D1 60) <- randomOfShape
18− rec512 :: S ('D1 512) <- randomOfShape
19− (lstm, lstm') :: (RecNet, RecInput) <- randomRecurrent
12+ lstm :: RecNet <- randomRecurrent
20−
21− let upIn60 :: H.R 3600 = H.randomVector 1 H.Uniform * 2 - 1
22− let upIn512 :: H.R 262144 = H.randomVector 1 H.Uniform * 2 - 1
23
13
24 defaultMain [
14 defaultMain [
25− bgroup "lstm" [ bench "forwards-60" $ nf (nfT2 . uncurry (testRun60 layer60)) (rec60, input40)
26− , bench "forwards-512" $ nf (nfT2 . uncurry (testRun512 layer512)) (rec512, input40)
27− , bench "backwards-60" $ nf (nfT3 . uncurry4 (testRun60' layer60)) (rec60, input40, rec60, rec60)
28− , bench "backwards-512" $ nf (nfT3 . uncurry4 (testRun512' layer512)) (rec512, input40, rec512, rec512)
29− ]
30− , bgroup "update" [ bench "matrix-60x60" $ nf (uncurry3 (decendVector 1 1 1)) (upIn60, upIn60, upIn60)
31− , bench "matrix-512x512" $ nf (uncurry3 (decendVector 1 1 1)) (upIn512, upIn512, upIn512)
32− ]
33− , bgroup "train" [ bench "one-time-step" $ whnf (nfT2 . trainRecurrent lp lstm lstm') [(input40, Just input40)]
15+ bgroup "train" [ bench "one-time-step" $ whnf (nfT2 . trainRecurrent lp lstm 0) [(input40, Just input40)]
34− , bench "ten-time-steps" $ whnf (nfT2 . trainRecurrent lp lstm lstm') $ replicate 10 (input40, Just input40)
16+ , bench "ten-time-steps" $ whnf (nfT2 . trainRecurrent lp lstm 0) $ replicate 10 (input40, Just input40)
35− , bench "fifty-time-steps" $ whnf (nfT2 . trainRecurrent lp lstm lstm') $ replicate 50 (input40, Just input40)
17+ , bench "fifty-time-steps" $ whnf (nfT2 . trainRecurrent lp lstm 0) $ replicate 50 (input40, Just input40)
36 ]
18 ]
37 ]
19 ]
38
20
39−testRun60 :: LSTM 40 60 -> S ('D1 60) -> S ('D1 40) -> (S ('D1 60), S ('D1 60))
40−testRun60 = runRecurrentForwards
41−
42−testRun60' :: LSTM 40 60 -> S ('D1 60) -> S ('D1 40) -> S ('D1 60) -> S ('D1 60) -> (Gradient (LSTM 40 60), S ('D1 60), S ('D1 40))
43−testRun60' = runRecurrentBackwards
44−
45−testRun512 :: LSTM 40 512 -> S ('D1 512) -> S ('D1 40) -> (S ('D1 512), S ('D1 512))
46−testRun512 = runRecurrentForwards
47−
48−testRun512' :: LSTM 40 512 -> S ('D1 512) -> S ('D1 40) -> S ('D1 512) -> S ('D1 512) -> (Gradient (LSTM 40 512), S ('D1 512), S ('D1 40))
49−testRun512' = runRecurrentBackwards
50−
51−uncurry4 :: (t -> t1 -> t2 -> t3 -> t4) -> (t, t1, t2, t3) -> t4
52−uncurry4 f (a,b,c,d) = f a b c d
53−
54−uncurry3 :: (t -> t1 -> t2 -> t3) -> (t, t1, t2) -> t3
55−uncurry3 f (a,b,c) = f a b c
56−
57nfT2 :: (a, b) -> (a, b)
21nfT2 :: (a, b) -> (a, b)
58nfT2 (!a, !b) = (a, b)
22nfT2 (!a, !b) = (a, b)
59
23
60−nfT3 :: (a, b, c) -> (b, c)
61−nfT3 (!_, !b, !c) = (b, c)
62−
63
24
64−type F = FeedForward
65type R = Recurrent
25type R = Recurrent
66−type RecNet = RecurrentNetwork '[ R (LSTM 40 512), R (LSTM 512 40), F (FullyConnected 40 40), F Logit]
26+type RecNet = RecurrentNetwork '[ R (LSTM 40 512), R (LSTM 512 40) ]
67− '[ 'D1 40, 'D1 512, 'D1 40, 'D1 40, 'D1 40 ]
27+ '[ 'D1 40, 'D1 512, 'D1 40 ]
68−
69−type RecInput = RecurrentInputs '[ R (LSTM 40 512), R (LSTM 512 40), F (FullyConnected 40 40), F Logit]
70
28
71lp :: LearningParameters
29lp :: LearningParameters
72lp = LearningParameters 0.1 0 0
30lp = LearningParameters 0.1 0 0
bench/bench.hs+47−18
@@ -1,26 +1,55 @@
1+{-# LANGUAGE DataKinds #-}
2+{-# LANGUAGE ScopedTypeVariables #-}
1import Criterion.Main
3import Criterion.Main
2
4
5+import Grenade
6+
3import Grenade.Layers.Internal.Convolution
7import Grenade.Layers.Internal.Convolution
4import Grenade.Layers.Internal.Pooling
8import Grenade.Layers.Internal.Pooling
5
9
6import Numeric.LinearAlgebra
10import Numeric.LinearAlgebra
7
11
8main :: IO ()
12main :: IO ()
13+main = do
14+ x :: S ('D2 60 60 ) <- randomOfShape
15+ y :: S ('D3 60 60 1) <- randomOfShape
16+
9−main = defaultMain [
17+ defaultMain [
10− bgroup "im2col" [ bench "im2col 3x4" $ whnf (im2col 2 2 1 1) ((3><4) [1..])
18+ bgroup "im2col" [ bench "im2col 3x4" $ whnf (im2col 2 2 1 1) ((3><4) [1..])
11− , bench "im2col 28x28" $ whnf (im2col 5 5 1 1) ((28><28) [1..])
19+ , bench "im2col 28x28" $ whnf (im2col 5 5 1 1) ((28><28) [1..])
12− , bench "im2col 100x100" $ whnf (im2col 10 10 1 1) ((100><100) [1..])
20+ , bench "im2col 100x100" $ whnf (im2col 10 10 1 1) ((100><100) [1..])
13− ]
21+ ]
14− , bgroup "col2im" [ bench "col2im 3x4" $ whnf (col2im 2 2 1 1 3 4) ((6><4) [1..])
22+ , bgroup "col2im" [ bench "col2im 3x4" $ whnf (col2im 2 2 1 1 3 4) ((6><4) [1..])
15− , bench "col2im 28x28" $ whnf (col2im 5 5 1 1 28 28) ((576><25) [1..])
23+ , bench "col2im 28x28" $ whnf (col2im 5 5 1 1 28 28) ((576><25) [1..])
16− , bench "col2im 100x100" $ whnf (col2im 10 10 1 1 100 100) ((8281><100) [1..])
24+ , bench "col2im 100x100" $ whnf (col2im 10 10 1 1 100 100) ((8281><100) [1..])
17− ]
25+ ]
18− , bgroup "poolfw" [ bench "poolforwards 3x4" $ whnf (poolForward 1 3 4 2 2 1 1) ((3><4) [1..])
26+ , bgroup "poolfw" [ bench "poolforwards 3x4" $ whnf (poolForward 1 3 4 2 2 1 1) ((3><4) [1..])
19− , bench "poolforwards 28x28" $ whnf (poolForward 1 28 28 5 5 1 1) ((28><28) [1..])
27+ , bench "poolforwards 28x28" $ whnf (poolForward 1 28 28 5 5 1 1) ((28><28) [1..])
20− , bench "poolforwards 100x100" $ whnf (poolForward 1 100 100 10 10 1 1) ((100><100) [1..])
28+ , bench "poolforwards 100x100" $ whnf (poolForward 1 100 100 10 10 1 1) ((100><100) [1..])
21− ]
29+ ]
22− , bgroup "poolbw" [ bench "poolbackwards 3x4" $ whnf (poolBackward 1 3 4 2 2 1 1 ((3><4) [1..])) ((2><3) [1..])
30+ , bgroup "poolbw" [ bench "poolbackwards 3x4" $ whnf (poolBackward 1 3 4 2 2 1 1 ((3><4) [1..])) ((2><3) [1..])
23− , bench "poolbackwards 28x28" $ whnf (poolBackward 1 28 28 5 5 1 1 ((28><28) [1..])) ((24><24) [1..])
31+ , bench "poolbackwards 28x28" $ whnf (poolBackward 1 28 28 5 5 1 1 ((28><28) [1..])) ((24><24) [1..])
24− , bench "poolbackwards 100x100" $ whnf (poolBackward 1 100 100 10 10 1 1 ((100><100) [1..])) ((91><91) [1..])
32+ , bench "poolbackwards 100x100" $ whnf (poolBackward 1 100 100 10 10 1 1 ((100><100) [1..])) ((91><91) [1..])
33+ ]
34+ , bgroup "padcrop" [ bench "pad 2D 60x60" $ whnf (testRun2D Pad) x
35+ , bench "pad 3D 60x60" $ whnf (testRun3D Pad) y
36+ , bench "crop 2D 60x60" $ whnf (testRun2D' Crop) x
37+ , bench "crop 3D 60x60" $ whnf (testRun3D' Crop) y
25− ]
38+ ]
26− ]
39+ ]
40+
41+
42+testRun2D :: Pad 1 1 1 1 -> S ('D2 60 60) -> S ('D2 62 62)
43+testRun2D = snd ... runForwards
44+
45+testRun3D :: Pad 1 1 1 1 -> S ('D3 60 60 1) -> S ('D3 62 62 1)
46+testRun3D = snd ... runForwards
47+
48+testRun2D' :: Crop 1 1 1 1 -> S ('D2 60 60) -> S ('D2 58 58)
49+testRun2D' = snd ... runForwards
50+
51+testRun3D' :: Crop 1 1 1 1 -> S ('D3 60 60 1) -> S ('D3 58 58 1)
52+testRun3D' = snd ... runForwards
53+
54+(...) :: (a -> b) -> (c -> d -> a) -> c -> d -> b
55+(...) = (.) . (.)
cbits/gradient_decent.c → cbits/gradient_descent.crenamed+3−3
@@ -1,12 +1,12 @@
1−#include "gradient_decent.h"
1+#include "gradient_descent.h"
2
2
3−void decend_cpu(int len, double rate, double momentum, double regulariser,
3+void descend_cpu(int len, double rate, double momentum, double regulariser,
4 const double* weights,
4 const double* weights,
5 const double* gradient,
5 const double* gradient,
6 const double* last,
6 const double* last,
7 double* outputWeights, double* outputMomentum) {
7 double* outputWeights, double* outputMomentum) {
8
8
9− for (int i = 0; i <= len; i++) {
9+ for (int i = 0; i < len; i++) {
10 outputMomentum[i] = momentum * last[i] - rate * gradient[i];
10 outputMomentum[i] = momentum * last[i] - rate * gradient[i];
11 outputWeights[i] = weights[i] + outputMomentum[i] - (rate * regulariser) * weights[i];
11 outputWeights[i] = weights[i] + outputMomentum[i] - (rate * regulariser) * weights[i];
12 }
12 }
cbits/gradient_decent.h → cbits/gradient_descent.hrenamed+1−1
@@ -1,7 +1,7 @@
1#include <stdio.h>
1#include <stdio.h>
2#include <stdint.h>
2#include <stdint.h>
3
3
4−void decend_cpu(int len, double rate, double momentum, double regulariser,
4+void descend_cpu(int len, double rate, double momentum, double regulariser,
5 const double* weights,
5 const double* weights,
6 const double* gradient,
6 const double* gradient,
7 const double* last,
7 const double* last,
cbits/pad.cnew file+39
@@ -0,0 +1,39 @@
1+#include "pad.h"
2+
3+void pad_cpu(double* data, const int channels,
4+ const int height, const int width, const int pad_left, const int pad_top,
5+ const int pad_right, const int pad_bottom,
6+ double* data_padded) {
7+
8+ const int pad_width = width + pad_left + pad_right;
9+ const int pad_height = height + pad_top + pad_bottom;
10+
11+ memset(data_padded, 0, pad_height * pad_width * channels * sizeof(double));
12+
13+ for (int channel = 0; channel < channels; channel++) {
14+ double* px = data_padded + (pad_width * pad_top + pad_left) + channel * (pad_width * pad_height);
15+ for (int y = 0; y < height; y++) {
16+ memcpy(px, data, sizeof(double) * width);
17+ px += pad_width;
18+ data += width;
19+ }
20+ }
21+}
22+
23+void crop_cpu(double* data, const int channels,
24+ const int height, const int width, const int crop_left, const int crop_top,
25+ const int crop_right, const int crop_bottom,
26+ double* data_cropped) {
27+
28+ const int crop_width = width + crop_left + crop_right;
29+ const int crop_height = height + crop_top + crop_bottom;
30+
31+ for (int channel = 0; channel < channels; channel++) {
32+ double* px = data + (crop_width * crop_top + crop_left) + channel * (crop_width * crop_height);
33+ for (int y = 0; y < height; y++) {
34+ memcpy(data_cropped, px, sizeof(double) * width);
35+ px += crop_width;
36+ data_cropped += width;
37+ }
38+ }
39+}
cbits/pad.hnew file+13
@@ -0,0 +1,13 @@
1+#include <stdio.h>
2+#include <stdint.h>
3+#include <string.h>
4+
5+void pad_cpu(double* data_im, const int channels,
6+ const int height, const int width, const int pad_left, const int pad_top,
7+ const int pad_right, const int pad_bottom,
8+ double* data_col);
9+
10+void crop_cpu(double* data_im, const int channels,
11+ const int height, const int width, const int crop_left, const int crop_top,
12+ const int crop_right, const int crop_bottom,
13+ double* data_col);
examples/LICENSEnew file+10
@@ -0,0 +1,10 @@
1+Copyright (c) 2016-2017, Huw Campbell
2+All rights reserved.
3+
4+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
5+
6+1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
7+
8+2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
9+
10+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
examples/grenade-examples.cabalnew file+109
@@ -0,0 +1,109 @@
1+name: grenade-examples
2+version: 0.0.1
3+license: BSD2
4+license-file: LICENSE
5+author: Huw Campbell <huw.campbell@gmail.com>
6+maintainer: Huw Campbell <huw.campbell@gmail.com>
7+copyright: (c) 2016-2017 Huw Campbell.
8+synopsis: grenade-examples
9+category: System
10+cabal-version: >= 1.8
11+build-type: Simple
12+description: grenade-examples
13+
14+source-repository head
15+ type: git
16+ location: https://github.com/HuwCampbell/grenade.git
17+
18+library
19+
20+executable feedforward
21+ ghc-options: -Wall -threaded -O2
22+ main-is: main/feedforward.hs
23+ build-depends: base
24+ , grenade
25+ , attoparsec
26+ , bytestring
27+ , cereal
28+ , either
29+ , optparse-applicative >= 0.13 && < 0.18
30+ , text == 1.2.*
31+ , mtl >= 2.2.1 && < 2.3
32+ , hmatrix
33+ , transformers
34+ , singletons
35+ , semigroups
36+ , MonadRandom
37+
38+executable mnist
39+ ghc-options: -Wall -threaded -O2
40+ main-is: main/mnist.hs
41+ build-depends: base
42+ , grenade
43+ , attoparsec
44+ , either
45+ , optparse-applicative >= 0.13 && < 0.18
46+ , text == 1.2.*
47+ , mtl >= 2.2.1 && < 2.3
48+ , hmatrix >= 0.18 && < 0.21
49+ , transformers
50+ , semigroups
51+ , singletons
52+ , MonadRandom
53+ , vector
54+
55+executable gan-mnist
56+ ghc-options: -Wall -threaded -O2
57+ main-is: main/gan-mnist.hs
58+ build-depends: base
59+ , grenade
60+ , attoparsec
61+ , bytestring
62+ , cereal
63+ , either
64+ , optparse-applicative >= 0.13 && < 0.18
65+ , text == 1.2.*
66+ , mtl >= 2.2.1 && < 2.3
67+ , hmatrix >= 0.18 && < 0.21
68+ , transformers
69+ , semigroups
70+ , singletons
71+ , MonadRandom
72+ , vector
73+
74+executable recurrent
75+ ghc-options: -Wall -threaded -O2
76+ main-is: main/recurrent.hs
77+ build-depends: base
78+ , grenade
79+ , attoparsec
80+ , either
81+ , optparse-applicative >= 0.13 && < 0.18
82+ , text == 1.2.*
83+ , mtl >= 2.2.1 && < 2.3
84+ , hmatrix >= 0.18 && < 0.21
85+ , transformers
86+ , semigroups
87+ , singletons
88+ , MonadRandom
89+
90+executable shakespeare
91+ ghc-options: -Wall -threaded -O2
92+ main-is: main/shakespeare.hs
93+ build-depends: base
94+ , grenade
95+ , attoparsec
96+ , bytestring
97+ , cereal
98+ , either
99+ , optparse-applicative >= 0.13 && < 0.18
100+ , text == 1.2.*
101+ , mtl >= 2.2.1 && < 2.3
102+ , hmatrix >= 0.18 && < 0.21
103+ , transformers
104+ , semigroups
105+ , singletons
106+ , singletons-base
107+ , vector
108+ , MonadRandom
109+ , containers
examples/mafianew file+62
@@ -0,0 +1,62 @@
1+#!/bin/sh -eu
2+
3+: ${MAFIA_HOME:=$HOME/.mafia}
4+: ${MAFIA_VERSIONS:=$MAFIA_HOME/versions}
5+
6+latest_version () {
7+ git ls-remote https://github.com/haskell-mafia/mafia | grep refs/heads/master | cut -f 1
8+}
9+
10+build_version() {
11+ MAFIA_VERSION="$1"
12+ MAFIA_TEMP=$(mktemp -d 2>/dev/null || mktemp -d -t 'exec_mafia')
13+ MAFIA_FILE=mafia-$MAFIA_VERSION
14+ MAFIA_PATH=$MAFIA_VERSIONS/$MAFIA_FILE
15+ mkdir -p $MAFIA_VERSIONS
16+ echo "Building $MAFIA_FILE in $MAFIA_TEMP"
17+ git clone https://github.com/haskell-mafia/mafia $MAFIA_TEMP
18+ git --git-dir="$MAFIA_TEMP/.git" --work-tree="$MAFIA_TEMP" reset --hard $MAFIA_VERSION || {
19+ echo "mafia version ($MAFIA_VERSION) could not be found." >&2
20+ exit 1
21+ }
22+ (cd "$MAFIA_TEMP" && ./bin/bootstrap) || {
23+ got=$?
24+ echo "mafia version ($MAFIA_VERSION) could not be built." >&2
25+ exit "$got"
26+ }
27+ chmod +x "$MAFIA_TEMP/.cabal-sandbox/bin/mafia"
28+ # Ensure executable is on same file-system so final mv is atomic.
29+ mv -f "$MAFIA_TEMP/.cabal-sandbox/bin/mafia" "$MAFIA_PATH.$$"
30+ mv "$MAFIA_PATH.$$" "$MAFIA_PATH" || {
31+ rm -f "$MAFIA_PATH.$$"
32+ echo "INFO: mafia version ($MAFIA_VERSION) already exists not overiding," >&2
33+ echo "INFO: this is expected if parallel builds of the same version of" >&2
34+ echo "INFO: mafia occur, we are playing by first in, wins." >&2
35+ exit 0
36+ }
37+}
38+
39+enable_version() {
40+ if [ $# -eq 0 ]; then
41+ MAFIA_VERSION="$(latest_version)"
42+ echo "INFO: No explicit mafia version requested installing latest ($MAFIA_VERSION)." >&2
43+ else
44+ MAFIA_VERSION="$1"
45+ fi
46+ [ -x "$MAFIA_HOME/versions/mafia-$MAFIA_VERSION" ] || build_version "$MAFIA_VERSION"
47+ ln -sf "$MAFIA_HOME/versions/mafia-$MAFIA_VERSION" "$MAFIA_HOME/versions/mafia"
48+}
49+
50+exec_mafia () {
51+ [ -x "$MAFIA_HOME/versions/mafia" ] || enable_version
52+ "$MAFIA_HOME/versions/mafia" "$@"
53+}
54+
55+#
56+# The actual start of the script.....
57+#
58+
59+case "${1:-}" in
60+upgrade) shift; enable_version "$@" ;;
61+*) exec_mafia "$@"
62+esac
main/feedforward.hs → examples/main/feedforward.hsrenamed+4−1
@@ -1,4 +1,5 @@
1{-# LANGUAGE BangPatterns #-}
1{-# LANGUAGE BangPatterns #-}
2+{-# LANGUAGE CPP #-}
2{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE ScopedTypeVariables #-}
4{-# LANGUAGE ScopedTypeVariables #-}
4{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeOperators #-}
@@ -10,7 +11,9 @@ import Data.List ( foldl' )
10
11
11import qualified Data.ByteString as B
12import qualified Data.ByteString as B
12import Data.Serialize
13import Data.Serialize
13−
14+#if ! MIN_VERSION_base(4,13,0)
15+import Data.Semigroup ( (<>) )
16+#endif
14import GHC.TypeLits
17import GHC.TypeLits
15
18
16import qualified Numeric.LinearAlgebra.Static as SA
19import qualified Numeric.LinearAlgebra.Static as SA
examples/main/gan-mnist.hsnew file+185
@@ -0,0 +1,185 @@
1+{-# LANGUAGE BangPatterns #-}
2+{-# LANGUAGE CPP #-}
3+{-# LANGUAGE DataKinds #-}
4+{-# LANGUAGE GADTs #-}
5+{-# LANGUAGE ScopedTypeVariables #-}
6+{-# LANGUAGE TypeOperators #-}
7+{-# LANGUAGE TupleSections #-}
8+{-# LANGUAGE TypeFamilies #-}
9+{-# LANGUAGE FlexibleContexts #-}
10+
11+-- This is a simple generative adversarial network to make pictures
12+-- of numbers similar to those in MNIST.
13+--
14+-- It demonstrates a different usage of the library. Within about 15
15+-- minutes it was producing examples like this:
16+--
17+-- --.
18+-- .=-.--..#=###
19+-- -##==#########.
20+-- #############-
21+-- -###-.=..-.-==
22+-- ###-
23+-- .###-
24+-- .####...==-.
25+-- -####=--.=##=
26+-- -##=- -##
27+-- =##
28+-- -##=
29+-- -###-
30+-- .####.
31+-- .#####.
32+-- ...---=#####-
33+-- .=#########. .
34+-- .#######=. .
35+-- . =-.
36+--
37+-- It's a 5!
38+--
39+import Control.Applicative
40+import Control.Monad
41+import Control.Monad.Random
42+import Control.Monad.Trans.Except
43+
44+import qualified Data.Attoparsec.Text as A
45+import qualified Data.ByteString as B
46+import Data.List ( foldl' )
47+#if ! MIN_VERSION_base(4,13,0)
48+import Data.Semigroup ( (<>) )
49+#endif
50+import Data.Serialize
51+import qualified Data.Text as T
52+import qualified Data.Text.IO as T
53+import qualified Data.Vector.Storable as V
54+
55+import qualified Numeric.LinearAlgebra.Static as SA
56+import Numeric.LinearAlgebra.Data ( toLists )
57+
58+import Options.Applicative
59+
60+import Grenade
61+import Grenade.Utils.OneHot
62+
63+type Discriminator =
64+ Network
65+ '[ Convolution 1 10 5 5 1 1, Pooling 2 2 2 2, Relu
66+ , Convolution 10 16 5 5 1 1, Pooling 2 2 2 2, Relu
67+ , Reshape, FullyConnected 256 80, Logit, FullyConnected 80 1, Logit]
68+ '[ 'D2 28 28
69+ , 'D3 24 24 10, 'D3 12 12 10, 'D3 12 12 10
70+ , 'D3 8 8 16, 'D3 4 4 16, 'D3 4 4 16
71+ , 'D1 256, 'D1 80, 'D1 80, 'D1 1, 'D1 1]
72+
73+type Generator =
74+ Network
75+ '[ FullyConnected 80 256, Relu, Reshape
76+ , Deconvolution 16 10 5 5 2 2, Relu
77+ , Deconvolution 10 1 8 8 2 2, Logit]
78+ '[ 'D1 80
79+ , 'D1 256, 'D1 256, 'D3 4 4 16
80+ , 'D3 11 11 10, 'D3 11 11 10
81+ , 'D2 28 28, 'D2 28 28 ]
82+
83+randomDiscriminator :: MonadRandom m => m Discriminator
84+randomDiscriminator = randomNetwork
85+
86+randomGenerator :: MonadRandom m => m Generator
87+randomGenerator = randomNetwork
88+
89+trainExample :: LearningParameters -> Discriminator -> Generator -> S ('D2 28 28) -> S ('D1 80) -> ( Discriminator, Generator )
90+trainExample rate discriminator generator realExample noiseSource
91+ = let (generatorTape, fakeExample) = runNetwork generator noiseSource
92+
93+ (discriminatorTapeReal, guessReal) = runNetwork discriminator realExample
94+ (discriminatorTapeFake, guessFake) = runNetwork discriminator fakeExample
95+
96+ (discriminator'real, _) = runGradient discriminator discriminatorTapeReal ( guessReal - 1 )
97+ (discriminator'fake, _) = runGradient discriminator discriminatorTapeFake guessFake
98+ (_, push) = runGradient discriminator discriminatorTapeFake ( guessFake - 1)
99+
100+ (generator', _) = runGradient generator generatorTape push
101+
102+ newDiscriminator = foldl' (applyUpdate rate { learningRegulariser = learningRegulariser rate * 10}) discriminator [ discriminator'real, discriminator'fake ]
103+ newGenerator = applyUpdate rate generator generator'
104+ in ( newDiscriminator, newGenerator )
105+
106+
107+ganTest :: (Discriminator, Generator) -> Int -> FilePath -> LearningParameters -> ExceptT String IO (Discriminator, Generator)
108+ganTest (discriminator0, generator0) iterations trainFile rate = do
109+ trainData <- fmap fst <$> readMNIST trainFile
110+
111+ lift $ foldM (runIteration trainData) ( discriminator0, generator0 ) [1..iterations]
112+
113+ where
114+
115+ showShape' :: S ('D2 a b) -> IO ()
116+ showShape' (S2D mm) = putStrLn $
117+ let m = SA.extract mm
118+ ms = toLists m
119+ render n' | n' <= 0.2 = ' '
120+ | n' <= 0.4 = '.'
121+ | n' <= 0.6 = '-'
122+ | n' <= 0.8 = '='
123+ | otherwise = '#'
124+
125+ px = (fmap . fmap) render ms
126+ in unlines px
127+
128+ runIteration :: [S ('D2 28 28)] -> (Discriminator, Generator) -> Int -> IO (Discriminator, Generator)
129+ runIteration trainData ( !discriminator, !generator ) _ = do
130+ trained' <- foldM ( \(!discriminatorX, !generatorX ) realExample -> do
131+ fakeExample <- randomOfShape
132+ return $ trainExample rate discriminatorX generatorX realExample fakeExample
133+ ) ( discriminator, generator ) trainData
134+
135+
136+ showShape' . snd . runNetwork (snd trained') =<< randomOfShape
137+
138+ return trained'
139+
140+data GanOpts = GanOpts FilePath Int LearningParameters (Maybe FilePath) (Maybe FilePath)
141+
142+mnist' :: Parser GanOpts
143+mnist' = GanOpts <$> argument str (metavar "TRAIN")
144+ <*> option auto (long "iterations" <> short 'i' <> value 15)
145+ <*> (LearningParameters
146+ <$> option auto (long "train_rate" <> short 'r' <> value 0.01)
147+ <*> option auto (long "momentum" <> value 0.9)
148+ <*> option auto (long "l2" <> value 0.0005)
149+ )
150+ <*> optional (strOption (long "load"))
151+ <*> optional (strOption (long "save"))
152+
153+
154+main :: IO ()
155+main = do
156+ GanOpts mnist iter rate load save <- execParser (info (mnist' <**> helper) idm)
157+ putStrLn "Training stupidly simply GAN"
158+ nets0 <- case load of
159+ Just loadFile -> netLoad loadFile
160+ Nothing -> (,) <$> randomDiscriminator <*> randomGenerator
161+
162+ res <- runExceptT $ ganTest nets0 iter mnist rate
163+ case res of
164+ Right nets1 -> case save of
165+ Just saveFile -> B.writeFile saveFile $ runPut (put nets1)
166+ Nothing -> return ()
167+
168+ Left err -> putStrLn err
169+
170+readMNIST :: FilePath -> ExceptT String IO [(S ('D2 28 28), S ('D1 10))]
171+readMNIST mnist = ExceptT $ do
172+ mnistdata <- T.readFile mnist
173+ return $ traverse (A.parseOnly parseMNIST) (T.lines mnistdata)
174+
175+parseMNIST :: A.Parser (S ('D2 28 28), S ('D1 10))
176+parseMNIST = do
177+ Just lab <- oneHot <$> A.decimal
178+ pixels <- many (A.char ',' >> A.double)
179+ image <- maybe (fail "Parsed row was of an incorrect size") pure (fromStorable . V.fromList $ pixels)
180+ return (image, lab)
181+
182+netLoad :: FilePath -> IO (Discriminator, Generator)
183+netLoad modelPath = do
184+ modelData <- B.readFile modelPath
185+ either fail return $ runGet (get :: Get (Discriminator, Generator)) modelData
main/mnist.hs → examples/main/mnist.hsrenamed+30−2
@@ -1,10 +1,12 @@
1{-# LANGUAGE BangPatterns #-}
1{-# LANGUAGE BangPatterns #-}
2+{-# LANGUAGE CPP #-}
2{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE ScopedTypeVariables #-}
4{-# LANGUAGE ScopedTypeVariables #-}
4{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TupleSections #-}
6{-# LANGUAGE TupleSections #-}
6{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE FlexibleContexts #-}
8{-# LANGUAGE FlexibleContexts #-}
9+
8import Control.Applicative
10import Control.Applicative
9import Control.Monad
11import Control.Monad
10import Control.Monad.Random
12import Control.Monad.Random
@@ -12,6 +14,10 @@ import Control.Monad.Trans.Except
12
14
13import qualified Data.Attoparsec.Text as A
15import qualified Data.Attoparsec.Text as A
14import Data.List ( foldl' )
16import Data.List ( foldl' )
17+#if ! MIN_VERSION_base(4,13,0)
18+import Data.Semigroup ( (<>) )
19+#endif
20+
15import qualified Data.Text as T
21import qualified Data.Text as T
16import qualified Data.Text.IO as T
22import qualified Data.Text.IO as T
17import qualified Data.Vector.Storable as V
23import qualified Data.Vector.Storable as V
@@ -24,6 +30,15 @@ import Options.Applicative
24import Grenade
30import Grenade
25import Grenade.Utils.OneHot
31import Grenade.Utils.OneHot
26
32
33+-- It's logistic regression!
34+--
35+-- This network is used to show how we can embed a Network as a layer in the larger MNIST
36+-- type.
37+type FL i o =
38+ Network
39+ '[ FullyConnected i o, Logit ]
40+ '[ 'D1 i, 'D1 o, 'D1 o ]
41+
27-- The definition of our convolutional neural network.
42-- The definition of our convolutional neural network.
28-- In the type signature, we have a type level list of shapes which are passed between the layers.
43-- In the type signature, we have a type level list of shapes which are passed between the layers.
29-- One can see that the images we are inputing are two dimensional with 28 * 28 pixels.
44-- One can see that the images we are inputing are two dimensional with 28 * 28 pixels.
@@ -33,8 +48,21 @@ import Grenade.Utils.OneHot
33
48
34-- With the mnist data from Kaggle normalised to doubles between 0 and 1, learning rate of 0.01 and 15 iterations,
49-- With the mnist data from Kaggle normalised to doubles between 0 and 1, learning rate of 0.01 and 15 iterations,
35-- this network should get down to about a 1.3% error rate.
50-- this network should get down to about a 1.3% error rate.
36−type MNIST = Network '[ Convolution 1 10 5 5 1 1, Pooling 2 2 2 2, Relu, Convolution 10 16 5 5 1 1, Pooling 2 2 2 2, FlattenLayer, Relu, FullyConnected 256 80, Logit, FullyConnected 80 10, Logit]
51+--
37− '[ 'D2 28 28, 'D3 24 24 10, 'D3 12 12 10, 'D3 12 12 10, 'D3 8 8 16, 'D3 4 4 16, 'D1 256, 'D1 256, 'D1 80, 'D1 80, 'D1 10, 'D1 10]
52+-- /NOTE:/ This model is actually too complex for MNIST, and one should use the type given in the readme instead.
53+-- This one is just here to demonstrate Inception layers in use.
54+--
55+type MNIST =
56+ Network
57+ '[ Reshape,
58+ Concat ('D3 28 28 1) Trivial ('D3 28 28 14) (InceptionMini 28 28 1 5 9),
59+ Pooling 2 2 2 2, Relu,
60+ Concat ('D3 14 14 3) (Convolution 15 3 1 1 1 1) ('D3 14 14 15) (InceptionMini 14 14 15 5 10), Crop 1 1 1 1, Pooling 3 3 3 3, Relu,
61+ Reshape, FL 288 80, FL 80 10 ]
62+ '[ 'D2 28 28, 'D3 28 28 1,
63+ 'D3 28 28 15, 'D3 14 14 15, 'D3 14 14 15, 'D3 14 14 18,
64+ 'D3 12 12 18, 'D3 4 4 18, 'D3 4 4 18,
65+ 'D1 288, 'D1 80, 'D1 10 ]
38
66
39randomMnist :: MonadRandom m => m MNIST
67randomMnist :: MonadRandom m => m MNIST
40randomMnist = randomNetwork
68randomMnist = randomNetwork
main/recurrent.hs → examples/main/recurrent.hsrenamed+16−10
@@ -1,4 +1,5 @@
1{-# LANGUAGE BangPatterns #-}
1{-# LANGUAGE BangPatterns #-}
2+{-# LANGUAGE CPP #-}
2{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE ScopedTypeVariables #-}
4{-# LANGUAGE ScopedTypeVariables #-}
4{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeOperators #-}
@@ -8,7 +9,13 @@
8import Control.Monad ( foldM )
9import Control.Monad ( foldM )
9import Control.Monad.Random ( MonadRandom, getRandomR )
10import Control.Monad.Random ( MonadRandom, getRandomR )
10
11
12+#if MIN_VERSION_base(4,13,0)
13+import Data.List ( unfoldr )
14+#else
11import Data.List ( cycle, unfoldr )
15import Data.List ( cycle, unfoldr )
16+import Data.Semigroup ( (<>) )
17+#endif
18+
12import qualified Numeric.LinearAlgebra.Static as SA
19import qualified Numeric.LinearAlgebra.Static as SA
13
20
14import Options.Applicative
21import Options.Applicative
@@ -22,15 +29,14 @@ import Grenade.Recurrent
22--
29--
23-- The F and R types are Tagging types to ensure that the runner and
30-- The F and R types are Tagging types to ensure that the runner and
24-- creation function know how to treat the layers.
31-- creation function know how to treat the layers.
25−type F = FeedForward
26type R = Recurrent
32type R = Recurrent
27
33
28−type RecNet = RecurrentNetwork '[ R (LSTM 1 4), R (LSTM 4 1), F Trivial]
34+type RecNet = RecurrentNetwork '[ R (LSTM 1 1)]
29− '[ 'D1 1, 'D1 4, 'D1 1, 'D1 1 ]
35+ '[ 'D1 1, 'D1 1]
30
36
31−type RecInput = RecurrentInputs '[ R (LSTM 1 4), R (LSTM 4 1), F Trivial]
37+type RecInput = RecurrentInputs '[ R (LSTM 1 1)]
32
38
33−randomNet :: MonadRandom m => m (RecNet, RecInput)
39+randomNet :: MonadRandom m => m RecNet
34randomNet = randomRecurrent
40randomNet = randomRecurrent
35
41
36netTest :: MonadRandom m => RecNet -> RecInput -> LearningParameters -> Int -> m (RecNet, RecInput)
42netTest :: MonadRandom m => RecNet -> RecInput -> LearningParameters -> Int -> m (RecNet, RecInput)
@@ -51,7 +57,7 @@ netTest net0 i0 rate iterations =
51data FeedForwardOpts = FeedForwardOpts Int LearningParameters
57data FeedForwardOpts = FeedForwardOpts Int LearningParameters
52
58
53feedForward' :: Parser FeedForwardOpts
59feedForward' :: Parser FeedForwardOpts
54−feedForward' = FeedForwardOpts <$> option auto (long "examples" <> short 'e' <> value 20000)
60+feedForward' = FeedForwardOpts <$> option auto (long "examples" <> short 'e' <> value 40000)
55 <*> (LearningParameters
61 <*> (LearningParameters
56 <$> option auto (long "train_rate" <> short 'r' <> value 0.01)
62 <$> option auto (long "train_rate" <> short 'r' <> value 0.01)
57 <*> option auto (long "momentum" <> value 0.9)
63 <*> option auto (long "momentum" <> value 0.9)
@@ -63,8 +69,8 @@ generateRecurrent n s i =
63 unfoldr go (s, i)
69 unfoldr go (s, i)
64 where
70 where
65 go (x, y) =
71 go (x, y) =
66− do let (ns, o) = runRecurrent n x y
72+ do let (_, ns, o) = runRecurrent n x y
67− o' = heat o
73+ o' = heat o
68 Just (o', (ns, fromIntegral o'))
74 Just (o', (ns, fromIntegral o'))
69
75
70 heat :: S ('D1 1) -> Int
76 heat :: S ('D1 1) -> Int
@@ -76,8 +82,8 @@ main = do
76 FeedForwardOpts examples rate <- execParser (info (feedForward' <**> helper) idm)
82 FeedForwardOpts examples rate <- execParser (info (feedForward' <**> helper) idm)
77 putStrLn "Training network..."
83 putStrLn "Training network..."
78
84
79− (net0, i0) <- randomNet
85+ net0 <- randomNet
80− (trained, bestInput) <- netTest net0 i0 rate examples
86+ (trained, bestInput) <- netTest net0 0 rate examples
81
87
82 let results = generateRecurrent trained bestInput (c 1)
88 let results = generateRecurrent trained bestInput (c 1)
83
89
main/shakespeare.hs → examples/main/shakespeare.hsrenamed+26−22
@@ -1,5 +1,5 @@
1{-# LANGUAGE BangPatterns #-}
1{-# LANGUAGE BangPatterns #-}
2−{-# LANGUAGE RecordWildCards #-}
2+{-# LANGUAGE CPP #-}
3{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE DataKinds #-}
4{-# LANGUAGE ScopedTypeVariables #-}
4{-# LANGUAGE ScopedTypeVariables #-}
5{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeOperators #-}
@@ -14,16 +14,15 @@ import Data.Char ( isUpper, toUpper, toLower )
14import Data.List ( foldl' )
14import Data.List ( foldl' )
15import Data.Maybe ( fromMaybe )
15import Data.Maybe ( fromMaybe )
16
16
17+
17import qualified Data.Vector as V
18import qualified Data.Vector as V
18import Data.Vector ( Vector )
19import Data.Vector ( Vector )
19
20
20import qualified Data.Map as M
21import qualified Data.Map as M
21−import Data.Proxy ( Proxy (..) )
22
22
23import qualified Data.ByteString as B
23import qualified Data.ByteString as B
24import Data.Serialize
24import Data.Serialize
25
25
26−import Data.Singletons.Prelude
27import GHC.TypeLits
26import GHC.TypeLits
28
27
29import Numeric.LinearAlgebra.Static ( konst )
28import Numeric.LinearAlgebra.Static ( konst )
@@ -35,11 +34,16 @@ import Grenade.Recurrent
35import Grenade.Utils.OneHot
34import Grenade.Utils.OneHot
36
35
37import System.IO.Unsafe ( unsafeInterleaveIO )
36import System.IO.Unsafe ( unsafeInterleaveIO )
37+import Data.Proxy
38+import Prelude.Singletons
38
39
39-- The defininition for our natural language recurrent network.
40-- The defininition for our natural language recurrent network.
40-- This network is able to learn and generate simple words in
41-- This network is able to learn and generate simple words in
41-- about an hour.
42-- about an hour.
42--
43--
44+-- Grab the input from
45+-- https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt
46+--
43-- This is a first class recurrent net.
47-- This is a first class recurrent net.
44--
48--
45-- The F and R types are tagging types to ensure that the runner and
49-- The F and R types are tagging types to ensure that the runner and
@@ -60,7 +64,7 @@ type Shakespeare = RecurrentNetwork '[ R (LSTM 40 80), R (LSTM 80 40), F (FullyC
60-- The definition of the "sideways" input, which the network is fed recurrently.
64-- The definition of the "sideways" input, which the network is fed recurrently.
61type Shakespearian = RecurrentInputs '[ R (LSTM 40 80), R (LSTM 80 40), F (FullyConnected 40 40), F Logit]
65type Shakespearian = RecurrentInputs '[ R (LSTM 40 80), R (LSTM 80 40), F (FullyConnected 40 40), F Logit]
62
66
63−randomNet :: MonadRandom m => m (Shakespeare, Shakespearian)
67+randomNet :: MonadRandom m => m Shakespeare
64randomNet = randomRecurrent
68randomNet = randomRecurrent
65
69
66-- | Load the data files and prepare a map of characters to a compressed int representation.
70-- | Load the data files and prepare a map of characters to a compressed int representation.
@@ -68,40 +72,40 @@ loadShakespeare :: FilePath -> ExceptT String IO (Vector Int, M.Map Char Int, Ve
68loadShakespeare path = do
72loadShakespeare path = do
69 contents <- lift $ readFile path
73 contents <- lift $ readFile path
70 let annotated = annotateCapitals contents
74 let annotated = annotateCapitals contents
71− (m,cs) <- ExceptT . return . note "Couldn't fit data in hotMap" $ hotMap (Proxy :: Proxy 40) annotated
75+ (m,cs) <- ExceptT . return $ hotMap (Proxy :: Proxy 40) annotated
72 hot <- ExceptT . return . note "Couldn't generate hot values" $ traverse (`M.lookup` m) annotated
76 hot <- ExceptT . return . note "Couldn't generate hot values" $ traverse (`M.lookup` m) annotated
73 return (V.fromList hot, m, cs)
77 return (V.fromList hot, m, cs)
74
78
75trainSlice :: LearningParameters -> Shakespeare -> Shakespearian -> Vector Int -> Int -> Int -> (Shakespeare, Shakespearian)
79trainSlice :: LearningParameters -> Shakespeare -> Shakespearian -> Vector Int -> Int -> Int -> (Shakespeare, Shakespearian)
76−trainSlice !rate !net !recIns input offset size =
80+trainSlice !lrate !net !recIns input offset size =
77 let e = fmap (x . oneHot) . V.toList $ V.slice offset size input
81 let e = fmap (x . oneHot) . V.toList $ V.slice offset size input
78 in case reverse e of
82 in case reverse e of
79 (o : l : xs) ->
83 (o : l : xs) ->
80 let examples = reverse $ (l, Just o) : ((,Nothing) <$> xs)
84 let examples = reverse $ (l, Just o) : ((,Nothing) <$> xs)
81− in trainRecurrent rate net recIns examples
85+ in trainRecurrent lrate net recIns examples
82 _ -> error "Not enough input"
86 _ -> error "Not enough input"
83 where
87 where
84 x = fromMaybe (error "Hot variable didn't fit.")
88 x = fromMaybe (error "Hot variable didn't fit.")
85
89
86runShakespeare :: ShakespeareOpts -> ExceptT String IO ()
90runShakespeare :: ShakespeareOpts -> ExceptT String IO ()
87−runShakespeare ShakespeareOpts {..} = do
91+runShakespeare opts = do
88− (shakespeare, oneHotMap, oneHotDictionary) <- loadShakespeare trainingFile
92+ (shakespeare, oneHotMap, oneHotDictionary) <- loadShakespeare $ trainingFile opts
89 (net0, i0) <- lift $
93 (net0, i0) <- lift $
90− case loadPath of
94+ case loadPath opts of
91 Just loadFile -> netLoad loadFile
95 Just loadFile -> netLoad loadFile
92− Nothing -> randomNet
96+ Nothing -> (,0) <$> randomNet
93
97
94 (trained, bestInput) <- lift $ foldM (\(!net, !io) size -> do
98 (trained, bestInput) <- lift $ foldM (\(!net, !io) size -> do
95− xs <- take (iterations `div` 10) <$> getRandomRs (0, length shakespeare - size - 1)
99+ xs <- take (iterations opts `div` 10) <$> getRandomRs (0, length shakespeare - size - 1)
96− let (!trained, !bestInput) = foldl' (\(!n, !i) offset -> trainSlice rate n i shakespeare offset size) (net, io) xs
100+ let (!trained, !bestInput) = foldl' (\(!n, !i) offset -> trainSlice (rate opts) n i shakespeare offset size) (net, io) xs
97− results <- take 1000 <$> generateParagraph trained bestInput temperature oneHotMap oneHotDictionary ( S1D $ konst 0)
101+ results <- take 1000 <$> generateParagraph trained bestInput (temperature opts) oneHotMap oneHotDictionary ( S1D $ konst 0)
98 putStrLn ("TRAINING STEP WITH SIZE: " ++ show size)
102 putStrLn ("TRAINING STEP WITH SIZE: " ++ show size)
99 putStrLn (unAnnotateCapitals results)
103 putStrLn (unAnnotateCapitals results)
100 return (trained, bestInput)
104 return (trained, bestInput)
101− ) (net0, i0) $ replicate 10 sequenceSize
105+ ) (net0, i0) $ replicate 10 (sequenceSize opts)
102
106
103− case savePath of
107+ case savePath opts of
104− Just saveFile -> lift . B.writeFile saveFile $ runPut (put (trained, bestInput))
108+ Just saveFile -> lift . B.writeFile saveFile $ runPut (put trained >> put bestInput)
105 Nothing -> return ()
109 Nothing -> return ()
106
110
107generateParagraph :: forall layers shapes n a. (Last shapes ~ 'D1 n, Head shapes ~ 'D1 n, KnownNat n, Ord a)
111generateParagraph :: forall layers shapes n a. (Last shapes ~ 'D1 n, Head shapes ~ 'D1 n, KnownNat n, Ord a)
@@ -112,14 +116,14 @@ generateParagraph :: forall layers shapes n a. (Last shapes ~ 'D1 n, Head shapes
112 -> Vector a
116 -> Vector a
113 -> S ('D1 n)
117 -> S ('D1 n)
114 -> IO [a]
118 -> IO [a]
115−generateParagraph n s temperature hotmap hotdict =
119+generateParagraph n s temp hotmap hotdict =
116 go s
120 go s
117 where
121 where
118 go x y =
122 go x y =
119− do let (ns, o) = runRecurrent n x y
123+ do let (_, ns, o) = runRecurrent n x y
120− un <- sample temperature hotdict o
124+ un <- sample temp hotdict o
121− Just re <- return $ makeHot hotmap un
125+ Just re <- return $ makeHot hotmap un
122− rest <- unsafeInterleaveIO $ go ns re
126+ rest <- unsafeInterleaveIO $ go ns re
123 return (un : rest)
127 return (un : rest)
124
128
125data ShakespeareOpts = ShakespeareOpts {
129data ShakespeareOpts = ShakespeareOpts {
framework/mafianew file+62
@@ -0,0 +1,62 @@
1+#!/bin/sh -eu
2+
3+: ${MAFIA_HOME:=$HOME/.mafia}
4+: ${MAFIA_VERSIONS:=$MAFIA_HOME/versions}
5+
6+latest_version () {
7+ git ls-remote https://github.com/haskell-mafia/mafia | grep refs/heads/master | cut -f 1
8+}
9+
10+build_version() {
11+ MAFIA_VERSION="$1"
12+ MAFIA_TEMP=$(mktemp -d 2>/dev/null || mktemp -d -t 'exec_mafia')
13+ MAFIA_FILE=mafia-$MAFIA_VERSION
14+ MAFIA_PATH=$MAFIA_VERSIONS/$MAFIA_FILE
15+ mkdir -p $MAFIA_VERSIONS
16+ echo "Building $MAFIA_FILE in $MAFIA_TEMP"
17+ git clone https://github.com/haskell-mafia/mafia $MAFIA_TEMP
18+ git --git-dir="$MAFIA_TEMP/.git" --work-tree="$MAFIA_TEMP" reset --hard $MAFIA_VERSION || {
19+ echo "mafia version ($MAFIA_VERSION) could not be found." >&2
20+ exit 1
21+ }
22+ (cd "$MAFIA_TEMP" && ./bin/bootstrap) || {
23+ got=$?
24+ echo "mafia version ($MAFIA_VERSION) could not be built." >&2
25+ exit "$got"
26+ }
27+ chmod +x "$MAFIA_TEMP/.cabal-sandbox/bin/mafia"
28+ # Ensure executable is on same file-system so final mv is atomic.
29+ mv -f "$MAFIA_TEMP/.cabal-sandbox/bin/mafia" "$MAFIA_PATH.$$"
30+ mv "$MAFIA_PATH.$$" "$MAFIA_PATH" || {
31+ rm -f "$MAFIA_PATH.$$"
32+ echo "INFO: mafia version ($MAFIA_VERSION) already exists not overiding," >&2
33+ echo "INFO: this is expected if parallel builds of the same version of" >&2
34+ echo "INFO: mafia occur, we are playing by first in, wins." >&2
35+ exit 0
36+ }
37+}
38+
39+enable_version() {
40+ if [ $# -eq 0 ]; then
41+ MAFIA_VERSION="$(latest_version)"
42+ echo "INFO: No explicit mafia version requested installing latest ($MAFIA_VERSION)." >&2
43+ else
44+ MAFIA_VERSION="$1"
45+ fi
46+ [ -x "$MAFIA_HOME/versions/mafia-$MAFIA_VERSION" ] || build_version "$MAFIA_VERSION"
47+ ln -sf "$MAFIA_HOME/versions/mafia-$MAFIA_VERSION" "$MAFIA_HOME/versions/mafia"
48+}
49+
50+exec_mafia () {
51+ [ -x "$MAFIA_HOME/versions/mafia" ] || enable_version
52+ "$MAFIA_HOME/versions/mafia" "$@"
53+}
54+
55+#
56+# The actual start of the script.....
57+#
58+
59+case "${1:-}" in
60+upgrade) shift; enable_version "$@" ;;
61+*) exec_mafia "$@"
62+esac
grenade.cabal+118−116
@@ -1,153 +1,126 @@
1name: grenade
1name: grenade
2−version: 0.0.1
2+version: 0.1.0
3license: BSD2
3license: BSD2
4+license-file: LICENSE
4author: Huw Campbell <huw.campbell@gmail.com>
5author: Huw Campbell <huw.campbell@gmail.com>
5maintainer: Huw Campbell <huw.campbell@gmail.com>
6maintainer: Huw Campbell <huw.campbell@gmail.com>
6−copyright: (c) 2015 Huw Campbell.
7+copyright: (c) 2016-2017 Huw Campbell.
7−synopsis: grenade
8+synopsis: Practical Deep Learning in Haskell
8−category: System
9+category: AI, Machine Learning
9−cabal-version: >= 1.8
10+cabal-version: >= 1.10
10build-type: Simple
11build-type: Simple
11−description: grenade.
12+description:
13+ Grenade is a composable, dependently typed, practical, and fast
14+ recurrent neural network library for precise specifications and
15+ complex deep neural networks in Haskell.
16+ .
17+ Grenade provides an API for composing layers of a neural network
18+ into a sequence parallel graph in a type safe manner; running
19+ networks with reverse automatic differentiation to calculate their
20+ gradients; and applying gradient descent for learning.
21+ .
22+ Documentation and examples are available on github
23+ <https://github.com/HuwCampbell/grenade>.
12
24
13extra-source-files:
25extra-source-files:
26+ README.md
14− cbits/im2col.h
27+ cbits/im2col.h
15− cbits/im2col.c
28+ cbits/im2col.c
29+ cbits/gradient_descent.h
30+ cbits/gradient_descent.c
31+ cbits/pad.h
32+ cbits/pad.c
33+
34+source-repository head
35+ type: git
36+ location: https://github.com/HuwCampbell/grenade.git
16
37
17library
38library
18 build-depends:
39 build-depends:
19 base >= 4.8 && < 5
40 base >= 4.8 && < 5
20− , bytestring == 0.10.*
41+ , bytestring >= 0.10.0
21− , async
22− , containers
42+ , containers >= 0.5 && < 0.7
23− , deepseq
24− , either == 4.4.*
25− , cereal
43+ , cereal >= 0.5 && < 0.6
26− , exceptions == 0.8.*
44+ , deepseq >= 1.4 && < 1.5
27− , hmatrix
45+ , hmatrix >= 0.18 && < 0.21
28− , MonadRandom
46+ , MonadRandom >= 0.4 && < 0.6
29− , mtl >= 2.2.1 && < 2.3
30− , parallel == 3.2.*
31− , primitive
47+ , primitive >= 0.6 && < 0.8
32− , reflection
48+ -- Versions of singletons are *tightly* coupled with the
33− , text == 1.2.*
49+ -- GHC version so its fine to drop version bounds.
34− , transformers
35 , singletons
50 , singletons
51+ , singletons-base
36− , vector
52+ , vector >= 0.11 && < 0.13
37
53
38 ghc-options:
54 ghc-options:
39 -Wall
55 -Wall
40 hs-source-dirs:
56 hs-source-dirs:
41 src
57 src
42
58
59+ default-language: Haskell2010
60+
61+ if impl(ghc < 8.0)
62+ ghc-options: -fno-warn-incomplete-patterns
63+ cpp-options: -DType=*
64+
65+ if impl(ghc >= 8.6)
66+ default-extensions: NoStarIsType
43
67
44 exposed-modules:
68 exposed-modules:
45 Grenade
69 Grenade
70+ Grenade.Core
71+ Grenade.Core.Layer
72+ Grenade.Core.LearningParameters
46 Grenade.Core.Network
73 Grenade.Core.Network
47 Grenade.Core.Runner
74 Grenade.Core.Runner
48 Grenade.Core.Shape
75 Grenade.Core.Shape
76+
77+ Grenade.Layers
49− Grenade.Layers.Crop
78+ Grenade.Layers.Concat
50 Grenade.Layers.Convolution
79 Grenade.Layers.Convolution
80+ Grenade.Layers.Crop
81+ Grenade.Layers.Deconvolution
51 Grenade.Layers.Dropout
82 Grenade.Layers.Dropout
83+ Grenade.Layers.Elu
52 Grenade.Layers.FullyConnected
84 Grenade.Layers.FullyConnected
53− Grenade.Layers.Flatten
85+ Grenade.Layers.Inception
54− Grenade.Layers.Fuse
55 Grenade.Layers.Logit
86 Grenade.Layers.Logit
56− Grenade.Layers.Relu
87+ Grenade.Layers.Merge
57− Grenade.Layers.Tanh
58 Grenade.Layers.Pad
88 Grenade.Layers.Pad
59 Grenade.Layers.Pooling
89 Grenade.Layers.Pooling
90+ Grenade.Layers.Relu
91+ Grenade.Layers.Reshape
92+ Grenade.Layers.Sinusoid
93+ Grenade.Layers.Softmax
94+ Grenade.Layers.Tanh
95+ Grenade.Layers.Trivial
60
96
61 Grenade.Layers.Internal.Convolution
97 Grenade.Layers.Internal.Convolution
98+ Grenade.Layers.Internal.Pad
62 Grenade.Layers.Internal.Pooling
99 Grenade.Layers.Internal.Pooling
63 Grenade.Layers.Internal.Update
100 Grenade.Layers.Internal.Update
64
101
65 Grenade.Recurrent
102 Grenade.Recurrent
66
103
104+ Grenade.Recurrent.Core
105+ Grenade.Recurrent.Core.Layer
67 Grenade.Recurrent.Core.Network
106 Grenade.Recurrent.Core.Network
68 Grenade.Recurrent.Core.Runner
107 Grenade.Recurrent.Core.Runner
69
108
109+ Grenade.Recurrent.Layers
70 Grenade.Recurrent.Layers.BasicRecurrent
110 Grenade.Recurrent.Layers.BasicRecurrent
111+ Grenade.Recurrent.Layers.ConcatRecurrent
71 Grenade.Recurrent.Layers.LSTM
112 Grenade.Recurrent.Layers.LSTM
72− Grenade.Recurrent.Layers.Trivial
73
113
74 Grenade.Utils.OneHot
114 Grenade.Utils.OneHot
75
115
76− includes: cbits/im2col.h
116+ includes: cbits/im2col.h
77− cbits/gradient_decent.h
117+ cbits/gradient_descent.h
118+ cbits/pad.h
78− c-sources: cbits/im2col.c
119+ c-sources: cbits/im2col.c
79− cbits/gradient_decent.c
120+ cbits/gradient_descent.c
80−
121+ cbits/pad.c
81− cc-options: -std=c99 -O3 -msse4.2 -Wall -Werror -DCABAL=1
82−
83−executable feedforward
84− ghc-options: -Wall -threaded -O2
85− main-is: main/feedforward.hs
86− build-depends: base
87− , grenade
88− , attoparsec
89− , bytestring
90− , cereal
91− , either
92− , optparse-applicative == 0.12.*
93− , text == 1.2.*
94− , mtl >= 2.2.1 && < 2.3
95− , hmatrix
96− , transformers
97− , singletons
98− , MonadRandom
99−
100−executable mnist
101− ghc-options: -Wall -threaded -O2
102− main-is: main/mnist.hs
103− build-depends: base
104− , grenade
105− , attoparsec
106− , either
107− , optparse-applicative == 0.12.*
108− , text == 1.2.*
109− , mtl >= 2.2.1 && < 2.3
110− , hmatrix >= 0.18 && < 0.19
111− , transformers
112− , singletons
113− , MonadRandom
114− , vector
115−
116−executable recurrent
117− ghc-options: -Wall -threaded -O2
118− main-is: main/recurrent.hs
119− build-depends: base
120− , grenade
121− , attoparsec
122− , either
123− , optparse-applicative == 0.12.*
124− , text == 1.2.*
125− , mtl >= 2.2.1 && < 2.3
126− , hmatrix >= 0.18 && < 0.19
127− , transformers
128− , singletons
129− , MonadRandom
130−
131−
132−executable shakespeare
133− ghc-options: -Wall -threaded -O2
134− main-is: main/shakespeare.hs
135− build-depends: base
136− , grenade
137− , attoparsec
138− , bytestring
139− , cereal
140− , either
141− , optparse-applicative == 0.12.*
142− , text == 1.2.*
143− , mtl >= 2.2.1 && < 2.3
144− , hmatrix >= 0.18 && < 0.19
145− , transformers
146− , singletons
147− , vector
148− , MonadRandom
149− , containers
150
122
123+ cc-options: -std=c99 -O3 -msse4.2 -Wall -Werror -DCABAL=1
151
124
152test-suite test
125test-suite test
153 type: exitcode-stdio-1.0
126 type: exitcode-stdio-1.0
@@ -159,23 +132,48 @@ test-suite test
159 hs-source-dirs:
132 hs-source-dirs:
160 test
133 test
161
134
135+ default-language: Haskell2010
136+
137+ other-modules: Test.Hedgehog.Compat
138+ Test.Hedgehog.Hmatrix
139+ Test.Hedgehog.TypeLits
140+
141+ Test.Grenade.Network
142+ Test.Grenade.Layers.Convolution
143+ Test.Grenade.Layers.FullyConnected
144+ Test.Grenade.Layers.Nonlinear
145+ Test.Grenade.Layers.PadCrop
146+ Test.Grenade.Layers.Pooling
147+ Test.Grenade.Layers.Internal.Convolution
148+ Test.Grenade.Layers.Internal.Pooling
149+ Test.Grenade.Layers.Internal.Reference
150+
151+ Test.Grenade.Recurrent.Layers.LSTM
152+ Test.Grenade.Recurrent.Layers.LSTM.Reference
153+
154+ if impl(ghc < 8.0)
155+ ghc-options: -fno-warn-incomplete-patterns
156+ cpp-options: -DType=*
157+
158+ if impl(ghc >= 8.6)
159+ default-extensions: NoStarIsType
160+
162 build-depends:
161 build-depends:
163− base >= 4.8 && < 5
162+ base
164 , grenade
163 , grenade
165− , ambiata-disorder-core
164+ , hedgehog >= 1.0 && < 1.1
166− , ambiata-disorder-jack
167 , hmatrix
165 , hmatrix
168 , mtl
166 , mtl
169 , singletons
167 , singletons
170 , text == 1.2.*
168 , text == 1.2.*
171− , typelits-witnesses
169+ , typelits-witnesses < 0.4
170+ , transformers
172 , constraints
171 , constraints
173− , QuickCheck >= 2.7 && < 2.9
174− , quickcheck-instances == 0.3.*
175 , MonadRandom
172 , MonadRandom
176 , random
173 , random
177 , ad
174 , ad
178 , reflection
175 , reflection
176+ , vector
179
177
180
178
181benchmark bench
179benchmark bench
@@ -188,10 +186,12 @@ benchmark bench
188 hs-source-dirs:
186 hs-source-dirs:
189 bench
187 bench
190
188
189+ default-language: Haskell2010
190+
191 build-depends:
191 build-depends:
192− base >= 3 && < 5
192+ base
193− , bytestring == 0.10.*
193+ , bytestring
194− , criterion == 1.1.*
194+ , criterion >= 1.1 && < 1.6
195 , grenade
195 , grenade
196 , hmatrix
196 , hmatrix
197
197
@@ -205,9 +205,11 @@ benchmark bench-lstm
205 hs-source-dirs:
205 hs-source-dirs:
206 bench
206 bench
207
207
208+ default-language: Haskell2010
209+
208 build-depends:
210 build-depends:
209− base >= 3 && < 5
211+ base
210− , bytestring == 0.10.*
212+ , bytestring
211− , criterion == 1.1.*
213+ , criterion
212 , grenade
214 , grenade
213 , hmatrix
215 , hmatrix
lib/disorder.hsdeleted−1

Submodule changed.

lib/pdeleted−1

Submodule changed.

mafiamode 100755 → 120000+1−123
@@ -1,123 +0,0 @@
1−#!/bin/sh -eu
2−
3−: ${MAFIA_HOME:=$HOME/.mafia}
4−
5−fetch_latest () {
6− if [ -z ${MAFIA_TEST_MODE+x} ]; then
7− TZ=$(date +"%T")
8− curl --silent "https://raw.githubusercontent.com/ambiata/mafia/master/script/mafia?$TZ"
9− else
10− cat ../script/mafia
11− fi
12−}
13−
14−latest_version () {
15− git ls-remote https://github.com/ambiata/mafia | grep refs/heads/master | cut -f 1
16−}
17−
18−local_version () {
19− awk '/^# Version: / { print $3; exit 0; }' $0
20−}
21−
22−run_upgrade () {
23− MAFIA_TEMP=$(mktemp 2>/dev/null || mktemp -t 'upgrade_mafia')
24−
25− clean_up () {
26− rm -f "$MAFIA_TEMP"
27− }
28−
29− trap clean_up EXIT
30−
31− MAFIA_CUR="$0"
32−
33− if [ -L "$MAFIA_CUR" ]; then
34− echo 'Refusing to overwrite a symlink; run `upgrade` from the canonical path.' >&2
35− exit 1
36− fi
37−
38− echo "Checking for a new version of mafia ..."
39− fetch_latest > $MAFIA_TEMP
40−
41− LATEST_VERSION=$(latest_version)
42− echo "# Version: $LATEST_VERSION" >> $MAFIA_TEMP
43−
44− if ! cmp $MAFIA_CUR $MAFIA_TEMP >/dev/null 2>&1; then
45− mv $MAFIA_TEMP $MAFIA_CUR
46− chmod +x $MAFIA_CUR
47− echo "New version found and upgraded. You can now commit it to your git repo."
48− else
49− echo "You have latest mafia."
50− fi
51−}
52−
53−exec_mafia () {
54− MAFIA_VERSION=$(local_version)
55−
56− if [ "x$MAFIA_VERSION" = "x" ]; then
57− # If we can't find the mafia version, then we need to upgrade the script.
58− run_upgrade
59− else
60− MAFIA_BIN=$MAFIA_HOME/bin
61− MAFIA_FILE=mafia-$MAFIA_VERSION
62− MAFIA_PATH=$MAFIA_BIN/$MAFIA_FILE
63−
64− [ -f "$MAFIA_PATH" ] || {
65− # Create a temporary directory which will be deleted when the script
66− # terminates. Unfortunately `mktemp` doesn't behave the same on
67− # Linux and OS/X so we need to try two different approaches.
68− MAFIA_TEMP=$(mktemp -d 2>/dev/null || mktemp -d -t 'exec_mafia')
69−
70− # Create a temporary file in MAFIA_BIN so we can do an atomic copy/move dance.
71− mkdir -p $MAFIA_BIN
72−
73− clean_up () {
74− rm -rf "$MAFIA_TEMP"
75− }
76−
77− trap clean_up EXIT
78−
79− echo "Building $MAFIA_FILE in $MAFIA_TEMP"
80−
81− ( cd "$MAFIA_TEMP"
82−
83− git clone https://github.com/ambiata/mafia
84− cd mafia
85−
86− git reset --hard $MAFIA_VERSION
87−
88− bin/bootstrap ) || exit $?
89−
90− MAFIA_PATH_TEMP=$(mktemp --tmpdir=$MAFIA_BIN $MAFIA_FILE-XXXXXX 2>/dev/null || TMPDIR=$MAFIA_BIN mktemp -t $MAFIA_FILE)
91−
92− clean_up_temp () {
93− clean_up
94− rm -f "$MAFIA_PATH_TEMP"
95− }
96− trap clean_up_temp EXIT
97−
98− cp "$MAFIA_TEMP/mafia/.cabal-sandbox/bin/mafia" "$MAFIA_PATH_TEMP"
99− chmod 755 "$MAFIA_PATH_TEMP"
100− mv "$MAFIA_PATH_TEMP" "$MAFIA_PATH"
101−
102− clean_up_temp
103− }
104−
105− exec $MAFIA_PATH "$@"
106− fi
107−}
108−
109−#
110−# The actual start of the script.....
111−#
112−
113−if [ $# -gt 0 ]; then
114− MODE="$1"
115−else
116− MODE=""
117−fi
118−
119−case "$MODE" in
120−upgrade) shift; run_upgrade "$@" ;;
121−*) exec_mafia "$@"
122−esac
123−# Version: 3044e63eb472fb9e16926d4ab2ca9dd9e255829c
@@ -0,0 +1 @@
1+framework/mafia
src/Grenade.hs+50−15
@@ -1,18 +1,53 @@
1module Grenade (
1module Grenade (
2+ -- | This is an empty module which simply re-exports public definitions
3+ -- for machine learning with Grenade.
4+
5+ -- * Exported modules
6+ --
7+ -- | The core types and runners for Grenade.
2− module X
8+ module Grenade.Core
9+
10+ -- | The neural network layer zoo
11+ , module Grenade.Layers
12+
13+
14+ -- * Overview of the library
15+ -- $library
16+
17+ -- * Example usage
18+ -- $example
19+
3 ) where
20 ) where
4
21
5−import Grenade.Core.Network as X
22+import Grenade.Core
6−import Grenade.Core.Runner as X
23+import Grenade.Layers
7−import Grenade.Core.Shape as X
24+
8−import Grenade.Layers.Crop as X
25+{- $library
9−import Grenade.Layers.Dropout as X
26+Grenade is a purely functional deep learning library.
10−import Grenade.Layers.Pad as X
27+
11−import Grenade.Layers.Pooling as X
28+It provides an expressive type level API for the construction
12−import Grenade.Layers.Flatten as X
29+of complex neural network architectures. Backing this API is and
13−import Grenade.Layers.Fuse as X
30+implementation written using BLAS and LAPACK, mostly provided by
14−import Grenade.Layers.FullyConnected as X
31+the hmatrix library.
15−import Grenade.Layers.Logit as X
32+
16−import Grenade.Layers.Convolution as X
33+-}
17−import Grenade.Layers.Relu as X
34+
18−import Grenade.Layers.Tanh as X
35+{- $example
36+A few examples are provided at https://github.com/HuwCampbell/grenade
37+under the examples folder.
38+
39+The starting place is to write your neural network type and a
40+function to create a random layer of that type. The following
41+is a simple example which runs a logistic regression.
42+
43+> type MyNet = Network '[ FullyConnected 10 1, Logit ] '[ 'D1 10, 'D1 1, 'D1 1 ]
44+>
45+> randomMyNet :: MonadRandom MyNet
46+> randomMyNet = randomNetwork
47+
48+The function `randomMyNet` witnesses the `CreatableNetwork`
49+constraint of the neural network, and in doing so, ensures the network
50+can be built, and hence, that the architecture is sound.
51+-}
52+
53+
src/Grenade/Core.hsnew file+13
@@ -0,0 +1,13 @@
1+module Grenade.Core (
2+ module Grenade.Core.Layer
3+ , module Grenade.Core.LearningParameters
4+ , module Grenade.Core.Network
5+ , module Grenade.Core.Runner
6+ , module Grenade.Core.Shape
7+ ) where
8+
9+import Grenade.Core.Layer
10+import Grenade.Core.LearningParameters
11+import Grenade.Core.Network
12+import Grenade.Core.Runner
13+import Grenade.Core.Shape
src/Grenade/Core/Layer.hsnew file+92
@@ -0,0 +1,92 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE DataKinds #-}
3+{-# LANGUAGE GADTs #-}
4+{-# LANGUAGE TypeOperators #-}
5+{-# LANGUAGE TypeFamilies #-}
6+{-# LANGUAGE MultiParamTypeClasses #-}
7+{-# LANGUAGE FlexibleContexts #-}
8+{-# LANGUAGE RankNTypes #-}
9+{-# LANGUAGE FlexibleInstances #-}
10+{-|
11+Module : Grenade.Core.Layer
12+Description : Defines the Layer Classes
13+Copyright : (c) Huw Campbell, 2016-2017
14+License : BSD2
15+Stability : experimental
16+
17+This module defines what a Layer is in a Grenade
18+neural network.
19+
20+There are two classes of interest: `UpdateLayer` and `Layer`.
21+
22+`UpdateLayer` is required for all types which are used as a layer
23+in a network. Having no shape information, this class is agnostic
24+to the input and output data of the layer.
25+
26+An instance of `Layer` on the other hand is required for usage in
27+a neural network, but also specifies the shapes of data that the
28+network can transform. Multiple instance of `Layer` are permitted
29+for a single type, to transform different shapes. The `Reshape` layer
30+for example can act as a flattening layer, and its inverse, projecting
31+a 1D shape up to 2 or 3 dimensions.
32+
33+Instances of `Layer` should be as strict as possible, and not emit
34+runtime errors.
35+-}
36+module Grenade.Core.Layer (
37+ Layer (..)
38+ , UpdateLayer (..)
39+ ) where
40+
41+import Control.Monad.Random ( MonadRandom )
42+
43+import Data.List ( foldl' )
44+
45+#if MIN_VERSION_base(4,9,0)
46+import Data.Kind (Type)
47+#endif
48+
49+import Grenade.Core.Shape
50+import Grenade.Core.LearningParameters
51+
52+-- | Class for updating a layer. All layers implement this, as it
53+-- describes how to create and update the layer.
54+--
55+class UpdateLayer x where
56+ -- | The type for the gradient for this layer.
57+ -- Unit if there isn't a gradient to pass back.
58+ type Gradient x :: Type
59+
60+ -- | Update a layer with its gradient and learning parameters
61+ runUpdate :: LearningParameters -> x -> Gradient x -> x
62+
63+ -- | Create a random layer, many layers will use pure
64+ createRandom :: MonadRandom m => m x
65+
66+ -- | Update a layer with many Gradients
67+ runUpdates :: LearningParameters -> x -> [Gradient x] -> x
68+ runUpdates rate = foldl' (runUpdate rate)
69+
70+ {-# MINIMAL runUpdate, createRandom #-}
71+
72+-- | Class for a layer. All layers implement this, however, they don't
73+-- need to implement it for all shapes, only ones which are
74+-- appropriate.
75+--
76+class UpdateLayer x => Layer x (i :: Shape) (o :: Shape) where
77+ -- | The Wengert tape for this layer. Includes all that is required
78+ -- to generate the back propagated gradients efficiently. As a
79+ -- default, `S i` is fine.
80+ type Tape x i o :: Type
81+
82+ -- | Used in training and scoring. Take the input from the previous
83+ -- layer, and give the output from this layer.
84+ runForwards :: x -> S i -> (Tape x i o, S o)
85+
86+ -- | Back propagate a step. Takes the current layer, the input that
87+ -- the layer gave from the input and the back propagated derivatives
88+ -- from the layer above.
89+ --
90+ -- Returns the gradient layer and the derivatives to push back
91+ -- further.
92+ runBackwards :: x -> Tape x i o -> S o -> (Gradient x, S i)
src/Grenade/Core/LearningParameters.hsnew file+21
@@ -0,0 +1,21 @@
1+{-|
2+Module : Grenade.Core.LearningParameters
3+Description : Stochastic gradient descent learning parameters
4+Copyright : (c) Huw Campbell, 2016-2017
5+License : BSD2
6+Stability : experimental
7+-}
8+module Grenade.Core.LearningParameters (
9+ -- | This module contains learning algorithm specific
10+ -- code. Currently, this module should be considered
11+ -- unstable, due to issue #26.
12+
13+ LearningParameters (..)
14+ ) where
15+
16+-- | Learning parameters for stochastic gradient descent.
17+data LearningParameters = LearningParameters {
18+ learningRate :: Double
19+ , learningMomentum :: Double
20+ , learningRegulariser :: Double
21+ } deriving (Eq, Show)
src/Grenade/Core/Network.hs+155−78
@@ -1,5 +1,8 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
3+{-# LANGUAGE BangPatterns #-}
2{-# LANGUAGE GADTs #-}
4{-# LANGUAGE GADTs #-}
5+{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE TypeOperators #-}
6{-# LANGUAGE TypeOperators #-}
4{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE TypeFamilies #-}
5{-# LANGUAGE MultiParamTypeClasses #-}
8{-# LANGUAGE MultiParamTypeClasses #-}
@@ -7,114 +10,188 @@
7{-# LANGUAGE FlexibleInstances #-}
10{-# LANGUAGE FlexibleInstances #-}
8{-|
11{-|
9Module : Grenade.Core.Network
12Module : Grenade.Core.Network
10−Description : Core definition a simple neural etwork
13+Description : Core definition of a Neural Network
11Copyright : (c) Huw Campbell, 2016-2017
14Copyright : (c) Huw Campbell, 2016-2017
12License : BSD2
15License : BSD2
13Stability : experimental
16Stability : experimental
14
17
15−This module defines the core data type for the simplest
18+This module defines the core data types and functions
16−Neural network we support.
19+for non-recurrent neural networks.
17−
18-}
20-}
21+
19module Grenade.Core.Network (
22module Grenade.Core.Network (
20− Layer (..)
21− , Network (..)
23+ Network (..)
22− , UpdateLayer (..)
23− , LearningParameters (..)
24 , Gradients (..)
24 , Gradients (..)
25− , CreatableNetwork (..)
25+ , Tapes (..)
26+
27+ , runNetwork
28+ , runGradient
29+ , applyUpdate
30+
31+ , randomNetwork
26 ) where
32 ) where
27
33
28−import Control.Monad.Random (MonadRandom)
34+import Control.Monad.Random ( MonadRandom )
29−import Data.List ( foldl' )
30−import Data.Singletons
31
35
36+import Data.Singletons
32import Data.Serialize
37import Data.Serialize
33
38
34−import Grenade.Core.Shape
39+#if MIN_VERSION_base(4,9,0)
40+import Data.Kind (Type)
41+#endif
35
42
36−-- | Learning parameters for stochastic gradient descent.
43+import Grenade.Core.Layer
37−data LearningParameters = LearningParameters {
44+import Grenade.Core.LearningParameters
38− learningRate :: Double
45+import Grenade.Core.Shape
39− , learningMomentum :: Double
46+import Prelude.Singletons
40− , learningRegulariser :: Double
41− } deriving (Eq, Show)
42−
43−-- | Class for updating a layer. All layers implement this, and it is
44−-- shape independent.
45−class Show x => UpdateLayer x where
46− {-# MINIMAL runUpdate, createRandom #-}
47− -- | The type for the gradient for this layer.
48− -- Unit if there isn't a gradient to pass back.
49− type Gradient x :: *
50− -- | Update a layer with its gradient and learning parameters
51− runUpdate :: LearningParameters -> x -> Gradient x -> x
52−
53− -- | Create a random layer, many layers will use pure
54− createRandom :: MonadRandom m => m x
55−
56− -- | Update a layer with many Gradients
57− runUpdates :: LearningParameters -> x -> [Gradient x] -> x
58− runUpdates rate = foldl' (runUpdate rate)
59−
60−-- | Class for a layer. All layers implement this, however, they don't
61−-- need to implement it for all shapes, only ones which are appropriate.
62−class UpdateLayer x => Layer x (i :: Shape) (o :: Shape) where
63− -- | The Wengert tape for this layer. Includes all that is required
64− -- to generate the back propagated gradients efficiently. As a
65− -- default, `S i` is fine.
66− type Tape x i o :: *
67− -- | Used in training and scoring. Take the input from the previous
68− -- layer, and give the output from this layer.
69− runForwards :: x -> S i -> (Tape x i o, S o)
70− -- | Back propagate a step. Takes the current layer, the input that the
71− -- layer gave from the input and the back propagated derivatives from
72− -- the layer above.
73− -- Returns the gradient layer and the derivatives to push back further.
74− runBackwards :: x -> Tape x i o -> S o -> (Gradient x, S i)
75
47
76-- | Type of a network.
48-- | Type of a network.
77--
49--
78−-- The [*] type specifies the types of the layers.
50+-- The @[*]@ type specifies the types of the layers.
79--
51--
80−-- The [Shape] type specifies the shapes of data passed between the layers.
52+-- The @[Shape]@ type specifies the shapes of data passed between the layers.
81--
53--
82-- Can be considered to be a heterogeneous list of layers which are able to
54-- Can be considered to be a heterogeneous list of layers which are able to
83-- transform the data shapes of the network.
55-- transform the data shapes of the network.
84−data Network :: [*] -> [Shape] -> * where
56+data Network :: [Type] -> [Shape] -> Type where
57+ NNil :: SingI i
85− O :: (SingI i, SingI o, Layer x i o) => !x -> Network '[x] '[i, o]
58+ => Network '[] '[i]
59+
86− (:~>) :: (SingI i, SingI h, Layer x i h) => !x -> !(Network xs (h ': hs)) -> Network (x ': xs) (i ': h ': hs)
60+ (:~>) :: (SingI i, SingI h, Layer x i h)
61+ => !x
62+ -> !(Network xs (h ': hs))
63+ -> Network (x ': xs) (i ': h ': hs)
87infixr 5 :~>
64infixr 5 :~>
88
65
89−instance Show (Network l h) where
66+instance Show (Network '[] '[i]) where
90− show (O a) = "O " ++ show a
67+ show NNil = "NNil"
68+instance (Show x, Show (Network xs rs)) => Show (Network (x ': xs) (i ': rs)) where
91− show (i :~> o) = show i ++ "\n:~>\n" ++ show o
69+ show (x :~> xs) = show x ++ "\n~>\n" ++ show xs
70+
71+-- | Gradient of a network.
72+--
73+-- Parameterised on the layers of the network.
74+data Gradients :: [Type] -> Type where
75+ GNil :: Gradients '[]
76+
77+ (:/>) :: UpdateLayer x
78+ => Gradient x
79+ -> Gradients xs
80+ -> Gradients (x ': xs)
92
81
93−-- | Gradients of a network.
82+-- | Wegnert Tape of a network.
83+--
94−-- Parameterised on the layers of a Network.
84+-- Parameterised on the layers and shapes of the network.
95−data Gradients :: [*] -> * where
85+data Tapes :: [Type] -> [Shape] -> Type where
96− OG :: UpdateLayer x => Gradient x -> Gradients '[x]
86+ TNil :: SingI i
97− (:/>) :: UpdateLayer x => Gradient x -> Gradients xs -> Gradients (x ': xs)
87+ => Tapes '[] '[i]
98
88
89+ (:\>) :: (SingI i, SingI h, Layer x i h)
90+ => !(Tape x i h)
91+ -> !(Tapes xs (h ': hs))
92+ -> Tapes (x ': xs) (i ': h ': hs)
93+
94+
95+-- | Running a network forwards with some input data.
96+--
97+-- This gives the output, and the Wengert tape required for back
98+-- propagation.
99+runNetwork :: forall layers shapes.
100+ Network layers shapes
101+ -> S (Head shapes)
102+ -> (Tapes layers shapes, S (Last shapes))
103+runNetwork =
104+ go
105+ where
106+ go :: forall js ss. (Last js ~ Last shapes)
107+ => Network ss js
108+ -> S (Head js)
109+ -> (Tapes ss js, S (Last js))
110+ go (layer :~> n) !x =
111+ let (tape, forward) = runForwards layer x
112+ (tapes, answer) = go n forward
113+ in (tape :\> tapes, answer)
114+
115+ go NNil !x
116+ = (TNil, x)
117+
118+
119+-- | Running a loss gradient back through the network.
120+--
121+-- This requires a Wengert tape, generated with the appropriate input
122+-- for the loss.
123+--
124+-- Gives the gradients for the layer, and the gradient across the
125+-- input (which may not be required).
126+runGradient :: forall layers shapes.
127+ Network layers shapes
128+ -> Tapes layers shapes
129+ -> S (Last shapes)
130+ -> (Gradients layers, S (Head shapes))
131+runGradient net tapes o =
132+ go net tapes
133+ where
134+ go :: forall js ss. (Last js ~ Last shapes)
135+ => Network ss js
136+ -> Tapes ss js
137+ -> (Gradients ss, S (Head js))
138+ go (layer :~> n) (tape :\> nt) =
139+ let (gradients, feed) = go n nt
140+ (layer', backGrad) = runBackwards layer tape feed
141+ in (layer' :/> gradients, backGrad)
142+
143+ go NNil TNil
144+ = (GNil, o)
145+
146+
147+-- | Apply one step of stochastic gradient descent across the network.
148+applyUpdate :: LearningParameters
149+ -> Network layers shapes
150+ -> Gradients layers
151+ -> Network layers shapes
152+applyUpdate rate (layer :~> rest) (gradient :/> grest)
153+ = runUpdate rate layer gradient :~> applyUpdate rate rest grest
154+
155+applyUpdate _ NNil GNil
156+ = NNil
157+
99−-- | A network can easily be created by hand with (:~>), but an easy way to initialise a random
158+-- | A network can easily be created by hand with (:~>), but an easy way to
100−-- network is with the randomNetwork.
159+-- initialise a random network is with the randomNetwork.
101−class CreatableNetwork (xs :: [*]) (ss :: [Shape]) where
160+class CreatableNetwork (xs :: [Type]) (ss :: [Shape]) where
102− -- | Create a network of the types requested
161+ -- | Create a network with randomly initialised weights.
162+ --
163+ -- Calls to this function will not compile if the type of the neural
164+ -- network is not sound.
103 randomNetwork :: MonadRandom m => m (Network xs ss)
165 randomNetwork :: MonadRandom m => m (Network xs ss)
104
166
105−instance (SingI i, SingI o, Layer x i o) => CreatableNetwork (x ': '[]) (i ': o ': '[]) where
167+instance SingI i => CreatableNetwork '[] '[i] where
106− randomNetwork = O <$> createRandom
168+ randomNetwork = return NNil
107
169
108−instance (SingI i, SingI o, Layer x i o, CreatableNetwork xs (o ': r ': rs)) => CreatableNetwork (x ': xs) (i ': o ': r ': rs) where
170+instance (SingI i, SingI o, Layer x i o, CreatableNetwork xs (o ': rs)) => CreatableNetwork (x ': xs) (i ': o ': rs) where
109 randomNetwork = (:~>) <$> createRandom <*> randomNetwork
171 randomNetwork = (:~>) <$> createRandom <*> randomNetwork
110
172
111−
112-- | Add very simple serialisation to the network
173-- | Add very simple serialisation to the network
113−instance (SingI i, SingI o, Layer x i o, Serialize x) => Serialize (Network '[x] '[i, o]) where
174+instance SingI i => Serialize (Network '[] '[i]) where
114− put (O x) = put x
175+ put NNil = pure ()
115− put _ = error "impossible"
176+ get = return NNil
116− get = O <$> get
117
177
118−instance (SingI i, SingI o, Layer x i o, Serialize x, Serialize (Network xs (o ': r ': rs))) => Serialize (Network (x ': xs) (i ': o ': r ': rs)) where
178+instance (SingI i, SingI o, Layer x i o, Serialize x, Serialize (Network xs (o ': rs))) => Serialize (Network (x ': xs) (i ': o ': rs)) where
119 put (x :~> r) = put x >> put r
179 put (x :~> r) = put x >> put r
120 get = (:~>) <$> get <*> get
180 get = (:~>) <$> get <*> get
181+
182+
183+-- | Ultimate composition.
184+--
185+-- This allows a complete network to be treated as a layer in a larger network.
186+instance CreatableNetwork sublayers subshapes => UpdateLayer (Network sublayers subshapes) where
187+ type Gradient (Network sublayers subshapes) = Gradients sublayers
188+ runUpdate = applyUpdate
189+ createRandom = randomNetwork
190+
191+-- | Ultimate composition.
192+--
193+-- This allows a complete network to be treated as a layer in a larger network.
194+instance (CreatableNetwork sublayers subshapes, i ~ (Head subshapes), o ~ (Last subshapes)) => Layer (Network sublayers subshapes) i o where
195+ type Tape (Network sublayers subshapes) i o = Tapes sublayers subshapes
196+ runForwards = runNetwork
197+ runBackwards = runGradient
src/Grenade/Core/Runner.hs+28−50
@@ -1,80 +1,58 @@
1−{-# LANGUAGE BangPatterns #-}
2−{-# LANGUAGE GADTs #-}
3{-# LANGUAGE DataKinds #-}
1{-# LANGUAGE DataKinds #-}
4−{-# LANGUAGE ScopedTypeVariables #-}
2+{-# LANGUAGE FlexibleContexts #-}
5−{-# LANGUAGE TypeOperators #-}
6−{-# LANGUAGE TypeFamilies #-}
7{-|
3{-|
8−Module : Grenade.Core.Shape
4+Module : Grenade.Core.Runner
9−Description : Core definition of the Shapes of data we understand
5+Description : Functions to perform training and backpropagation
10Copyright : (c) Huw Campbell, 2016-2017
6Copyright : (c) Huw Campbell, 2016-2017
11License : BSD2
7License : BSD2
12Stability : experimental
8Stability : experimental
13−
14−This module defines simple back propagation and training functions
15−for a network.
16-}
9-}
17module Grenade.Core.Runner (
10module Grenade.Core.Runner (
18 train
11 train
19 , backPropagate
12 , backPropagate
20 , runNet
13 , runNet
21− , applyUpdate
22 ) where
14 ) where
23
15
24−import Data.Singletons.Prelude
16+
17+import Grenade.Core.LearningParameters
25import Grenade.Core.Network
18import Grenade.Core.Network
26import Grenade.Core.Shape
19import Grenade.Core.Shape
20+import Data.Singletons
21+import Prelude.Singletons
27
22
28-- | Perform reverse automatic differentiation on the network
23-- | Perform reverse automatic differentiation on the network
29-- for the current input and expected output.
24-- for the current input and expected output.
30--
25--
31-- /Note:/ The loss function pushed backwards is appropriate
26-- /Note:/ The loss function pushed backwards is appropriate
32-- for both regression and classification as a squared loss
27-- for both regression and classification as a squared loss
33−-- or log-loss respectively. Other loss functions are not yet
28+-- or log-loss respectively.
34−-- implemented.
29+--
30+-- For other loss functions, use runNetwork and runGradient
31+-- with the back propagated gradient of your loss.
32+--
35−backPropagate :: forall shapes layers.
33+backPropagate :: SingI (Last shapes)
36− Network layers shapes -> S (Head shapes) -> S (Last shapes) -> Gradients layers
34+ => Network layers shapes
35+ -> S (Head shapes)
36+ -> S (Last shapes)
37+ -> Gradients layers
37backPropagate network input target =
38backPropagate network input target =
38− fst $ go input network
39+ let (tapes, output) = runNetwork network input
39− where
40+ (grads, _) = runGradient network tapes (output - target)
40− go :: forall js sublayers. (Last js ~ Last shapes)
41+ in grads
41− => S (Head js) -- ^ input vector
42− -> Network sublayers js -- ^ network to train
43− -> (Gradients sublayers, S (Head js))
44− -- handle input from the beginning, feeding upwards.
45− go !x (layer :~> n)
46− = let (tape, y) = runForwards layer x
47− -- recursively run the rest of the network, and get the gradients from above.
48− (n', dWs') = go y n
49− -- calculate the gradient for this layer to pass down,
50− (layer', dWs) = runBackwards layer tape dWs'
51
42
52− in (layer' :/> n', dWs)
53−
54− -- handle the output layer, bouncing the derivatives back down.
55− go !x (O layer)
56− = let (tape, y) = runForwards layer x
57− -- the gradient (how much y affects the error)
58− (layer', dWs) = runBackwards layer tape (y - target)
59−
60− in (OG layer', dWs)
61−
62−-- | Apply one step of stochastic gradient decent across the network.
63−applyUpdate :: LearningParameters -> Network ls ss -> Gradients ls -> Network ls ss
64−applyUpdate rate (O layer) (OG gradient)
65− = O (runUpdate rate layer gradient)
66−applyUpdate rate (layer :~> rest) (gradient :/> grest)
67− = runUpdate rate layer gradient :~> applyUpdate rate rest grest
68−applyUpdate _ _ _
69− = error "Impossible for the gradients of a network to have a different length to the network"
70
43
71-- | Update a network with new weights after training with an instance.
44-- | Update a network with new weights after training with an instance.
72−train :: LearningParameters -> Network layers shapes -> S (Head shapes) -> S (Last shapes) -> Network layers shapes
45+train :: SingI (Last shapes)
46+ => LearningParameters
47+ -> Network layers shapes
48+ -> S (Head shapes)
49+ -> S (Last shapes)
50+ -> Network layers shapes
73train rate network input output =
51train rate network input output =
74 let grads = backPropagate network input output
52 let grads = backPropagate network input output
75 in applyUpdate rate network grads
53 in applyUpdate rate network grads
76
54
55+
77-- | Run the network with input and return the given output.
56-- | Run the network with input and return the given output.
78runNet :: Network layers shapes -> S (Head shapes) -> S (Last shapes)
57runNet :: Network layers shapes -> S (Head shapes) -> S (Last shapes)
79−runNet (layer :~> n) !x = let (_, y) = runForwards layer x in runNet n y
58+runNet net = snd . runNetwork net
80−runNet (O layer) !x = snd (runForwards layer x)
src/Grenade/Core/Shape.hs+80−33
@@ -1,3 +1,4 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE GADTs #-}
3{-# LANGUAGE GADTs #-}
3{-# LANGUAGE KindSignatures #-}
4{-# LANGUAGE KindSignatures #-}
@@ -7,39 +8,33 @@
7{-# LANGUAGE FlexibleContexts #-}
8{-# LANGUAGE FlexibleContexts #-}
8{-# LANGUAGE ScopedTypeVariables #-}
9{-# LANGUAGE ScopedTypeVariables #-}
9{-# LANGUAGE RankNTypes #-}
10{-# LANGUAGE RankNTypes #-}
10−
11+{-# LANGUAGE UndecidableInstances #-}
11−-- Ghc 8.0 gives a warning on `n2 _ _ = error ...` but ghc 7.10 fails to
12−-- compile without this default pattern.
13−{-# OPTIONS_GHC -fno-warn-overlapping-patterns #-}
14−
15{-|
12{-|
16Module : Grenade.Core.Shape
13Module : Grenade.Core.Shape
17−Description : Core definition of the Shapes of data we understand
14+Description : Dependently typed shapes of data which are passed between layers of a network
18Copyright : (c) Huw Campbell, 2016-2017
15Copyright : (c) Huw Campbell, 2016-2017
19License : BSD2
16License : BSD2
20Stability : experimental
17Stability : experimental
21
18
22−This module defines the core data types for the shapes of data that
19+
23−are understood by Grenade.
24-}
20-}
25module Grenade.Core.Shape (
21module Grenade.Core.Shape (
22+ S (..)
26− Shape (..)
23+ , Shape (..)
27− , S (..)
24+ , SShape (..)
28 , randomOfShape
25 , randomOfShape
29 , fromStorable
26 , fromStorable
30 ) where
27 ) where
31
28
32import Control.DeepSeq (NFData (..))
29import Control.DeepSeq (NFData (..))
33import Control.Monad.Random ( MonadRandom, getRandom )
30import Control.Monad.Random ( MonadRandom, getRandom )
34−
31+import Data.Kind (Type)
35import Data.Proxy
32import Data.Proxy
33+import Data.Serialize
36import Data.Singletons
34import Data.Singletons
37−import Data.Singletons.TypeLits
38import Data.Vector.Storable ( Vector )
35import Data.Vector.Storable ( Vector )
39import qualified Data.Vector.Storable as V
36import qualified Data.Vector.Storable as V
40−
41import GHC.TypeLits
37import GHC.TypeLits
42−
43import qualified Numeric.LinearAlgebra.Static as H
38import qualified Numeric.LinearAlgebra.Static as H
44import Numeric.LinearAlgebra.Static
39import Numeric.LinearAlgebra.Static
45import qualified Numeric.LinearAlgebra as NLA
40import qualified Numeric.LinearAlgebra as NLA
@@ -47,31 +42,57 @@ import qualified Numeric.LinearAlgebra as NLA
47-- | The current shapes we accept.
42-- | The current shapes we accept.
48-- at the moment this is just one, two, and three dimensional
43-- at the moment this is just one, two, and three dimensional
49-- Vectors/Matricies.
44-- Vectors/Matricies.
45+--
46+-- These are only used with DataKinds, as Kind `Shape`, with Types 'D1, 'D2, 'D3.
50data Shape
47data Shape
51 = D1 Nat
48 = D1 Nat
49+ -- ^ One dimensional vector
52 | D2 Nat Nat
50 | D2 Nat Nat
51+ -- ^ Two dimensional matrix. Row, Column.
53 | D3 Nat Nat Nat
52 | D3 Nat Nat Nat
53+ -- ^ Three dimensional matrix. Row, Column, Channels.
54
54
55−-- | Given a Shape n, these are the possible data structures with that shape.
55+-- | Concrete data structures for a Shape.
56+--
56-- All shapes are held in contiguous memory.
57-- All shapes are held in contiguous memory.
57-- 3D is held in a matrix (usually row oriented) which has height depth * rows.
58-- 3D is held in a matrix (usually row oriented) which has height depth * rows.
58data S (n :: Shape) where
59data S (n :: Shape) where
59− S1D :: ( KnownNat o ) => R o -> S ('D1 o)
60+ S1D :: ( KnownNat len )
61+ => R len
62+ -> S ('D1 len)
63+
60− S2D :: ( KnownNat rows, KnownNat columns ) => L rows columns -> S ('D2 rows columns)
64+ S2D :: ( KnownNat rows, KnownNat columns )
65+ => L rows columns
66+ -> S ('D2 rows columns)
67+
61 S3D :: ( KnownNat rows
68 S3D :: ( KnownNat rows
62 , KnownNat columns
69 , KnownNat columns
63 , KnownNat depth
70 , KnownNat depth
71+ , KnownNat (rows * depth))
72+ => L (rows * depth) columns
64− , KnownNat (rows * depth)) => L (rows * depth) columns -> S ('D3 rows columns depth)
73+ -> S ('D3 rows columns depth)
65
74
66deriving instance Show (S n)
75deriving instance Show (S n)
67
76
68−-- Singletons
77+-- Singleton instances.
78+--
69-- These could probably be derived with template haskell, but this seems
79-- These could probably be derived with template haskell, but this seems
70-- clear and makes adding the KnownNat constraints simple.
80-- clear and makes adding the KnownNat constraints simple.
81+-- We can also keep our code TH free, which is great.
82+#if MIN_VERSION_singletons(2,6,0)
83+-- In singletons 2.6 Sing switched from a data family to a type family.
84+type instance Sing = SShape
85+
86+data SShape :: Shape -> Type where
87+ D1Sing :: KnownNat a => SShape ('D1 a)
88+ D2Sing :: (KnownNat a, KnownNat b) => SShape ('D2 a b)
89+ D3Sing :: (KnownNat (a * c), KnownNat a, KnownNat b, KnownNat c) => SShape ('D3 a b c)
90+#else
71data instance Sing (n :: Shape) where
91data instance Sing (n :: Shape) where
72− D1Sing :: KnownNat a => Sing ('D1 a)
92+ D1Sing :: Sing a -> Sing ('D1 a)
73− D2Sing :: (KnownNat a, KnownNat b) => Sing ('D2 a b)
93+ D2Sing :: Sing a -> Sing b -> Sing ('D2 a b)
74− D3Sing :: (KnownNat a, KnownNat b, KnownNat c, KnownNat (a * c)) => Sing ('D3 a b c)
94+ D3Sing :: KnownNat (a * c) => Sing a -> Sing b -> Sing c -> Sing ('D3 a b c)
95+#endif
75
96
76instance KnownNat a => SingI ('D1 a) where
97instance KnownNat a => SingI ('D1 a) where
77 sing = D1Sing
98 sing = D1Sing
@@ -127,18 +148,28 @@ randomOfShape :: forall x m. ( MonadRandom m, SingI x ) => m (S x)
127randomOfShape = do
148randomOfShape = do
128 seed :: Int <- getRandom
149 seed :: Int <- getRandom
129 return $ case (sing :: Sing x) of
150 return $ case (sing :: Sing x) of
151+ D1Sing ->
130− D1Sing -> S1D (randomVector seed Uniform * 2 - 1)
152+ S1D (randomVector seed Uniform * 2 - 1)
153+
154+ D2Sing ->
131− D2Sing -> S2D (uniformSample seed (-1) 1)
155+ S2D (uniformSample seed (-1) 1)
156+
157+ D3Sing ->
132− D3Sing -> S3D (uniformSample seed (-1) 1)
158+ S3D (uniformSample seed (-1) 1)
133
159
134-- | Generate a shape from a Storable Vector.
160-- | Generate a shape from a Storable Vector.
135--
161--
136-- Returns Nothing if the vector is of the wrong size.
162-- Returns Nothing if the vector is of the wrong size.
137fromStorable :: forall x. SingI x => Vector Double -> Maybe (S x)
163fromStorable :: forall x. SingI x => Vector Double -> Maybe (S x)
138fromStorable xs = case sing :: Sing x of
164fromStorable xs = case sing :: Sing x of
165+ D1Sing ->
139− D1Sing -> S1D <$> H.create xs
166+ S1D <$> H.create xs
167+
168+ D2Sing ->
140− D2Sing -> S2D <$> mkL xs
169+ S2D <$> mkL xs
170+
171+ D3Sing ->
141− D3Sing -> S3D <$> mkL xs
172+ S3D <$> mkL xs
142 where
173 where
143 mkL :: forall rows columns. (KnownNat rows, KnownNat columns)
174 mkL :: forall rows columns. (KnownNat rows, KnownNat columns)
144 => Vector Double -> Maybe (L rows columns)
175 => Vector Double -> Maybe (L rows columns)
@@ -149,6 +180,18 @@ fromStorable xs = case sing :: Sing x of
149 then H.create $ NLA.reshape columns v
180 then H.create $ NLA.reshape columns v
150 else Nothing
181 else Nothing
151
182
183+
184+instance SingI x => Serialize (S x) where
185+ put i = (case i of
186+ (S1D x) -> putListOf put . NLA.toList . H.extract $ x
187+ (S2D x) -> putListOf put . NLA.toList . NLA.flatten . H.extract $ x
188+ (S3D x) -> putListOf put . NLA.toList . NLA.flatten . H.extract $ x
189+ ) :: PutM ()
190+
191+ get = do
192+ Just i <- fromStorable . V.fromList <$> getListOf get
193+ return i
194+
152-- Helper function for creating the number instances
195-- Helper function for creating the number instances
153n1 :: ( forall a. Floating a => a -> a ) -> S x -> S x
196n1 :: ( forall a. Floating a => a -> a ) -> S x -> S x
154n1 f (S1D x) = S1D (f x)
197n1 f (S1D x) = S1D (f x)
@@ -160,11 +203,15 @@ n2 :: ( forall a. Floating a => a -> a -> a ) -> S x -> S x -> S x
160n2 f (S1D x) (S1D y) = S1D (f x y)
203n2 f (S1D x) (S1D y) = S1D (f x y)
161n2 f (S2D x) (S2D y) = S2D (f x y)
204n2 f (S2D x) (S2D y) = S2D (f x y)
162n2 f (S3D x) (S3D y) = S3D (f x y)
205n2 f (S3D x) (S3D y) = S3D (f x y)
163−n2 _ _ _ = error "Impossible to have different constructors for the same shaped network"
164
206
165-- Helper function for creating the number instances
207-- Helper function for creating the number instances
166−nk :: forall x. SingI x => Double -> S x
208+nk :: forall x. (SingI x) => Double -> S x
167nk x = case (sing :: Sing x) of
209nk x = case (sing :: Sing x) of
210+ D1Sing ->
168− D1Sing -> S1D (konst x)
211+ S1D (konst x)
212+
213+ D2Sing ->
169− D2Sing -> S2D (konst x)
214+ S2D (konst x)
215+
216+ D3Sing ->
170− D3Sing -> S3D (konst x)
217+ S3D (konst x)
src/Grenade/Layers.hsnew file+35
@@ -0,0 +1,35 @@
1+module Grenade.Layers (
2+ module Grenade.Layers.Concat
3+ , module Grenade.Layers.Convolution
4+ , module Grenade.Layers.Crop
5+ , module Grenade.Layers.Deconvolution
6+ , module Grenade.Layers.Elu
7+ , module Grenade.Layers.FullyConnected
8+ , module Grenade.Layers.Inception
9+ , module Grenade.Layers.Logit
10+ , module Grenade.Layers.Merge
11+ , module Grenade.Layers.Pad
12+ , module Grenade.Layers.Pooling
13+ , module Grenade.Layers.Reshape
14+ , module Grenade.Layers.Relu
15+ , module Grenade.Layers.Softmax
16+ , module Grenade.Layers.Tanh
17+ , module Grenade.Layers.Trivial
18+ ) where
19+
20+import Grenade.Layers.Concat
21+import Grenade.Layers.Convolution
22+import Grenade.Layers.Crop
23+import Grenade.Layers.Deconvolution
24+import Grenade.Layers.Elu
25+import Grenade.Layers.Pad
26+import Grenade.Layers.FullyConnected
27+import Grenade.Layers.Inception
28+import Grenade.Layers.Logit
29+import Grenade.Layers.Merge
30+import Grenade.Layers.Pooling
31+import Grenade.Layers.Reshape
32+import Grenade.Layers.Relu
33+import Grenade.Layers.Softmax
34+import Grenade.Layers.Tanh
35+import Grenade.Layers.Trivial
src/Grenade/Layers/Concat.hsnew file+139
@@ -0,0 +1,139 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE DataKinds #-}
3+{-# LANGUAGE GADTs #-}
4+{-# LANGUAGE TypeOperators #-}
5+{-# LANGUAGE TypeFamilies #-}
6+{-# LANGUAGE MultiParamTypeClasses #-}
7+{-# LANGUAGE FlexibleContexts #-}
8+{-# LANGUAGE RankNTypes #-}
9+{-# LANGUAGE FlexibleInstances #-}
10+{-# LANGUAGE ScopedTypeVariables #-}
11+{-# LANGUAGE StandaloneDeriving #-}
12+{-# LANGUAGE UndecidableInstances #-}
13+{-|
14+Module : Grenade.Layers.Concat
15+Description : Concatenation layer
16+Copyright : (c) Huw Campbell, 2016-2017
17+License : BSD2
18+Stability : experimental
19+
20+This module provides the concatenation layer, which runs two chilld layers in parallel and combines their outputs.
21+-}
22+module Grenade.Layers.Concat (
23+ Concat (..)
24+ ) where
25+
26+import Data.Serialize
27+
28+import Data.Singletons
29+import GHC.TypeLits
30+
31+#if MIN_VERSION_base(4,9,0)
32+import Data.Kind (Type)
33+#endif
34+
35+import Grenade.Core
36+
37+import Numeric.LinearAlgebra.Static ( row, (===), splitRows, unrow, (#), split, R )
38+
39+-- | A Concatentating Layer.
40+--
41+-- This layer shares it's input state between two sublayers, and concatenates their output.
42+--
43+-- With Networks able to be Layers, this allows for very expressive composition of complex Networks.
44+--
45+-- The Concat layer has a few instances, which allow one to flexibly "bash" together the outputs.
46+--
47+-- Two 1D vectors, can go to a 2D shape with 2 rows if their lengths are identical.
48+-- Any 2 1D vectors can also become a longer 1D Vector.
49+--
50+-- 3D images become 3D images with more channels. The sizes must be the same, one can use Pad
51+-- and Crop layers to ensure this is the case.
52+data Concat :: Shape -> Type -> Shape -> Type -> Type where
53+ Concat :: x -> y -> Concat m x n y
54+
55+instance (Show x, Show y) => Show (Concat m x n y) where
56+ show (Concat x y) = "Concat\n" ++ show x ++ "\n" ++ show y
57+
58+-- | Run two layers in parallel, combining their outputs.
59+instance (UpdateLayer x, UpdateLayer y) => UpdateLayer (Concat m x n y) where
60+ type Gradient (Concat m x n y) = (Gradient x, Gradient y)
61+ runUpdate lr (Concat x y) (x', y') = Concat (runUpdate lr x x') (runUpdate lr y y')
62+ createRandom = Concat <$> createRandom <*> createRandom
63+
64+instance ( SingI i
65+ , Layer x i ('D1 o)
66+ , Layer y i ('D1 o)
67+ ) => Layer (Concat ('D1 o) x ('D1 o) y) i ('D2 2 o) where
68+ type Tape (Concat ('D1 o) x ('D1 o) y) i ('D2 2 o) = (Tape x i ('D1 o), Tape y i ('D1 o))
69+
70+ runForwards (Concat x y) input =
71+ let (xT, xOut :: S ('D1 o)) = runForwards x input
72+ (yT, yOut :: S ('D1 o)) = runForwards y input
73+ in case (xOut, yOut) of
74+ (S1D xOut', S1D yOut') ->
75+ ((xT, yT), S2D (row xOut' === row yOut'))
76+
77+ runBackwards (Concat x y) (xTape, yTape) (S2D o) =
78+ let (ox, oy) = splitRows o
79+ (x', xB :: S i) = runBackwards x xTape (S1D $ unrow ox)
80+ (y', yB :: S i) = runBackwards y yTape (S1D $ unrow oy)
81+ in ((x', y'), xB + yB)
82+
83+instance ( SingI i
84+ , Layer x i ('D1 m)
85+ , Layer y i ('D1 n)
86+ , KnownNat o
87+ , KnownNat m
88+ , KnownNat n
89+ , o ~ (m + n)
90+ , n ~ (o - m)
91+ , (m <=? o) ~ 'True
92+ ) => Layer (Concat ('D1 m) x ('D1 n) y) i ('D1 o) where
93+ type Tape (Concat ('D1 m) x ('D1 n) y) i ('D1 o) = (Tape x i ('D1 m), Tape y i ('D1 n))
94+
95+ runForwards (Concat x y) input =
96+ let (xT, xOut :: S ('D1 m)) = runForwards x input
97+ (yT, yOut :: S ('D1 n)) = runForwards y input
98+ in case (xOut, yOut) of
99+ (S1D xOut', S1D yOut') ->
100+ ((xT, yT), S1D (xOut' # yOut'))
101+
102+ runBackwards (Concat x y) (xTape, yTape) (S1D o) =
103+ let (ox :: R m , oy :: R n) = split o
104+ (x', xB :: S i) = runBackwards x xTape (S1D ox)
105+ (y', yB :: S i) = runBackwards y yTape (S1D oy)
106+ in ((x', y'), xB + yB)
107+
108+-- | Concat 3D shapes, increasing the number of channels.
109+instance ( SingI i
110+ , Layer x i ('D3 rows cols m)
111+ , Layer y i ('D3 rows cols n)
112+ , KnownNat (rows * n)
113+ , KnownNat (rows * m)
114+ , KnownNat (rows * o)
115+ , KnownNat o
116+ , KnownNat m
117+ , KnownNat n
118+ , ((rows * m) + (rows * n)) ~ (rows * o)
119+ , ((rows * o) - (rows * m)) ~ (rows * n)
120+ , ((rows * m) <=? (rows * o)) ~ 'True
121+ ) => Layer (Concat ('D3 rows cols m) x ('D3 rows cols n) y) i ('D3 rows cols o) where
122+ type Tape (Concat ('D3 rows cols m) x ('D3 rows cols n) y) i ('D3 rows cols o) = (Tape x i ('D3 rows cols m), Tape y i ('D3 rows cols n))
123+
124+ runForwards (Concat x y) input =
125+ let (xT, xOut :: S ('D3 rows cols m)) = runForwards x input
126+ (yT, yOut :: S ('D3 rows cols n)) = runForwards y input
127+ in case (xOut, yOut) of
128+ (S3D xOut', S3D yOut') ->
129+ ((xT, yT), S3D (xOut' === yOut'))
130+
131+ runBackwards (Concat x y) (xTape, yTape) (S3D o) =
132+ let (ox, oy) = splitRows o
133+ (x', xB :: S i) = runBackwards x xTape (S3D ox :: S ('D3 rows cols m))
134+ (y', yB :: S i) = runBackwards y yTape (S3D oy :: S ('D3 rows cols n))
135+ in ((x', y'), xB + yB)
136+
137+instance (Serialize a, Serialize b) => Serialize (Concat sa a sb b) where
138+ put (Concat a b) = put a *> put b
139+ get = Concat <$> get <*> get
src/Grenade/Layers/Convolution.hs+90−64
@@ -1,12 +1,23 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE ScopedTypeVariables #-}
3−{-# LANGUAGE RecordWildCards #-}
4{-# LANGUAGE GADTs #-}
4{-# LANGUAGE GADTs #-}
5{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeOperators #-}
6{-# LANGUAGE TypeFamilies #-}
6{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE MultiParamTypeClasses #-}
7{-# LANGUAGE MultiParamTypeClasses #-}
8{-# LANGUAGE FlexibleInstances #-}
8{-# LANGUAGE FlexibleInstances #-}
9{-# LANGUAGE FlexibleContexts #-}
9{-# LANGUAGE FlexibleContexts #-}
10+{-# LANGUAGE UndecidableInstances #-}
11+{-|
12+Module : Grenade.Layers.Convolution
13+Description : Convolution layer
14+Copyright : (c) Huw Campbell, 2016-2017
15+License : BSD2
16+Stability : experimental
17+
18+This module provides the Convolution layer, which is critical in many computer vision tasks.
19+
20+-}
10module Grenade.Layers.Convolution (
21module Grenade.Layers.Convolution (
11 Convolution (..)
22 Convolution (..)
12 , Convolution' (..)
23 , Convolution' (..)
@@ -17,16 +28,15 @@ import Control.Monad.Random hiding ( fromList )
17import Data.Maybe
28import Data.Maybe
18import Data.Proxy
29import Data.Proxy
19import Data.Serialize
30import Data.Serialize
20−import Data.Singletons.TypeLits
21
31
22import GHC.TypeLits
32import GHC.TypeLits
33+import Data.Kind (Type)
23
34
24import Numeric.LinearAlgebra hiding ( uniformSample, konst )
35import Numeric.LinearAlgebra hiding ( uniformSample, konst )
25import qualified Numeric.LinearAlgebra as LA
36import qualified Numeric.LinearAlgebra as LA
26import Numeric.LinearAlgebra.Static hiding ((|||), build, toRows)
37import Numeric.LinearAlgebra.Static hiding ((|||), build, toRows)
27
38
28−import Grenade.Core.Network
39+import Grenade.Core
29−import Grenade.Core.Shape
30import Grenade.Layers.Internal.Convolution
40import Grenade.Layers.Internal.Convolution
31import Grenade.Layers.Internal.Update
41import Grenade.Layers.Internal.Update
32
42
@@ -47,7 +57,7 @@ data Convolution :: Nat -- Number of channels, for the first layer this could be
47 -> Nat -- The number of column in the kernel filter
57 -> Nat -- The number of column in the kernel filter
48 -> Nat -- The row stride of the convolution filter
58 -> Nat -- The row stride of the convolution filter
49 -> Nat -- The columns stride of the convolution filter
59 -> Nat -- The columns stride of the convolution filter
50− -> * where
60+ -> Type where
51 Convolution :: ( KnownNat channels
61 Convolution :: ( KnownNat channels
52 , KnownNat filters
62 , KnownNat filters
53 , KnownNat kernelRows
63 , KnownNat kernelRows
@@ -66,7 +76,7 @@ data Convolution' :: Nat -- Number of channels, for the first layer this could b
66 -> Nat -- The number of column in the kernel filter
76 -> Nat -- The number of column in the kernel filter
67 -> Nat -- The row stride of the convolution filter
77 -> Nat -- The row stride of the convolution filter
68 -> Nat -- The columns stride of the convolution filter
78 -> Nat -- The columns stride of the convolution filter
69− -> * where
79+ -> Type where
70 Convolution' :: ( KnownNat channels
80 Convolution' :: ( KnownNat channels
71 , KnownNat filters
81 , KnownNat filters
72 , KnownNat kernelRows
82 , KnownNat kernelRows
@@ -120,9 +130,9 @@ instance ( KnownNat channels
120 , KnownNat strideColumns
130 , KnownNat strideColumns
121 , KnownNat (kernelRows * kernelColumns * channels)
131 , KnownNat (kernelRows * kernelColumns * channels)
122 ) => UpdateLayer (Convolution channels filters kernelRows kernelColumns strideRows strideColumns) where
132 ) => UpdateLayer (Convolution channels filters kernelRows kernelColumns strideRows strideColumns) where
123− type Gradient (Convolution channels filters kernelRows kernelCols strideRows strideCols) = (Convolution' channels filters kernelRows kernelCols strideRows strideCols)
133+ type Gradient (Convolution channels filters kernelRows kernelColumns strideRows strideColumns) = (Convolution' channels filters kernelRows kernelColumns strideRows strideColumns)
124− runUpdate LearningParameters {..} (Convolution oldKernel oldMomentum) (Convolution' kernelGradient) =
134+ runUpdate lp (Convolution oldKernel oldMomentum) (Convolution' kernelGradient) =
125− let (newKernel, newMomentum) = decendMatrix learningRate learningMomentum learningRegulariser oldKernel kernelGradient oldMomentum
135+ let (newKernel, newMomentum) = descendMatrix (learningRate lp) (learningMomentum lp) (learningRegulariser lp) oldKernel kernelGradient oldMomentum
126 in Convolution newKernel newMomentum
136 in Convolution newKernel newMomentum
127
137
128 createRandom = randomConvolution
138 createRandom = randomConvolution
@@ -142,7 +152,8 @@ instance ( KnownNat channels
142 let mm = konst 0
152 let mm = konst 0
143 return $ Convolution wN mm
153 return $ Convolution wN mm
144
154
145−-- | A two dimentional image may have a convolution filter applied to it
155+-- | A three dimensional image (or 2d with many channels) can have
156+-- an appropriately sized convolution filter run across it.
146instance ( KnownNat kernelRows
157instance ( KnownNat kernelRows
147 , KnownNat kernelCols
158 , KnownNat kernelCols
148 , KnownNat filters
159 , KnownNat filters
@@ -152,28 +163,33 @@ instance ( KnownNat kernelRows
152 , KnownNat inputCols
163 , KnownNat inputCols
153 , KnownNat outputRows
164 , KnownNat outputRows
154 , KnownNat outputCols
165 , KnownNat outputCols
166+ , KnownNat channels
155 , ((outputRows - 1) * strideRows) ~ (inputRows - kernelRows)
167 , ((outputRows - 1) * strideRows) ~ (inputRows - kernelRows)
156 , ((outputCols - 1) * strideCols) ~ (inputCols - kernelCols)
168 , ((outputCols - 1) * strideCols) ~ (inputCols - kernelCols)
157− , KnownNat (kernelRows * kernelCols * 1)
169+ , KnownNat (kernelRows * kernelCols * channels)
158 , KnownNat (outputRows * filters)
170 , KnownNat (outputRows * filters)
159− ) => Layer (Convolution 1 filters kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D3 outputRows outputCols filters) where
171+ ) => Layer (Convolution channels filters kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D3 outputRows outputCols filters) where
172+
160− type Tape (Convolution 1 filters kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D3 outputRows outputCols filters) = S ('D2 inputRows inputCols)
173+ type Tape (Convolution channels filters kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D3 outputRows outputCols filters) = S ('D3 inputRows inputCols channels)
174+
161− runForwards (Convolution kernel _) (S2D input) =
175+ runForwards (Convolution kernel _) (S3D input) =
162 let ex = extract input
176 let ex = extract input
163 ek = extract kernel
177 ek = extract kernel
178+ ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)
179+ iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)
164 kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
180 kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
165 ky = fromIntegral $ natVal (Proxy :: Proxy kernelCols)
181 ky = fromIntegral $ natVal (Proxy :: Proxy kernelCols)
166 sx = fromIntegral $ natVal (Proxy :: Proxy strideRows)
182 sx = fromIntegral $ natVal (Proxy :: Proxy strideRows)
167 sy = fromIntegral $ natVal (Proxy :: Proxy strideCols)
183 sy = fromIntegral $ natVal (Proxy :: Proxy strideCols)
168 ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
184 ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
169 oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
185 oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
186+
170− c = im2col kx ky sx sy ex
187+ c = vid2col kx ky sx sy ix iy ex
171 mt = c LA.<> ek
188 mt = c LA.<> ek
172 r = col2vid 1 1 1 1 ox oy mt
189 r = col2vid 1 1 1 1 ox oy mt
173 rs = fromJust . create $ r
190 rs = fromJust . create $ r
174− in (S2D input, S3D rs)
191+ in (S3D input, S3D rs)
175−
176− runBackwards (Convolution kernel _) (S2D input) (S3D dEdy) =
192+ runBackwards (Convolution kernel _) (S3D input) (S3D dEdy) =
177 let ex = extract input
193 let ex = extract input
178 ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)
194 ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)
179 iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)
195 iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)
@@ -184,7 +200,7 @@ instance ( KnownNat kernelRows
184 ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
200 ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
185 oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
201 oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
186
202
187− c = im2col kx ky sx sy ex
203+ c = vid2col kx ky sx sy ix iy ex
188
204
189 eo = extract dEdy
205 eo = extract dEdy
190 ek = extract kernel
206 ek = extract kernel
@@ -192,14 +208,14 @@ instance ( KnownNat kernelRows
192 vs = vid2col 1 1 1 1 ox oy eo
208 vs = vid2col 1 1 1 1 ox oy eo
193
209
194 kN = fromJust . create $ tr c LA.<> vs
210 kN = fromJust . create $ tr c LA.<> vs
211+
195 dW = vs LA.<> tr ek
212 dW = vs LA.<> tr ek
196
213
197− xW = col2im kx ky sx sy ix iy dW
214+ xW = col2vid kx ky sx sy ix iy dW
198− in (Convolution' kN, S2D . fromJust . create $ xW)
215+ in (Convolution' kN, S3D . fromJust . create $ xW)
199
216
200
217
201−-- | A three dimensional image (or 2d with many channels) can have
218+-- | A two dimentional image may have a convolution filter applied to it
202−-- an appropriately sized convolution filter run across it.
203instance ( KnownNat kernelRows
219instance ( KnownNat kernelRows
204 , KnownNat kernelCols
220 , KnownNat kernelCols
205 , KnownNat filters
221 , KnownNat filters
@@ -209,53 +225,63 @@ instance ( KnownNat kernelRows
209 , KnownNat inputCols
225 , KnownNat inputCols
210 , KnownNat outputRows
226 , KnownNat outputRows
211 , KnownNat outputCols
227 , KnownNat outputCols
212− , KnownNat channels
213 , ((outputRows - 1) * strideRows) ~ (inputRows - kernelRows)
228 , ((outputRows - 1) * strideRows) ~ (inputRows - kernelRows)
214 , ((outputCols - 1) * strideCols) ~ (inputCols - kernelCols)
229 , ((outputCols - 1) * strideCols) ~ (inputCols - kernelCols)
215− , KnownNat (kernelRows * kernelCols * channels)
230+ , KnownNat (kernelRows * kernelCols * 1)
216 , KnownNat (outputRows * filters)
231 , KnownNat (outputRows * filters)
217− ) => Layer (Convolution channels filters kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D3 outputRows outputCols filters) where
232+ ) => Layer (Convolution 1 filters kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D3 outputRows outputCols filters) where
218−
219− type Tape (Convolution channels filters kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D3 outputRows outputCols filters) = S ('D3 inputRows inputCols channels)
233+ type Tape (Convolution 1 filters kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D3 outputRows outputCols filters) = S ('D3 inputRows inputCols 1)
220−
221− runForwards (Convolution kernel _) (S3D input) =
234+ runForwards c (S2D input) =
222− let ex = extract input
235+ runForwards c (S3D input :: S ('D3 inputRows inputCols 1))
223− ek = extract kernel
224− ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)
225− iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)
226− kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
227− ky = fromIntegral $ natVal (Proxy :: Proxy kernelCols)
228− sx = fromIntegral $ natVal (Proxy :: Proxy strideRows)
229− sy = fromIntegral $ natVal (Proxy :: Proxy strideCols)
230− ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
231− oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
232−
233− c = vid2col kx ky sx sy ix iy ex
234− mt = c LA.<> ek
235− r = col2vid 1 1 1 1 ox oy mt
236− rs = fromJust . create $ r
237− in (S3D input, S3D rs)
238− runBackwards (Convolution kernel _) (S3D input) (S3D dEdy) =
239− let ex = extract input
240− ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)
241− iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)
242− kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
243− ky = fromIntegral $ natVal (Proxy :: Proxy kernelCols)
244− sx = fromIntegral $ natVal (Proxy :: Proxy strideRows)
245− sy = fromIntegral $ natVal (Proxy :: Proxy strideCols)
246− ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
247− oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
248
236
249− c = vid2col kx ky sx sy ix iy ex
237+ runBackwards c tape grads =
238+ case runBackwards c tape grads of
239+ (c', S3D back :: S ('D3 inputRows inputCols 1)) -> (c', S2D back)
250
240
251− eo = extract dEdy
252− ek = extract kernel
253
241
254− vs = vid2col 1 1 1 1 ox oy eo
242+-- | A two dimensional image may have a convolution filter applied to it producing
243+-- a two dimensional image if both channels and filters is 1.
244+instance ( KnownNat kernelRows
245+ , KnownNat kernelCols
246+ , KnownNat strideRows
247+ , KnownNat strideCols
248+ , KnownNat inputRows
249+ , KnownNat inputCols
250+ , KnownNat outputRows
251+ , KnownNat outputCols
252+ , ((outputRows - 1) * strideRows) ~ (inputRows - kernelRows)
253+ , ((outputCols - 1) * strideCols) ~ (inputCols - kernelCols)
254+ , KnownNat (kernelRows * kernelCols * 1)
255+ , KnownNat (outputRows * 1)
256+ ) => Layer (Convolution 1 1 kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D2 outputRows outputCols) where
257+ type Tape (Convolution 1 1 kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D2 outputRows outputCols) = S ('D3 inputRows inputCols 1)
258+ runForwards c (S2D input) =
259+ case runForwards c (S3D input :: S ('D3 inputRows inputCols 1)) of
260+ (tps, S3D back :: S ('D3 outputRows outputCols 1)) -> (tps, S2D back)
255
261
256− kN = fromJust . create $ tr c LA.<> vs
262+ runBackwards c tape (S2D grads) =
263+ case runBackwards c tape (S3D grads :: S ('D3 outputRows outputCols 1)) of
264+ (c', S3D back :: S ('D3 inputRows inputCols 1)) -> (c', S2D back)
257
265
258− dW = vs LA.<> tr ek
266+-- | A three dimensional image can produce a 2D image from a convolution with 1 filter
267+instance ( KnownNat kernelRows
268+ , KnownNat kernelCols
269+ , KnownNat strideRows
270+ , KnownNat strideCols
271+ , KnownNat inputRows
272+ , KnownNat inputCols
273+ , KnownNat outputRows
274+ , KnownNat outputCols
275+ , KnownNat channels
276+ , ((outputRows - 1) * strideRows) ~ (inputRows - kernelRows)
277+ , ((outputCols - 1) * strideCols) ~ (inputCols - kernelCols)
278+ , KnownNat (kernelRows * kernelCols * channels)
279+ , KnownNat (outputRows * 1)
280+ ) => Layer (Convolution channels 1 kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D2 outputRows outputCols) where
281+ type Tape (Convolution channels 1 kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D2 outputRows outputCols) = S ('D3 inputRows inputCols channels)
282+ runForwards c input =
283+ case runForwards c input of
284+ (tps, S3D back :: S ('D3 outputRows outputCols 1)) -> (tps, S2D back)
259
285
260− xW = col2vid kx ky sx sy ix iy dW
286+ runBackwards c tape (S2D grads) =
261− in (Convolution' kN, S3D . fromJust . create $ xW)
287+ runBackwards c tape (S3D grads :: S ('D3 outputRows outputCols 1))
src/Grenade/Layers/Crop.hs+67−6
@@ -1,20 +1,32 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE GADTs #-}
4{-# LANGUAGE GADTs #-}
4{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeFamilies #-}
6{-# LANGUAGE TypeFamilies #-}
7+{-# LANGUAGE FlexibleContexts #-}
6{-# LANGUAGE MultiParamTypeClasses #-}
8{-# LANGUAGE MultiParamTypeClasses #-}
9+{-# LANGUAGE UndecidableInstances #-}
10+{-|
11+Module : Grenade.Layers.Crop
12+Description : Cropping layer
13+Copyright : (c) Huw Campbell, 2016-2017
14+License : BSD2
15+Stability : experimental
16+-}
7module Grenade.Layers.Crop (
17module Grenade.Layers.Crop (
8 Crop (..)
18 Crop (..)
9 ) where
19 ) where
10
20
11import Data.Maybe
21import Data.Maybe
12import Data.Proxy
22import Data.Proxy
13−import Data.Singletons.TypeLits
23+import Data.Serialize
24+
14import GHC.TypeLits
25import GHC.TypeLits
26+import Data.Kind (Type)
15
27
16−import Grenade.Core.Network
28+import Grenade.Core
17−import Grenade.Core.Shape
29+import Grenade.Layers.Internal.Pad
18
30
19import Numeric.LinearAlgebra (konst, subMatrix, diagBlock)
31import Numeric.LinearAlgebra (konst, subMatrix, diagBlock)
20import Numeric.LinearAlgebra.Static (extract, create)
32import Numeric.LinearAlgebra.Static (extract, create)
@@ -23,7 +35,7 @@ import Numeric.LinearAlgebra.Static (extract, create)
23data Crop :: Nat
35data Crop :: Nat
24 -> Nat
36 -> Nat
25 -> Nat
37 -> Nat
26− -> Nat -> * where
38+ -> Nat -> Type where
27 Crop :: Crop cropLeft cropTop cropRight cropBottom
39 Crop :: Crop cropLeft cropTop cropRight cropBottom
28
40
29instance Show (Crop cropLeft cropTop cropRight cropBottom) where
41instance Show (Crop cropLeft cropTop cropRight cropBottom) where
@@ -34,6 +46,10 @@ instance UpdateLayer (Crop l t r b) where
34 runUpdate _ x _ = x
46 runUpdate _ x _ = x
35 createRandom = return Crop
47 createRandom = return Crop
36
48
49+instance Serialize (Crop l t r b) where
50+ put _ = return ()
51+ get = return Crop
52+
37-- | A two dimentional image can be cropped.
53-- | A two dimentional image can be cropped.
38instance ( KnownNat cropLeft
54instance ( KnownNat cropLeft
39 , KnownNat cropTop
55 , KnownNat cropTop
@@ -43,8 +59,8 @@ instance ( KnownNat cropLeft
43 , KnownNat inputColumns
59 , KnownNat inputColumns
44 , KnownNat outputRows
60 , KnownNat outputRows
45 , KnownNat outputColumns
61 , KnownNat outputColumns
46− , (inputRows - cropTop - cropBottom) ~ outputRows
62+ , (outputRows + cropTop + cropBottom) ~ inputRows
47− , (inputColumns - cropLeft - cropRight) ~ outputColumns
63+ , (outputColumns + cropLeft + cropRight) ~ inputColumns
48 ) => Layer (Crop cropLeft cropTop cropRight cropBottom) ('D2 inputRows inputColumns) ('D2 outputRows outputColumns) where
64 ) => Layer (Crop cropLeft cropTop cropRight cropBottom) ('D2 inputRows inputColumns) ('D2 outputRows outputColumns) where
49 type Tape (Crop cropLeft cropTop cropRight cropBottom) ('D2 inputRows inputColumns) ('D2 outputRows outputColumns) = ()
65 type Tape (Crop cropLeft cropTop cropRight cropBottom) ('D2 inputRows inputColumns) ('D2 outputRows outputColumns) = ()
50 runForwards Crop (S2D input) =
66 runForwards Crop (S2D input) =
@@ -63,3 +79,48 @@ instance ( KnownNat cropLeft
63 eo = extract dEdy
79 eo = extract dEdy
64 vs = diagBlock [konst 0 (cropt,cropl), eo, konst 0 (cropb,cropr)]
80 vs = diagBlock [konst 0 (cropt,cropl), eo, konst 0 (cropb,cropr)]
65 in ((), S2D . fromJust . create $ vs)
81 in ((), S2D . fromJust . create $ vs)
82+
83+
84+-- | A two dimentional image can be cropped.
85+instance ( KnownNat cropLeft
86+ , KnownNat cropTop
87+ , KnownNat cropRight
88+ , KnownNat cropBottom
89+ , KnownNat inputRows
90+ , KnownNat inputColumns
91+ , KnownNat outputRows
92+ , KnownNat outputColumns
93+ , KnownNat channels
94+ , KnownNat (inputRows * channels)
95+ , KnownNat (outputRows * channels)
96+ , (outputRows + cropTop + cropBottom) ~ inputRows
97+ , (outputColumns + cropLeft + cropRight) ~ inputColumns
98+ ) => Layer (Crop cropLeft cropTop cropRight cropBottom) ('D3 inputRows inputColumns channels) ('D3 outputRows outputColumns channels) where
99+ type Tape (Crop cropLeft cropTop cropRight cropBottom) ('D3 inputRows inputColumns channels) ('D3 outputRows outputColumns channels) = ()
100+ runForwards Crop (S3D input) =
101+ let padl = fromIntegral $ natVal (Proxy :: Proxy cropLeft)
102+ padt = fromIntegral $ natVal (Proxy :: Proxy cropTop)
103+ padr = fromIntegral $ natVal (Proxy :: Proxy cropRight)
104+ padb = fromIntegral $ natVal (Proxy :: Proxy cropBottom)
105+ inr = fromIntegral $ natVal (Proxy :: Proxy inputRows)
106+ inc = fromIntegral $ natVal (Proxy :: Proxy inputColumns)
107+ outr = fromIntegral $ natVal (Proxy :: Proxy outputRows)
108+ outc = fromIntegral $ natVal (Proxy :: Proxy outputColumns)
109+ ch = fromIntegral $ natVal (Proxy :: Proxy channels)
110+ m = extract input
111+ cropped = crop ch padl padt padr padb outr outc inr inc m
112+ in ((), S3D . fromJust . create $ cropped)
113+
114+ runBackwards Crop () (S3D gradient) =
115+ let padl = fromIntegral $ natVal (Proxy :: Proxy cropLeft)
116+ padt = fromIntegral $ natVal (Proxy :: Proxy cropTop)
117+ padr = fromIntegral $ natVal (Proxy :: Proxy cropRight)
118+ padb = fromIntegral $ natVal (Proxy :: Proxy cropBottom)
119+ inr = fromIntegral $ natVal (Proxy :: Proxy inputRows)
120+ inc = fromIntegral $ natVal (Proxy :: Proxy inputColumns)
121+ outr = fromIntegral $ natVal (Proxy :: Proxy outputRows)
122+ outc = fromIntegral $ natVal (Proxy :: Proxy outputColumns)
123+ ch = fromIntegral $ natVal (Proxy :: Proxy channels)
124+ m = extract gradient
125+ padded = pad ch padl padt padr padb outr outc inr inc m
126+ in ((), S3D . fromJust . create $ padded)
src/Grenade/Layers/Deconvolution.hsnew file+285
@@ -0,0 +1,285 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE DataKinds #-}
3+{-# LANGUAGE ScopedTypeVariables #-}
4+{-# LANGUAGE GADTs #-}
5+{-# LANGUAGE TypeOperators #-}
6+{-# LANGUAGE TypeFamilies #-}
7+{-# LANGUAGE MultiParamTypeClasses #-}
8+{-# LANGUAGE FlexibleInstances #-}
9+{-# LANGUAGE FlexibleContexts #-}
10+{-# LANGUAGE UndecidableInstances #-}
11+{-|
12+Module : Grenade.Layers.Deconvolution
13+Description : Deconvolution layer
14+Copyright : (c) Huw Campbell, 2016-2017
15+License : BSD2
16+Stability : experimental
17+
18+A deconvolution layer is in many ways a convolution layer in reverse.
19+It learns a kernel to apply to each pixel location, spreading it out
20+into a larger layer.
21+
22+This layer is important for image generation tasks, such as GANs on
23+images.
24+-}
25+module Grenade.Layers.Deconvolution (
26+ Deconvolution (..)
27+ , Deconvolution' (..)
28+ , randomDeconvolution
29+ ) where
30+
31+import Control.Monad.Random hiding ( fromList )
32+import Data.Maybe
33+import Data.Proxy
34+import Data.Serialize
35+
36+import GHC.TypeLits
37+import Data.Kind (Type)
38+
39+import Numeric.LinearAlgebra hiding ( uniformSample, konst )
40+import qualified Numeric.LinearAlgebra as LA
41+import Numeric.LinearAlgebra.Static hiding ((|||), build, toRows)
42+
43+import Grenade.Core
44+import Grenade.Layers.Internal.Convolution
45+import Grenade.Layers.Internal.Update
46+
47+-- | A Deconvolution layer for a neural network.
48+-- This uses the im2col Convolution trick popularised by Caffe.
49+--
50+-- The Deconvolution layer is a way of spreading out a single response
51+-- into a larger image, and is useful in generating images.
52+--
53+data Deconvolution :: Nat -- Number of channels, for the first layer this could be RGB for instance.
54+ -> Nat -- Number of filters, this is the number of channels output by the layer.
55+ -> Nat -- The number of rows in the kernel filter
56+ -> Nat -- The number of column in the kernel filter
57+ -> Nat -- The row stride of the Deconvolution filter
58+ -> Nat -- The columns stride of the Deconvolution filter
59+ -> Type where
60+ Deconvolution :: ( KnownNat channels
61+ , KnownNat filters
62+ , KnownNat kernelRows
63+ , KnownNat kernelColumns
64+ , KnownNat strideRows
65+ , KnownNat strideColumns
66+ , KnownNat kernelFlattened
67+ , kernelFlattened ~ (kernelRows * kernelColumns * filters))
68+ => !(L kernelFlattened channels) -- The kernel filter weights
69+ -> !(L kernelFlattened channels) -- The last kernel update (or momentum)
70+ -> Deconvolution channels filters kernelRows kernelColumns strideRows strideColumns
71+
72+data Deconvolution' :: Nat -- Number of channels, for the first layer this could be RGB for instance.
73+ -> Nat -- Number of filters, this is the number of channels output by the layer.
74+ -> Nat -- The number of rows in the kernel filter
75+ -> Nat -- The number of column in the kernel filter
76+ -> Nat -- The row stride of the Deconvolution filter
77+ -> Nat -- The columns stride of the Deconvolution filter
78+ -> Type where
79+ Deconvolution' :: ( KnownNat channels
80+ , KnownNat filters
81+ , KnownNat kernelRows
82+ , KnownNat kernelColumns
83+ , KnownNat strideRows
84+ , KnownNat strideColumns
85+ , KnownNat kernelFlattened
86+ , kernelFlattened ~ (kernelRows * kernelColumns * filters))
87+ => !(L kernelFlattened channels) -- The kernel filter gradient
88+ -> Deconvolution' channels filters kernelRows kernelColumns strideRows strideColumns
89+
90+instance Show (Deconvolution c f k k' s s') where
91+ show (Deconvolution a _) = renderConv a
92+ where
93+ renderConv mm =
94+ let m = extract mm
95+ ky = fromIntegral $ natVal (Proxy :: Proxy k)
96+ rs = LA.toColumns m
97+ ms = map (take ky) $ toLists . reshape ky <$> rs
98+
99+ render n' | n' <= 0.2 = ' '
100+ | n' <= 0.4 = '.'
101+ | n' <= 0.6 = '-'
102+ | n' <= 0.8 = '='
103+ | otherwise = '#'
104+
105+ px = (fmap . fmap . fmap) render ms
106+ in unlines $ foldl1 (zipWith (\a' b' -> a' ++ " | " ++ b')) $ px
107+
108+randomDeconvolution :: ( MonadRandom m
109+ , KnownNat channels
110+ , KnownNat filters
111+ , KnownNat kernelRows
112+ , KnownNat kernelColumns
113+ , KnownNat strideRows
114+ , KnownNat strideColumns
115+ , KnownNat kernelFlattened
116+ , kernelFlattened ~ (kernelRows * kernelColumns * filters))
117+ => m (Deconvolution channels filters kernelRows kernelColumns strideRows strideColumns)
118+randomDeconvolution = do
119+ s <- getRandom
120+ let wN = uniformSample s (-1) 1
121+ mm = konst 0
122+ return $ Deconvolution wN mm
123+
124+instance ( KnownNat channels
125+ , KnownNat filters
126+ , KnownNat kernelRows
127+ , KnownNat kernelColumns
128+ , KnownNat strideRows
129+ , KnownNat strideColumns
130+ , KnownNat (kernelRows * kernelColumns * filters)
131+ ) => UpdateLayer (Deconvolution channels filters kernelRows kernelColumns strideRows strideColumns) where
132+ type Gradient (Deconvolution channels filters kernelRows kernelColumns strideRows strideColumns) = (Deconvolution' channels filters kernelRows kernelColumns strideRows strideColumns)
133+ runUpdate lp (Deconvolution oldKernel oldMomentum) (Deconvolution' kernelGradient) =
134+ let (newKernel, newMomentum) = descendMatrix (learningRate lp) (learningMomentum lp) (learningRegulariser lp) oldKernel kernelGradient oldMomentum
135+ in Deconvolution newKernel newMomentum
136+
137+ createRandom = randomDeconvolution
138+
139+instance ( KnownNat channels
140+ , KnownNat filters
141+ , KnownNat kernelRows
142+ , KnownNat kernelColumns
143+ , KnownNat strideRows
144+ , KnownNat strideColumns
145+ , KnownNat (kernelRows * kernelColumns * filters)
146+ ) => Serialize (Deconvolution channels filters kernelRows kernelColumns strideRows strideColumns) where
147+ put (Deconvolution w _) = putListOf put . toList . flatten . extract $ w
148+ get = do
149+ let f = fromIntegral $ natVal (Proxy :: Proxy channels)
150+ wN <- maybe (fail "Vector of incorrect size") return . create . reshape f . LA.fromList =<< getListOf get
151+ let mm = konst 0
152+ return $ Deconvolution wN mm
153+
154+-- | A two dimentional image may have a Deconvolution filter applied to it
155+instance ( KnownNat kernelRows
156+ , KnownNat kernelCols
157+ , KnownNat filters
158+ , KnownNat strideRows
159+ , KnownNat strideCols
160+ , KnownNat inputRows
161+ , KnownNat inputCols
162+ , KnownNat outputRows
163+ , KnownNat outputCols
164+ , ((inputRows - 1) * strideRows) ~ (outputRows - kernelRows)
165+ , ((inputCols - 1) * strideCols) ~ (outputCols - kernelCols)
166+ , KnownNat (kernelRows * kernelCols * filters)
167+ , KnownNat (outputRows * filters)
168+ ) => Layer (Deconvolution 1 filters kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D3 outputRows outputCols filters) where
169+ type Tape (Deconvolution 1 filters kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D3 outputRows outputCols filters) = S ('D3 inputRows inputCols 1)
170+ runForwards c (S2D input) =
171+ runForwards c (S3D input :: S ('D3 inputRows inputCols 1))
172+
173+ runBackwards c tape grads =
174+ case runBackwards c tape grads of
175+ (c', S3D back :: S ('D3 inputRows inputCols 1)) -> (c', S2D back)
176+
177+-- | A two dimentional image may have a Deconvolution filter applied to it
178+instance ( KnownNat kernelRows
179+ , KnownNat kernelCols
180+ , KnownNat strideRows
181+ , KnownNat strideCols
182+ , KnownNat inputRows
183+ , KnownNat inputCols
184+ , KnownNat outputRows
185+ , KnownNat outputCols
186+ , ((inputRows - 1) * strideRows) ~ (outputRows - kernelRows)
187+ , ((inputCols - 1) * strideCols) ~ (outputCols - kernelCols)
188+ , KnownNat (kernelRows * kernelCols * 1)
189+ , KnownNat (outputRows * 1)
190+ ) => Layer (Deconvolution 1 1 kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D2 outputRows outputCols) where
191+ type Tape (Deconvolution 1 1 kernelRows kernelCols strideRows strideCols) ('D2 inputRows inputCols) ('D2 outputRows outputCols) = S ('D3 inputRows inputCols 1)
192+ runForwards c (S2D input) =
193+ case runForwards c (S3D input :: S ('D3 inputRows inputCols 1)) of
194+ (tps, S3D fore :: S ('D3 outputRows outputCols 1)) -> (tps, S2D fore)
195+
196+ runBackwards c tape (S2D grads) =
197+ case runBackwards c tape (S3D grads :: S ('D3 outputRows outputCols 1)) of
198+ (c', S3D back :: S ('D3 inputRows inputCols 1)) -> (c', S2D back)
199+
200+-- | A two dimentional image may have a Deconvolution filter applied to it
201+instance ( KnownNat kernelRows
202+ , KnownNat kernelCols
203+ , KnownNat strideRows
204+ , KnownNat strideCols
205+ , KnownNat inputRows
206+ , KnownNat inputCols
207+ , KnownNat outputRows
208+ , KnownNat outputCols
209+ , ((inputRows - 1) * strideRows) ~ (outputRows - kernelRows)
210+ , ((inputCols - 1) * strideCols) ~ (outputCols - kernelCols)
211+ , KnownNat (kernelRows * kernelCols * 1)
212+ , KnownNat (outputRows * 1)
213+ , KnownNat channels
214+ ) => Layer (Deconvolution channels 1 kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D2 outputRows outputCols) where
215+ type Tape (Deconvolution channels 1 kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D2 outputRows outputCols) = S ('D3 inputRows inputCols channels)
216+ runForwards c input =
217+ case runForwards c input of
218+ (tps, S3D fore :: S ('D3 outputRows outputCols 1)) -> (tps, S2D fore)
219+
220+ runBackwards c tape (S2D grads) =
221+ runBackwards c tape (S3D grads :: S ('D3 outputRows outputCols 1))
222+
223+-- | A three dimensional image (or 2d with many channels) can have
224+-- an appropriately sized Deconvolution filter run across it.
225+instance ( KnownNat kernelRows
226+ , KnownNat kernelCols
227+ , KnownNat filters
228+ , KnownNat strideRows
229+ , KnownNat strideCols
230+ , KnownNat inputRows
231+ , KnownNat inputCols
232+ , KnownNat outputRows
233+ , KnownNat outputCols
234+ , KnownNat channels
235+ , ((inputRows - 1) * strideRows) ~ (outputRows - kernelRows)
236+ , ((inputCols - 1) * strideCols) ~ (outputCols - kernelCols)
237+ , KnownNat (kernelRows * kernelCols * filters)
238+ , KnownNat (outputRows * filters)
239+ ) => Layer (Deconvolution channels filters kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D3 outputRows outputCols filters) where
240+
241+ type Tape (Deconvolution channels filters kernelRows kernelCols strideRows strideCols) ('D3 inputRows inputCols channels) ('D3 outputRows outputCols filters) = S ('D3 inputRows inputCols channels)
242+
243+ runForwards (Deconvolution kernel _) (S3D input) =
244+ let ex = extract input
245+ ek = extract kernel
246+ ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)
247+ iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)
248+ kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
249+ ky = fromIntegral $ natVal (Proxy :: Proxy kernelCols)
250+ sx = fromIntegral $ natVal (Proxy :: Proxy strideRows)
251+ sy = fromIntegral $ natVal (Proxy :: Proxy strideCols)
252+ ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
253+ oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
254+
255+ c = vid2col 1 1 1 1 ix iy ex
256+
257+ mt = c LA.<> tr ek
258+
259+ r = col2vid kx ky sx sy ox oy mt
260+ rs = fromJust . create $ r
261+ in (S3D input, S3D rs)
262+ runBackwards (Deconvolution kernel _) (S3D input) (S3D dEdy) =
263+ let ex = extract input
264+ ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)
265+ iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)
266+ kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
267+ ky = fromIntegral $ natVal (Proxy :: Proxy kernelCols)
268+ sx = fromIntegral $ natVal (Proxy :: Proxy strideRows)
269+ sy = fromIntegral $ natVal (Proxy :: Proxy strideCols)
270+ ox = fromIntegral $ natVal (Proxy :: Proxy outputRows)
271+ oy = fromIntegral $ natVal (Proxy :: Proxy outputCols)
272+
273+ c = vid2col 1 1 1 1 ix iy ex
274+
275+ eo = extract dEdy
276+ ek = extract kernel
277+
278+ vs = vid2col kx ky sx sy ox oy eo
279+
280+ kN = fromJust . create . tr $ tr c LA.<> vs
281+
282+ dW = vs LA.<> ek
283+
284+ xW = col2vid 1 1 1 1 ix iy dW
285+ in (Deconvolution' kN, S3D . fromJust . create $ xW)
src/Grenade/Layers/Dropout.hs+14−21
@@ -10,35 +10,28 @@ module Grenade.Layers.Dropout (
10import Control.Monad.Random hiding (fromList)
10import Control.Monad.Random hiding (fromList)
11
11
12import GHC.TypeLits
12import GHC.TypeLits
13−import Grenade.Core.Shape
13+import Grenade.Core
14−import Grenade.Core.Network
15−
16−import Numeric.LinearAlgebra.Static
17−
18
14
19-- Dropout layer help to reduce overfitting.
15-- Dropout layer help to reduce overfitting.
20-- Idea here is that the vector is a shape of 1s and 0s, which we multiply the input by.
16-- Idea here is that the vector is a shape of 1s and 0s, which we multiply the input by.
21-- After backpropogation, we return a new matrix/vector, with different bits dropped out.
17-- After backpropogation, we return a new matrix/vector, with different bits dropped out.
22-- Double is the proportion to drop in each training iteration (like 1% or 5% would be
18-- Double is the proportion to drop in each training iteration (like 1% or 5% would be
23-- reasonable).
19-- reasonable).
24−data Dropout o =
20+data Dropout = Dropout {
25− Dropout (R o)
21+ dropoutRate :: Double
22+ , dropoutSeed :: Int
26− deriving Show
23+ } deriving Show
27
24
28−instance (KnownNat i) => UpdateLayer (Dropout i) where
25+instance UpdateLayer Dropout where
29− type Gradient (Dropout i) = ()
26+ type Gradient Dropout = ()
30 runUpdate _ x _ = x
27 runUpdate _ x _ = x
31 createRandom = randomDropout 0.95
28 createRandom = randomDropout 0.95
32
29
33−randomDropout :: (MonadRandom m, KnownNat i)
30+randomDropout :: MonadRandom m
34− => Double -> m (Dropout i)
31+ => Double -> m Dropout
35−randomDropout rate = do
32+randomDropout rate = Dropout rate <$> getRandom
36− seed <- getRandom
37− let wN = randomVector seed Uniform
38− xs = dvmap (\a -> if a <= rate then 0 else 1) wN
39− return $ Dropout xs
40
33
41−instance (KnownNat i) => Layer (Dropout i) ('D1 i) ('D1 i) where
34+instance (KnownNat i) => Layer Dropout ('D1 i) ('D1 i) where
42− type Tape (Dropout i) ('D1 i) ('D1 i) = ()
35+ type Tape Dropout ('D1 i) ('D1 i) = ()
43− runForwards (Dropout drops) (S1D x) = ((), S1D $ x * drops)
36+ runForwards (Dropout _ _) (S1D x) = ((), S1D x)
44− runBackwards (Dropout drops) _ (S1D x) = ((), S1D $ x * drops)
37+ runBackwards (Dropout _ _) _ (S1D x) = ((), S1D x)
src/Grenade/Layers/Elu.hsnew file+67
@@ -0,0 +1,67 @@
1+{-# LANGUAGE DataKinds #-}
2+{-# LANGUAGE TypeOperators #-}
3+{-# LANGUAGE TypeFamilies #-}
4+{-# LANGUAGE MultiParamTypeClasses #-}
5+{-|
6+Module : Grenade.Layers.Logit
7+Description : Exponential linear unit layer
8+Copyright : (c) Huw Campbell, 2016-2017
9+License : BSD2
10+Stability : experimental
11+-}
12+module Grenade.Layers.Elu (
13+ Elu (..)
14+ ) where
15+
16+import Data.Serialize
17+
18+import GHC.TypeLits
19+import Grenade.Core
20+
21+import qualified Numeric.LinearAlgebra.Static as LAS
22+
23+-- | An exponential linear unit.
24+-- A layer which can act between any shape of the same dimension, acting as a
25+-- diode on every neuron individually.
26+data Elu = Elu
27+ deriving Show
28+
29+instance UpdateLayer Elu where
30+ type Gradient Elu = ()
31+ runUpdate _ _ _ = Elu
32+ createRandom = return Elu
33+
34+instance Serialize Elu where
35+ put _ = return ()
36+ get = return Elu
37+
38+instance ( KnownNat i) => Layer Elu ('D1 i) ('D1 i) where
39+ type Tape Elu ('D1 i) ('D1 i) = LAS.R i
40+
41+ runForwards _ (S1D y) = (y, S1D (elu y))
42+ where
43+ elu = LAS.dvmap (\a -> if a <= 0 then exp a - 1 else a)
44+ runBackwards _ y (S1D dEdy) = ((), S1D (elu' y * dEdy))
45+ where
46+ elu' = LAS.dvmap (\a -> if a <= 0 then exp a else 1)
47+
48+instance (KnownNat i, KnownNat j) => Layer Elu ('D2 i j) ('D2 i j) where
49+ type Tape Elu ('D2 i j) ('D2 i j) = S ('D2 i j)
50+
51+ runForwards _ (S2D y) = (S2D y, S2D (elu y))
52+ where
53+ elu = LAS.dmmap (\a -> if a <= 0 then exp a - 1 else a)
54+ runBackwards _ (S2D y) (S2D dEdy) = ((), S2D (elu' y * dEdy))
55+ where
56+ elu' = LAS.dmmap (\a -> if a <= 0 then exp a else 1)
57+
58+instance (KnownNat i, KnownNat j, KnownNat k) => Layer Elu ('D3 i j k) ('D3 i j k) where
59+
60+ type Tape Elu ('D3 i j k) ('D3 i j k) = S ('D3 i j k)
61+
62+ runForwards _ (S3D y) = (S3D y, S3D (elu y))
63+ where
64+ elu = LAS.dmmap (\a -> if a <= 0 then exp a - 1 else a)
65+ runBackwards _ (S3D y) (S3D dEdy) = ((), S3D (elu' y * dEdy))
66+ where
67+ elu' = LAS.dmmap (\a -> if a <= 0 then exp a else 1)
src/Grenade/Layers/Flatten.hsdeleted−56
@@ -1,56 +0,0 @@
1−{-# LANGUAGE DataKinds #-}
2−{-# LANGUAGE TypeOperators #-}
3−{-# LANGUAGE TypeFamilies #-}
4−{-# LANGUAGE MultiParamTypeClasses #-}
5−{-# LANGUAGE FlexibleContexts #-}
6−module Grenade.Layers.Flatten (
7− FlattenLayer (..)
8− ) where
9−
10−import Data.Serialize
11−
12−import Data.Singletons.TypeLits
13−import GHC.TypeLits
14−
15−import Numeric.LinearAlgebra.Static
16−import Numeric.LinearAlgebra.Data as LA ( flatten )
17−
18−import Grenade.Core.Shape
19−import Grenade.Core.Network
20−
21−-- | Flatten Layer
22−--
23−-- Flattens input down to D1 from either 2D or 3D data.
24−--
25−-- Can also be used to turn a 3D image with only one channel into a 2D image.
26−data FlattenLayer = FlattenLayer
27− deriving Show
28−
29−instance UpdateLayer FlattenLayer where
30− type Gradient FlattenLayer = ()
31− runUpdate _ _ _ = FlattenLayer
32− createRandom = return FlattenLayer
33−
34−instance (KnownNat a, KnownNat x, KnownNat y, a ~ (x * z)) => Layer FlattenLayer ('D2 x y) ('D1 a) where
35− type Tape FlattenLayer ('D2 x y) ('D1 a) = ()
36− runForwards _ (S2D y) = ((), fromJust' . fromStorable . flatten . extract $ y)
37− runBackwards _ _ (S1D y) = ((), fromJust' . fromStorable . extract $ y)
38−
39−instance (KnownNat a, KnownNat x, KnownNat y, KnownNat (x * z), KnownNat z, a ~ (x * y * z)) => Layer FlattenLayer ('D3 x y z) ('D1 a) where
40− type Tape FlattenLayer ('D3 x y z) ('D1 a) = ()
41− runForwards _ (S3D y) = ((), fromJust' . fromStorable . flatten . extract $ y)
42− runBackwards _ _ (S1D y) = ((), fromJust' . fromStorable . extract $ y)
43−
44−instance (KnownNat y, KnownNat x, KnownNat z, z ~ 1) => Layer FlattenLayer ('D3 x y z) ('D2 x y) where
45− type Tape FlattenLayer ('D3 x y z) ('D2 x y) = ()
46− runForwards _ (S3D y) = ((), S2D y)
47− runBackwards _ _ (S2D y) = ((), S3D y)
48−
49−instance Serialize FlattenLayer where
50− put _ = return ()
51− get = return FlattenLayer
52−
53−
54−fromJust' :: Maybe x -> x
55−fromJust' (Just x) = x
56−fromJust' Nothing = error $ "FlattenLayer error: data shape couldn't be converted."
src/Grenade/Layers/FullyConnected.hs+10−11
@@ -1,11 +1,11 @@
1{-# LANGUAGE DataKinds #-}
1{-# LANGUAGE DataKinds #-}
2−{-# LANGUAGE RecordWildCards #-}
3{-# LANGUAGE TypeOperators #-}
2{-# LANGUAGE TypeOperators #-}
4{-# LANGUAGE TypeFamilies #-}
3{-# LANGUAGE TypeFamilies #-}
5{-# LANGUAGE MultiParamTypeClasses #-}
4{-# LANGUAGE MultiParamTypeClasses #-}
6{-# LANGUAGE ScopedTypeVariables #-}
5{-# LANGUAGE ScopedTypeVariables #-}
7module Grenade.Layers.FullyConnected (
6module Grenade.Layers.FullyConnected (
8 FullyConnected (..)
7 FullyConnected (..)
8+ , FullyConnected' (..)
9 , randomFullyConnected
9 , randomFullyConnected
10 ) where
10 ) where
11
11
@@ -13,19 +13,18 @@ import Control.Monad.Random hiding (fromList)
13
13
14import Data.Proxy
14import Data.Proxy
15import Data.Serialize
15import Data.Serialize
16−import Data.Singletons.TypeLits
17
16
18import qualified Numeric.LinearAlgebra as LA
17import qualified Numeric.LinearAlgebra as LA
19import Numeric.LinearAlgebra.Static
18import Numeric.LinearAlgebra.Static
20
19
21−import Grenade.Core.Network
20+import Grenade.Core
22−import Grenade.Core.Shape
23
21
24import Grenade.Layers.Internal.Update
22import Grenade.Layers.Internal.Update
23+import GHC.TypeLits
25
24
26-- | A basic fully connected (or inner product) neural network layer.
25-- | A basic fully connected (or inner product) neural network layer.
27data FullyConnected i o = FullyConnected
26data FullyConnected i o = FullyConnected
28− !(FullyConnected' i o) -- Weuron weights
27+ !(FullyConnected' i o) -- Neuron weights
29 !(FullyConnected' i o) -- Neuron momentum
28 !(FullyConnected' i o) -- Neuron momentum
30
29
31data FullyConnected' i o = FullyConnected'
30data FullyConnected' i o = FullyConnected'
@@ -38,20 +37,20 @@ instance Show (FullyConnected i o) where
38instance (KnownNat i, KnownNat o) => UpdateLayer (FullyConnected i o) where
37instance (KnownNat i, KnownNat o) => UpdateLayer (FullyConnected i o) where
39 type Gradient (FullyConnected i o) = (FullyConnected' i o)
38 type Gradient (FullyConnected i o) = (FullyConnected' i o)
40
39
41− runUpdate LearningParameters {..} (FullyConnected (FullyConnected' oldBias oldActivations) (FullyConnected' oldBiasMomentum oldMomentum)) (FullyConnected' biasGradient activationGradient) =
40+ runUpdate lp (FullyConnected (FullyConnected' oldBias oldActivations) (FullyConnected' oldBiasMomentum oldMomentum)) (FullyConnected' biasGradient activationGradient) =
42− let (newBias, newBiasMomentum) = decendVector learningRate learningMomentum learningRegulariser oldBias biasGradient oldBiasMomentum
41+ let (newBias, newBiasMomentum) = descendVector (learningRate lp) (learningMomentum lp) (learningRegulariser lp) oldBias biasGradient oldBiasMomentum
43− (newActivations, newMomentum) = decendMatrix learningRate learningMomentum learningRegulariser oldActivations activationGradient oldMomentum
42+ (newActivations, newMomentum) = descendMatrix (learningRate lp) (learningMomentum lp) (learningRegulariser lp) oldActivations activationGradient oldMomentum
44 in FullyConnected (FullyConnected' newBias newActivations) (FullyConnected' newBiasMomentum newMomentum)
43 in FullyConnected (FullyConnected' newBias newActivations) (FullyConnected' newBiasMomentum newMomentum)
45
44
46 createRandom = randomFullyConnected
45 createRandom = randomFullyConnected
47
46
48instance (KnownNat i, KnownNat o) => Layer (FullyConnected i o) ('D1 i) ('D1 o) where
47instance (KnownNat i, KnownNat o) => Layer (FullyConnected i o) ('D1 i) ('D1 o) where
49− type Tape (FullyConnected i o) ('D1 i) ('D1 o) = S ('D1 i)
48+ type Tape (FullyConnected i o) ('D1 i) ('D1 o) = R i
50 -- Do a matrix vector multiplication and return the result.
49 -- Do a matrix vector multiplication and return the result.
51− runForwards (FullyConnected (FullyConnected' wB wN) _) (S1D v) = (S1D v, S1D (wB + wN #> v))
50+ runForwards (FullyConnected (FullyConnected' wB wN) _) (S1D v) = (v, S1D (wB + wN #> v))
52
51
53 -- Run a backpropogation step for a full connected layer.
52 -- Run a backpropogation step for a full connected layer.
54− runBackwards (FullyConnected (FullyConnected' _ wN) _) (S1D x) (S1D dEdy) =
53+ runBackwards (FullyConnected (FullyConnected' _ wN) _) x (S1D dEdy) =
55 let wB' = dEdy
54 let wB' = dEdy
56 mm' = dEdy `outer` x
55 mm' = dEdy `outer` x
57 -- calcluate derivatives for next step
56 -- calcluate derivatives for next step
src/Grenade/Layers/Fuse.hsdeleted−46
@@ -1,46 +0,0 @@
1−{-# LANGUAGE DataKinds #-}
2−{-# LANGUAGE GADTs #-}
3−{-# LANGUAGE ScopedTypeVariables #-}
4−{-# LANGUAGE TypeOperators #-}
5−{-# LANGUAGE TypeFamilies #-}
6−{-# LANGUAGE MultiParamTypeClasses #-}
7−{-# LANGUAGE FlexibleContexts #-}
8−{-# LANGUAGE FlexibleInstances #-}
9−module Grenade.Layers.Fuse (
10− Fuse (..)
11− ) where
12−
13−import Grenade.Core.Network
14−import Grenade.Core.Shape
15−
16−-- | Fuse two layers into one layer.
17−-- This can be used to simplify a network if a complicated repeated structure is used.
18−data Fuse :: * -> * -> Shape -> Shape -> Shape -> * where
19− (:$$) :: (Layer x i h, Layer y h o)
20− => !x
21− -> !y
22− -> Fuse x y i h o
23−infixr 5 :$$
24−
25−instance (Show x, Show y) => Show (Fuse x y i h o) where
26− show (x :$$ y) = "(" ++ show x ++ " :$$ " ++ show y ++ ")"
27−
28−instance (Layer x i h, Layer y h o) => UpdateLayer (Fuse x y i h o) where
29− type Gradient (Fuse x y i h o) = (Gradient x, Gradient y)
30− runUpdate lr (x :$$ y) (x', y') =
31− let newX = runUpdate lr x x'
32− newY = runUpdate lr y y'
33− in (newX :$$ newY)
34− createRandom = (:$$) <$> createRandom <*> createRandom
35−
36−instance (Layer x i h, Layer y h o) => Layer (Fuse x y i h o) i o where
37− type Tape (Fuse x y i h o) i o = (Tape x i h, Tape y h o)
38− runForwards (x :$$ y) input =
39− let (xTape, yInput :: S h) = runForwards x input
40− (yTape, output :: S o) = runForwards y yInput
41− in ((xTape, yTape), output)
42−
43− runBackwards (x :$$ y) (xTape, yTape) backGradient =
44− let (y', yGrad :: S h) = runBackwards y yTape backGradient
45− (x', xGrad) = runBackwards x xTape yGrad
46− in ((x', y'), xGrad)
src/Grenade/Layers/Inception.hsnew file+67
@@ -0,0 +1,67 @@
1+{-# LANGUAGE DataKinds #-}
2+{-# LANGUAGE GADTs #-}
3+{-# LANGUAGE TypeOperators #-}
4+{-# LANGUAGE TypeFamilies #-}
5+{-# LANGUAGE MultiParamTypeClasses #-}
6+{-# LANGUAGE FlexibleContexts #-}
7+{-# LANGUAGE RankNTypes #-}
8+{-# LANGUAGE FlexibleInstances #-}
9+{-# LANGUAGE ScopedTypeVariables #-}
10+{-|
11+Module : Grenade.Core.Network
12+Description : Inception style parallel convolutional network composition.
13+Copyright : (c) Huw Campbell, 2016-2017
14+License : BSD2
15+Stability : experimental
16+
17+Export an Inception style type, which can be used to build up
18+complex multiconvolution size networks.
19+-}
20+module Grenade.Layers.Inception (
21+ Inception
22+ , InceptionMini
23+ , Resnet
24+ ) where
25+
26+import GHC.TypeLits
27+
28+import Grenade.Core
29+import Grenade.Layers.Convolution
30+import Grenade.Layers.Pad
31+import Grenade.Layers.Concat
32+import Grenade.Layers.Merge
33+import Grenade.Layers.Trivial
34+
35+-- | Type of an inception layer.
36+--
37+-- It looks like a bit of a handful, but is actually pretty easy to use.
38+--
39+-- The first three type parameters are the size of the (3D) data the
40+-- inception layer will take. It will emit 3D data with the number of
41+-- channels being the sum of @chx@, @chy@, @chz@, which are the number
42+-- of convolution filters in the 3x3, 5x5, and 7x7 convolutions Layers
43+-- respectively.
44+--
45+-- The network get padded effectively before each convolution filters
46+-- such that the output dimension is the same x and y as the input.
47+type Inception rows cols channels chx chy chz
48+ = Network '[ Concat ('D3 rows cols (chx + chy)) (InceptionMini rows cols channels chx chy) ('D3 rows cols chz) (Inception7x7 rows cols channels chz) ]
49+ '[ 'D3 rows cols channels, 'D3 rows cols (chx + chy + chz) ]
50+
51+type InceptionMini rows cols channels chx chy
52+ = Network '[ Concat ('D3 rows cols chx) (Inception3x3 rows cols channels chx) ('D3 rows cols chy) (Inception5x5 rows cols channels chy) ]
53+ '[ 'D3 rows cols channels, 'D3 rows cols (chx + chy) ]
54+
55+type Inception3x3 rows cols channels chx
56+ = Network '[ Pad 1 1 1 1, Convolution channels chx 3 3 1 1 ]
57+ '[ 'D3 rows cols channels, 'D3 (rows + 2) (cols + 2) channels, 'D3 rows cols chx ]
58+
59+type Inception5x5 rows cols channels chx
60+ = Network '[ Pad 2 2 2 2, Convolution channels chx 5 5 1 1 ]
61+ '[ 'D3 rows cols channels, 'D3 (rows + 4) (cols + 4) channels, 'D3 rows cols chx ]
62+
63+type Inception7x7 rows cols channels chx
64+ = Network '[ Pad 3 3 3 3, Convolution channels chx 7 7 1 1 ]
65+ '[ 'D3 rows cols channels, 'D3 (rows + 6) (cols + 6) channels, 'D3 rows cols chx ]
66+
67+type Resnet branch = Merge Trivial branch
src/Grenade/Layers/Internal/Pad.hsnew file+54
@@ -0,0 +1,54 @@
1+{-# LANGUAGE ForeignFunctionInterface #-}
2+module Grenade.Layers.Internal.Pad (
3+ pad
4+ , crop
5+ ) where
6+
7+import qualified Data.Vector.Storable as U ( unsafeToForeignPtr0, unsafeFromForeignPtr0 )
8+
9+import Foreign ( mallocForeignPtrArray, withForeignPtr )
10+import Foreign.Ptr ( Ptr )
11+
12+import Numeric.LinearAlgebra ( flatten, Matrix )
13+import qualified Numeric.LinearAlgebra.Devel as U
14+
15+import System.IO.Unsafe ( unsafePerformIO )
16+
17+pad :: Int -> Int -> Int -> Int -> Int -> Int -> Int -> Int -> Int -> Matrix Double -> Matrix Double
18+pad channels padLeft padTop padRight padBottom rows cols rows' cols' m
19+ = let outMatSize = rows' * cols' * channels
20+ vec = flatten m
21+ in unsafePerformIO $ do
22+ outPtr <- mallocForeignPtrArray outMatSize
23+ let (inPtr, _) = U.unsafeToForeignPtr0 vec
24+
25+ withForeignPtr inPtr $ \inPtr' ->
26+ withForeignPtr outPtr $ \outPtr' ->
27+ pad_cpu inPtr' channels rows cols padLeft padTop padRight padBottom outPtr'
28+
29+ let matVec = U.unsafeFromForeignPtr0 outPtr outMatSize
30+ return (U.matrixFromVector U.RowMajor (rows' * channels) cols' matVec)
31+{-# INLINE pad #-}
32+
33+foreign import ccall unsafe
34+ pad_cpu
35+ :: Ptr Double -> Int -> Int -> Int -> Int -> Int -> Int -> Int -> Ptr Double -> IO ()
36+
37+crop :: Int -> Int -> Int -> Int -> Int -> Int -> Int -> Int -> Int -> Matrix Double -> Matrix Double
38+crop channels padLeft padTop padRight padBottom rows cols _ _ m
39+ = let outMatSize = rows * cols * channels
40+ vec = flatten m
41+ in unsafePerformIO $ do
42+ outPtr <- mallocForeignPtrArray outMatSize
43+ let (inPtr, _) = U.unsafeToForeignPtr0 vec
44+
45+ withForeignPtr inPtr $ \inPtr' ->
46+ withForeignPtr outPtr $ \outPtr' ->
47+ crop_cpu inPtr' channels rows cols padLeft padTop padRight padBottom outPtr'
48+
49+ let matVec = U.unsafeFromForeignPtr0 outPtr outMatSize
50+ return (U.matrixFromVector U.RowMajor (rows * channels) cols matVec)
51+
52+foreign import ccall unsafe
53+ crop_cpu
54+ :: Ptr Double -> Int -> Int -> Int -> Int -> Int -> Int -> Int -> Ptr Double -> IO ()
src/Grenade/Layers/Internal/Update.hs+12−12
@@ -1,7 +1,7 @@
1{-# LANGUAGE ForeignFunctionInterface #-}
1{-# LANGUAGE ForeignFunctionInterface #-}
2module Grenade.Layers.Internal.Update (
2module Grenade.Layers.Internal.Update (
3− decendMatrix
3+ descendMatrix
4− , decendVector
4+ , descendVector
5 ) where
5 ) where
6
6
7import Data.Maybe ( fromJust )
7import Data.Maybe ( fromJust )
@@ -17,8 +17,8 @@ import qualified Numeric.LinearAlgebra.Devel as U
17
17
18import System.IO.Unsafe ( unsafePerformIO )
18import System.IO.Unsafe ( unsafePerformIO )
19
19
20−decendMatrix :: (KnownNat rows, KnownNat columns) => Double -> Double -> Double -> L rows columns -> L rows columns -> L rows columns -> (L rows columns, L rows columns)
20+descendMatrix :: (KnownNat rows, KnownNat columns) => Double -> Double -> Double -> L rows columns -> L rows columns -> L rows columns -> (L rows columns, L rows columns)
21−decendMatrix rate momentum regulariser weights gradient lastUpdate =
21+descendMatrix rate momentum regulariser weights gradient lastUpdate =
22 let (rows, cols) = size weights
22 let (rows, cols) = size weights
23 len = rows * cols
23 len = rows * cols
24 -- Most gradients come in in ColumnMajor,
24 -- Most gradients come in in ColumnMajor,
@@ -29,7 +29,7 @@ decendMatrix rate momentum regulariser weights gradient lastUpdate =
29 weights' = flatten . tr . extract $ weights
29 weights' = flatten . tr . extract $ weights
30 gradient' = flatten . tr . extract $ gradient
30 gradient' = flatten . tr . extract $ gradient
31 lastUpdate' = flatten . tr . extract $ lastUpdate
31 lastUpdate' = flatten . tr . extract $ lastUpdate
32− (vw, vm) = decendUnsafe len rate momentum regulariser weights' gradient' lastUpdate'
32+ (vw, vm) = descendUnsafe len rate momentum regulariser weights' gradient' lastUpdate'
33
33
34 -- Note that it's ColumnMajor, as we did a transpose before
34 -- Note that it's ColumnMajor, as we did a transpose before
35 -- using the internal vectors.
35 -- using the internal vectors.
@@ -37,17 +37,17 @@ decendMatrix rate momentum regulariser weights gradient lastUpdate =
37 mm = U.matrixFromVector U.ColumnMajor rows cols vm
37 mm = U.matrixFromVector U.ColumnMajor rows cols vm
38 in (fromJust . create $ mw, fromJust . create $ mm)
38 in (fromJust . create $ mw, fromJust . create $ mm)
39
39
40−decendVector :: (KnownNat r) => Double -> Double -> Double -> R r -> R r -> R r -> (R r, R r)
40+descendVector :: (KnownNat r) => Double -> Double -> Double -> R r -> R r -> R r -> (R r, R r)
41−decendVector rate momentum regulariser weights gradient lastUpdate =
41+descendVector rate momentum regulariser weights gradient lastUpdate =
42 let len = size weights
42 let len = size weights
43 weights' = extract weights
43 weights' = extract weights
44 gradient' = extract gradient
44 gradient' = extract gradient
45 lastUpdate' = extract lastUpdate
45 lastUpdate' = extract lastUpdate
46− (vw, vm) = decendUnsafe len rate momentum regulariser weights' gradient' lastUpdate'
46+ (vw, vm) = descendUnsafe len rate momentum regulariser weights' gradient' lastUpdate'
47 in (fromJust $ create vw, fromJust $ create vm)
47 in (fromJust $ create vw, fromJust $ create vm)
48
48
49−decendUnsafe :: Int -> Double -> Double -> Double -> Vector Double -> Vector Double -> Vector Double -> (Vector Double, Vector Double)
49+descendUnsafe :: Int -> Double -> Double -> Double -> Vector Double -> Vector Double -> Vector Double -> (Vector Double, Vector Double)
50−decendUnsafe len rate momentum regulariser weights gradient lastUpdate =
50+descendUnsafe len rate momentum regulariser weights gradient lastUpdate =
51 unsafePerformIO $ do
51 unsafePerformIO $ do
52 outWPtr <- mallocForeignPtrArray len
52 outWPtr <- mallocForeignPtrArray len
53 outMPtr <- mallocForeignPtrArray len
53 outMPtr <- mallocForeignPtrArray len
@@ -60,11 +60,11 @@ decendUnsafe len rate momentum regulariser weights gradient lastUpdate =
60 withForeignPtr lPtr $ \lPtr' ->
60 withForeignPtr lPtr $ \lPtr' ->
61 withForeignPtr outWPtr $ \outWPtr' ->
61 withForeignPtr outWPtr $ \outWPtr' ->
62 withForeignPtr outMPtr $ \outMPtr' ->
62 withForeignPtr outMPtr $ \outMPtr' ->
63− decend_cpu len rate momentum regulariser wPtr' gPtr' lPtr' outWPtr' outMPtr'
63+ descend_cpu len rate momentum regulariser wPtr' gPtr' lPtr' outWPtr' outMPtr'
64
64
65 return (U.unsafeFromForeignPtr0 outWPtr len, U.unsafeFromForeignPtr0 outMPtr len)
65 return (U.unsafeFromForeignPtr0 outWPtr len, U.unsafeFromForeignPtr0 outMPtr len)
66
66
67foreign import ccall unsafe
67foreign import ccall unsafe
68− decend_cpu
68+ descend_cpu
69 :: Int -> Double -> Double -> Double -> Ptr Double -> Ptr Double -> Ptr Double -> Ptr Double -> Ptr Double -> IO ()
69 :: Int -> Double -> Double -> Double -> Ptr Double -> Ptr Double -> Ptr Double -> Ptr Double -> Ptr Double -> IO ()
70
70
src/Grenade/Layers/Logit.hs+28−24
@@ -1,19 +1,27 @@
1{-# LANGUAGE DataKinds #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE TypeOperators #-}
2{-# LANGUAGE TypeOperators #-}
3{-# LANGUAGE TypeFamilies #-}
3{-# LANGUAGE TypeFamilies #-}
4+{-# LANGUAGE FlexibleInstances #-}
4{-# LANGUAGE MultiParamTypeClasses #-}
5{-# LANGUAGE MultiParamTypeClasses #-}
6+{-|
7+Module : Grenade.Layers.Logit
8+Description : Sigmoid nonlinear layer
9+Copyright : (c) Huw Campbell, 2016-2017
10+License : BSD2
11+Stability : experimental
12+-}
5module Grenade.Layers.Logit (
13module Grenade.Layers.Logit (
6 Logit (..)
14 Logit (..)
7 ) where
15 ) where
8
16
9
17
10import Data.Serialize
18import Data.Serialize
19+import Data.Singletons
11
20
12−import Data.Singletons.TypeLits
13−import Grenade.Core.Network
14−import Grenade.Core.Shape
21+import Grenade.Core
15
22
16-- | A Logit layer.
23-- | A Logit layer.
24+--
17-- A layer which can act between any shape of the same dimension, perfoming an sigmoid function.
25-- A layer which can act between any shape of the same dimension, perfoming an sigmoid function.
18-- This layer should be used as the output layer of a network for logistic regression (classification)
26-- This layer should be used as the output layer of a network for logistic regression (classification)
19-- problems.
27-- problems.
@@ -25,29 +33,25 @@ instance UpdateLayer Logit where
25 runUpdate _ _ _ = Logit
33 runUpdate _ _ _ = Logit
26 createRandom = return Logit
34 createRandom = return Logit
27
35
28−instance (KnownNat i) => Layer Logit ('D1 i) ('D1 i) where
36+instance (a ~ b, SingI a) => Layer Logit a b where
29− type Tape Logit ('D1 i) ('D1 i) = S ('D1 i)
37+ -- Wengert tape optimisation:
30− runForwards _ (S1D y) = (S1D y, S1D (logistic y))
38+ --
31− runBackwards _ (S1D y) (S1D dEdy) = ((), S1D (logistic' y * dEdy))
39+ -- Derivative of the sigmoid function is
32−
40+ -- d σ(x) / dx = σ(x) • (1 - σ(x))
33−instance (KnownNat i, KnownNat j) => Layer Logit ('D2 i j) ('D2 i j) where
41+ -- but we have already calculated σ(x) in
42+ -- the forward pass, so just store that
43+ -- and use it in the backwards pass.
34− type Tape Logit ('D2 i j) ('D2 i j) = S ('D2 i j)
44+ type Tape Logit a b = S a
35− runForwards _ (S2D y) = (S2D y, S2D (logistic y))
45+ runForwards _ a =
36− runBackwards _ (S2D y) (S2D dEdy) = ((), S2D (logistic' y * dEdy))
46+ let l = sigmoid a
37−
47+ in (l, l)
38−instance (KnownNat i, KnownNat j, KnownNat k) => Layer Logit ('D3 i j k) ('D3 i j k) where
48+ runBackwards _ l g =
39− type Tape Logit ('D3 i j k) ('D3 i j k) = S ('D3 i j k)
49+ let sigmoid' = l * (1 - l)
40− runForwards _ (S3D y) = (S3D y, S3D (logistic y))
50+ in ((), sigmoid' * g)
41− runBackwards _ (S3D y) (S3D dEdy) = ((), S3D (logistic' y * dEdy))
42
51
43instance Serialize Logit where
52instance Serialize Logit where
44 put _ = return ()
53 put _ = return ()
45 get = return Logit
54 get = return Logit
46
55
47−logistic :: Floating a => a -> a
56+sigmoid :: Floating a => a -> a
48−logistic x = 1 / (1 + exp (-x))
57+sigmoid x = 1 / (1 + exp (-x))
49−
50−logistic' :: Floating a => a -> a
51−logistic' x = logix * (1 - logix)
52− where
53− logix = logistic x
src/Grenade/Layers/Merge.hsnew file+66
@@ -0,0 +1,66 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE DataKinds #-}
3+{-# LANGUAGE GADTs #-}
4+{-# LANGUAGE TypeOperators #-}
5+{-# LANGUAGE TypeFamilies #-}
6+{-# LANGUAGE MultiParamTypeClasses #-}
7+{-# LANGUAGE FlexibleContexts #-}
8+{-# LANGUAGE RankNTypes #-}
9+{-# LANGUAGE FlexibleInstances #-}
10+{-# LANGUAGE ScopedTypeVariables #-}
11+{-# LANGUAGE StandaloneDeriving #-}
12+{-|
13+Module : Grenade.Core.Network
14+Description : Merging layer for parallel network composition
15+Copyright : (c) Huw Campbell, 2016-2017
16+License : BSD2
17+Stability : experimental
18+-}
19+module Grenade.Layers.Merge (
20+ Merge (..)
21+ ) where
22+
23+import Data.Serialize
24+
25+import Data.Singletons
26+
27+#if MIN_VERSION_base(4,9,0)
28+import Data.Kind (Type)
29+#endif
30+
31+import Grenade.Core
32+
33+-- | A Merging layer.
34+--
35+-- Similar to Concat layer, except sums the activations instead of creating a larger
36+-- shape.
37+data Merge :: Type -> Type -> Type where
38+ Merge :: x -> y -> Merge x y
39+
40+instance (Show x, Show y) => Show (Merge x y) where
41+ show (Merge x y) = "Merge\n" ++ show x ++ "\n" ++ show y
42+
43+-- | Run two layers in parallel, combining their outputs.
44+-- This just kind of "smooshes" the weights together.
45+instance (UpdateLayer x, UpdateLayer y) => UpdateLayer (Merge x y) where
46+ type Gradient (Merge x y) = (Gradient x, Gradient y)
47+ runUpdate lr (Merge x y) (x', y') = Merge (runUpdate lr x x') (runUpdate lr y y')
48+ createRandom = Merge <$> createRandom <*> createRandom
49+
50+-- | Combine the outputs and the inputs, summing the output shape
51+instance (SingI i, SingI o, Layer x i o, Layer y i o) => Layer (Merge x y) i o where
52+ type Tape (Merge x y) i o = (Tape x i o, Tape y i o)
53+
54+ runForwards (Merge x y) input =
55+ let (xT, xOut) = runForwards x input
56+ (yT, yOut) = runForwards y input
57+ in ((xT, yT), xOut + yOut)
58+
59+ runBackwards (Merge x y) (xTape, yTape) o =
60+ let (x', xB) = runBackwards x xTape o
61+ (y', yB) = runBackwards y yTape o
62+ in ((x', y'), xB + yB)
63+
64+instance (Serialize a, Serialize b) => Serialize (Merge a b) where
65+ put (Merge a b) = put a *> put b
66+ get = Merge <$> get <*> get
src/Grenade/Layers/Pad.hs+62−5
@@ -1,9 +1,19 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE GADTs #-}
4{-# LANGUAGE GADTs #-}
4{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeFamilies #-}
6{-# LANGUAGE TypeFamilies #-}
6{-# LANGUAGE MultiParamTypeClasses #-}
7{-# LANGUAGE MultiParamTypeClasses #-}
8+{-# LANGUAGE FlexibleContexts #-}
9+{-# LANGUAGE UndecidableInstances #-}
10+{-|
11+Module : Grenade.Core.Pad
12+Description : Padding layer for 2D and 3D images
13+Copyright : (c) Huw Campbell, 2016-2017
14+License : BSD2
15+Stability : experimental
16+-}
7module Grenade.Layers.Pad (
17module Grenade.Layers.Pad (
8 Pad (..)
18 Pad (..)
9 ) where
19 ) where
@@ -11,20 +21,23 @@ module Grenade.Layers.Pad (
11import Data.Maybe
21import Data.Maybe
12import Data.Proxy
22import Data.Proxy
13import Data.Serialize
23import Data.Serialize
14−import Data.Singletons.TypeLits
24+
15import GHC.TypeLits
25import GHC.TypeLits
26+import Data.Kind (Type)
16
27
17−import Grenade.Core.Network
28+import Grenade.Core
18−import Grenade.Core.Shape
29+import Grenade.Layers.Internal.Pad
19
30
20import Numeric.LinearAlgebra (konst, subMatrix, diagBlock)
31import Numeric.LinearAlgebra (konst, subMatrix, diagBlock)
21import Numeric.LinearAlgebra.Static (extract, create)
32import Numeric.LinearAlgebra.Static (extract, create)
22
33
23-- | A padding layer for a neural network.
34-- | A padding layer for a neural network.
35+--
36+-- Pads on the X and Y dimension of an image.
24data Pad :: Nat
37data Pad :: Nat
25 -> Nat
38 -> Nat
26 -> Nat
39 -> Nat
27− -> Nat -> * where
40+ -> Nat -> Type where
28 Pad :: Pad padLeft padTop padRight padBottom
41 Pad :: Pad padLeft padTop padRight padBottom
29
42
30instance Show (Pad padLeft padTop padRight padBottom) where
43instance Show (Pad padLeft padTop padRight padBottom) where
@@ -39,7 +52,7 @@ instance Serialize (Pad l t r b) where
39 put _ = return ()
52 put _ = return ()
40 get = return Pad
53 get = return Pad
41
54
42−-- | A two dimentional image can be padped.
55+-- | A two dimentional image can be padded.
43instance ( KnownNat padLeft
56instance ( KnownNat padLeft
44 , KnownNat padTop
57 , KnownNat padTop
45 , KnownNat padRight
58 , KnownNat padRight
@@ -68,3 +81,47 @@ instance ( KnownNat padLeft
68 m = extract dEdy
81 m = extract dEdy
69 vs = subMatrix (padt, padl) (nrows, ncols) m
82 vs = subMatrix (padt, padl) (nrows, ncols) m
70 in ((), S2D . fromJust . create $ vs)
83 in ((), S2D . fromJust . create $ vs)
84+
85+-- | A two dimentional image can be padded.
86+instance ( KnownNat padLeft
87+ , KnownNat padTop
88+ , KnownNat padRight
89+ , KnownNat padBottom
90+ , KnownNat inputRows
91+ , KnownNat inputColumns
92+ , KnownNat outputRows
93+ , KnownNat outputColumns
94+ , KnownNat channels
95+ , KnownNat (inputRows * channels)
96+ , KnownNat (outputRows * channels)
97+ , (inputRows + padTop + padBottom) ~ outputRows
98+ , (inputColumns + padLeft + padRight) ~ outputColumns
99+ ) => Layer (Pad padLeft padTop padRight padBottom) ('D3 inputRows inputColumns channels) ('D3 outputRows outputColumns channels) where
100+ type Tape (Pad padLeft padTop padRight padBottom) ('D3 inputRows inputColumns channels) ('D3 outputRows outputColumns channels) = ()
101+ runForwards Pad (S3D input) =
102+ let padl = fromIntegral $ natVal (Proxy :: Proxy padLeft)
103+ padt = fromIntegral $ natVal (Proxy :: Proxy padTop)
104+ padr = fromIntegral $ natVal (Proxy :: Proxy padRight)
105+ padb = fromIntegral $ natVal (Proxy :: Proxy padBottom)
106+ outr = fromIntegral $ natVal (Proxy :: Proxy outputRows)
107+ outc = fromIntegral $ natVal (Proxy :: Proxy outputColumns)
108+ inr = fromIntegral $ natVal (Proxy :: Proxy inputRows)
109+ inc = fromIntegral $ natVal (Proxy :: Proxy inputColumns)
110+ ch = fromIntegral $ natVal (Proxy :: Proxy channels)
111+ m = extract input
112+ padded = pad ch padl padt padr padb inr inc outr outc m
113+ in ((), S3D . fromJust . create $ padded)
114+
115+ runBackwards Pad () (S3D gradient) =
116+ let padl = fromIntegral $ natVal (Proxy :: Proxy padLeft)
117+ padt = fromIntegral $ natVal (Proxy :: Proxy padTop)
118+ padr = fromIntegral $ natVal (Proxy :: Proxy padRight)
119+ padb = fromIntegral $ natVal (Proxy :: Proxy padBottom)
120+ outr = fromIntegral $ natVal (Proxy :: Proxy outputRows)
121+ outc = fromIntegral $ natVal (Proxy :: Proxy outputColumns)
122+ inr = fromIntegral $ natVal (Proxy :: Proxy inputRows)
123+ inc = fromIntegral $ natVal (Proxy :: Proxy inputColumns)
124+ ch = fromIntegral $ natVal (Proxy :: Proxy channels)
125+ m = extract gradient
126+ cropped = crop ch padl padt padr padb inr inc outr outc m
127+ in ((), S3D . fromJust . create $ cropped)
src/Grenade/Layers/Pooling.hs+15−5
@@ -1,3 +1,4 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE ScopedTypeVariables #-}
3{-# LANGUAGE StandaloneDeriving #-}
4{-# LANGUAGE StandaloneDeriving #-}
@@ -6,6 +7,14 @@
6{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE MultiParamTypeClasses #-}
8{-# LANGUAGE MultiParamTypeClasses #-}
8{-# LANGUAGE FlexibleContexts #-}
9{-# LANGUAGE FlexibleContexts #-}
10+{-# LANGUAGE UndecidableInstances #-}
11+{-|
12+Module : Grenade.Core.Pooling
13+Description : Max Pooling layer for 2D and 3D images
14+Copyright : (c) Huw Campbell, 2016-2017
15+License : BSD2
16+Stability : experimental
17+-}
9module Grenade.Layers.Pooling (
18module Grenade.Layers.Pooling (
10 Pooling (..)
19 Pooling (..)
11 ) where
20 ) where
@@ -13,30 +22,31 @@ module Grenade.Layers.Pooling (
13import Data.Maybe
22import Data.Maybe
14import Data.Proxy
23import Data.Proxy
15import Data.Serialize
24import Data.Serialize
16−import Data.Singletons.TypeLits
25+
17import GHC.TypeLits
26import GHC.TypeLits
27+import Data.Kind (Type)
18
28
19−import Grenade.Core.Network
29+import Grenade.Core
20−import Grenade.Core.Shape
21import Grenade.Layers.Internal.Pooling
30import Grenade.Layers.Internal.Pooling
22
31
23import Numeric.LinearAlgebra.Static as LAS hiding ((|||), build, toRows)
32import Numeric.LinearAlgebra.Static as LAS hiding ((|||), build, toRows)
24
33
25-- | A pooling layer for a neural network.
34-- | A pooling layer for a neural network.
35+--
26-- Does a max pooling, looking over a kernel similarly to the convolution network, but returning
36-- Does a max pooling, looking over a kernel similarly to the convolution network, but returning
27-- maxarg only. This layer is often used to provide minor amounts of translational invariance.
37-- maxarg only. This layer is often used to provide minor amounts of translational invariance.
28--
38--
29-- The kernel size dictates which input and output sizes will "fit". Fitting the equation:
39-- The kernel size dictates which input and output sizes will "fit". Fitting the equation:
30-- `out = (in - kernel) / stride + 1` for both dimensions.
40-- `out = (in - kernel) / stride + 1` for both dimensions.
31--
41--
32−data Pooling :: Nat -> Nat -> Nat -> Nat -> * where
42+data Pooling :: Nat -> Nat -> Nat -> Nat -> Type where
33 Pooling :: Pooling kernelRows kernelColumns strideRows strideColumns
43 Pooling :: Pooling kernelRows kernelColumns strideRows strideColumns
34
44
35instance Show (Pooling k k' s s') where
45instance Show (Pooling k k' s s') where
36 show Pooling = "Pooling"
46 show Pooling = "Pooling"
37
47
38instance UpdateLayer (Pooling kernelRows kernelColumns strideRows strideColumns) where
48instance UpdateLayer (Pooling kernelRows kernelColumns strideRows strideColumns) where
39− type Gradient (Pooling kr kc sr sc) = ()
49+ type Gradient (Pooling kernelRows kernelColumns strideRows strideColumns) = ()
40 runUpdate _ Pooling _ = Pooling
50 runUpdate _ Pooling _ = Pooling
41 createRandom = return Pooling
51 createRandom = return Pooling
42
52
src/Grenade/Layers/Relu.hs+8−2
@@ -2,6 +2,13 @@
2{-# LANGUAGE TypeOperators #-}
2{-# LANGUAGE TypeOperators #-}
3{-# LANGUAGE TypeFamilies #-}
3{-# LANGUAGE TypeFamilies #-}
4{-# LANGUAGE MultiParamTypeClasses #-}
4{-# LANGUAGE MultiParamTypeClasses #-}
5+{-|
6+Module : Grenade.Layers.Relu
7+Description : Rectifying linear unit layer
8+Copyright : (c) Huw Campbell, 2016-2017
9+License : BSD2
10+Stability : experimental
11+-}
5module Grenade.Layers.Relu (
12module Grenade.Layers.Relu (
6 Relu (..)
13 Relu (..)
7 ) where
14 ) where
@@ -9,8 +16,7 @@ module Grenade.Layers.Relu (
9import Data.Serialize
16import Data.Serialize
10
17
11import GHC.TypeLits
18import GHC.TypeLits
12−import Grenade.Core.Network
19+import Grenade.Core
13−import Grenade.Core.Shape
14
20
15import qualified Numeric.LinearAlgebra.Static as LAS
21import qualified Numeric.LinearAlgebra.Static as LAS
16
22
src/Grenade/Layers/Reshape.hsnew file+80
@@ -0,0 +1,80 @@
1+{-# LANGUAGE DataKinds #-}
2+{-# LANGUAGE TypeOperators #-}
3+{-# LANGUAGE TypeFamilies #-}
4+{-# LANGUAGE MultiParamTypeClasses #-}
5+{-# LANGUAGE FlexibleContexts #-}
6+{-# LANGUAGE UndecidableInstances #-}
7+{-|
8+Module : Grenade.Layers.Reshape
9+Description : Multipurpose reshaping layer
10+Copyright : (c) Huw Campbell, 2016-2017
11+License : BSD2
12+Stability : experimental
13+-}
14+module Grenade.Layers.Reshape (
15+ Reshape (..)
16+ ) where
17+
18+import Data.Serialize
19+
20+import GHC.TypeLits
21+
22+import Numeric.LinearAlgebra.Static
23+import Numeric.LinearAlgebra.Data as LA ( flatten )
24+
25+import Grenade.Core
26+
27+-- | Reshape Layer
28+--
29+-- The Reshape layer can flatten any 2D or 3D image to 1D vector with the
30+-- same number of activations, as well as cast up from 1D to a 2D or 3D
31+-- shape.
32+--
33+-- Can also be used to turn a 3D image with only one channel into a 2D image
34+-- or vice versa.
35+data Reshape = Reshape
36+ deriving Show
37+
38+instance UpdateLayer Reshape where
39+ type Gradient Reshape = ()
40+ runUpdate _ _ _ = Reshape
41+ createRandom = return Reshape
42+
43+instance (KnownNat a, KnownNat x, KnownNat y, a ~ (x * y)) => Layer Reshape ('D2 x y) ('D1 a) where
44+ type Tape Reshape ('D2 x y) ('D1 a) = ()
45+ runForwards _ (S2D y) = ((), fromJust' . fromStorable . flatten . extract $ y)
46+ runBackwards _ _ (S1D y) = ((), fromJust' . fromStorable . extract $ y)
47+
48+instance (KnownNat a, KnownNat x, KnownNat y, KnownNat (x * z), KnownNat z, a ~ (x * y * z)) => Layer Reshape ('D3 x y z) ('D1 a) where
49+ type Tape Reshape ('D3 x y z) ('D1 a) = ()
50+ runForwards _ (S3D y) = ((), fromJust' . fromStorable . flatten . extract $ y)
51+ runBackwards _ _ (S1D y) = ((), fromJust' . fromStorable . extract $ y)
52+
53+instance (KnownNat y, KnownNat x, KnownNat z, z ~ 1) => Layer Reshape ('D3 x y z) ('D2 x y) where
54+ type Tape Reshape ('D3 x y z) ('D2 x y) = ()
55+ runForwards _ (S3D y) = ((), S2D y)
56+ runBackwards _ _ (S2D y) = ((), S3D y)
57+
58+instance (KnownNat y, KnownNat x, KnownNat z, z ~ 1) => Layer Reshape ('D2 x y) ('D3 x y z) where
59+ type Tape Reshape ('D2 x y) ('D3 x y z) = ()
60+ runForwards _ (S2D y) = ((), S3D y)
61+ runBackwards _ _ (S3D y) = ((), S2D y)
62+
63+instance (KnownNat a, KnownNat x, KnownNat y, a ~ (x * y)) => Layer Reshape ('D1 a) ('D2 x y) where
64+ type Tape Reshape ('D1 a) ('D2 x y) = ()
65+ runForwards _ (S1D y) = ((), fromJust' . fromStorable . extract $ y)
66+ runBackwards _ _ (S2D y) = ((), fromJust' . fromStorable . flatten . extract $ y)
67+
68+instance (KnownNat a, KnownNat x, KnownNat y, KnownNat (x * z), KnownNat z, a ~ (x * y * z)) => Layer Reshape ('D1 a) ('D3 x y z) where
69+ type Tape Reshape ('D1 a) ('D3 x y z) = ()
70+ runForwards _ (S1D y) = ((), fromJust' . fromStorable . extract $ y)
71+ runBackwards _ _ (S3D y) = ((), fromJust' . fromStorable . flatten . extract $ y)
72+
73+instance Serialize Reshape where
74+ put _ = return ()
75+ get = return Reshape
76+
77+
78+fromJust' :: Maybe x -> x
79+fromJust' (Just x) = x
80+fromJust' Nothing = error $ "Reshape error: data shape couldn't be converted."
src/Grenade/Layers/Sinusoid.hsnew file+39
@@ -0,0 +1,39 @@
1+{-# LANGUAGE DataKinds #-}
2+{-# LANGUAGE FlexibleInstances #-}
3+{-# LANGUAGE MultiParamTypeClasses #-}
4+{-# LANGUAGE TypeFamilies #-}
5+{-# LANGUAGE TypeOperators #-}
6+{-|
7+Module : Grenade.Layers.Sinusoid
8+Description : Sinusoid nonlinear layer
9+Copyright : (c) Manuel Schneckenreither, 2018
10+License : BSD2
11+Stability : experimental
12+-}
13+module Grenade.Layers.Sinusoid (
14+ Sinusoid (..)
15+ ) where
16+
17+import Data.Serialize
18+import Data.Singletons
19+
20+import Grenade.Core
21+
22+-- | A Sinusoid layer.
23+-- A layer which can act between any shape of the same dimension, performing a sin function.
24+data Sinusoid = Sinusoid
25+ deriving Show
26+
27+instance UpdateLayer Sinusoid where
28+ type Gradient Sinusoid = ()
29+ runUpdate _ _ _ = Sinusoid
30+ createRandom = return Sinusoid
31+
32+instance Serialize Sinusoid where
33+ put _ = return ()
34+ get = return Sinusoid
35+
36+instance (a ~ b, SingI a) => Layer Sinusoid a b where
37+ type Tape Sinusoid a b = S a
38+ runForwards _ a = (a, sin a)
39+ runBackwards _ a g = ((), cos a * g)
src/Grenade/Layers/Softmax.hsnew file+64
@@ -0,0 +1,64 @@
1+{-# LANGUAGE DataKinds #-}
2+{-# LANGUAGE TypeOperators #-}
3+{-# LANGUAGE TypeFamilies #-}
4+{-# LANGUAGE FlexibleInstances #-}
5+{-# LANGUAGE MultiParamTypeClasses #-}
6+{-|
7+Module : Grenade.Core.Softmax
8+Description : Softmax loss layer
9+Copyright : (c) Huw Campbell, 2016-2017
10+License : BSD2
11+Stability : experimental
12+-}
13+module Grenade.Layers.Softmax (
14+ Softmax (..)
15+ , softmax
16+ , softmax'
17+ ) where
18+
19+import Data.Serialize
20+
21+import GHC.TypeLits
22+import Grenade.Core
23+
24+import Numeric.LinearAlgebra.Static as LAS
25+
26+-- | A Softmax layer
27+--
28+-- This layer is like a logit layer, but normalises
29+-- a set of matricies to be probabilities.
30+--
31+-- One can use this layer as the last layer in a network
32+-- if they need normalised probabilities.
33+data Softmax = Softmax
34+ deriving Show
35+
36+instance UpdateLayer Softmax where
37+ type Gradient Softmax = ()
38+ runUpdate _ _ _ = Softmax
39+ createRandom = return Softmax
40+
41+instance ( KnownNat i ) => Layer Softmax ('D1 i) ('D1 i) where
42+ type Tape Softmax ('D1 i) ('D1 i) = S ('D1 i)
43+
44+ runForwards _ (S1D y) = (S1D y, S1D (softmax y))
45+ runBackwards _ (S1D y) (S1D dEdy) = ((), S1D (softmax' y dEdy))
46+
47+instance Serialize Softmax where
48+ put _ = return ()
49+ get = return Softmax
50+
51+softmax :: KnownNat i => LAS.R i -> LAS.R i
52+softmax xs =
53+ let xs' = LAS.dvmap exp xs
54+ s = LAS.dot xs' 1
55+ in LAS.dvmap (/ s) xs'
56+
57+softmax' :: KnownNat i => LAS.R i -> LAS.R i -> LAS.R i
58+softmax' x grad =
59+ let yTy = outer sm sm
60+ d = diag sm
61+ g = d - yTy
62+ in g #> grad
63+ where
64+ sm = softmax x
src/Grenade/Layers/Tanh.hs+12−6
@@ -1,8 +1,15 @@
1{-# LANGUAGE DataKinds #-}
1{-# LANGUAGE DataKinds #-}
2−{-# LANGUAGE TypeOperators #-}
3−{-# LANGUAGE TypeFamilies #-}
4{-# LANGUAGE FlexibleInstances #-}
2{-# LANGUAGE FlexibleInstances #-}
5{-# LANGUAGE MultiParamTypeClasses #-}
3{-# LANGUAGE MultiParamTypeClasses #-}
4+{-# LANGUAGE TypeFamilies #-}
5+{-# LANGUAGE TypeOperators #-}
6+{-|
7+Module : Grenade.Layers.Tanh
8+Description : Hyperbolic tangent nonlinear layer
9+Copyright : (c) Huw Campbell, 2016-2017
10+License : BSD2
11+Stability : experimental
12+-}
6module Grenade.Layers.Tanh (
13module Grenade.Layers.Tanh (
7 Tanh (..)
14 Tanh (..)
8 ) where
15 ) where
@@ -10,11 +17,10 @@ module Grenade.Layers.Tanh (
10import Data.Serialize
17import Data.Serialize
11import Data.Singletons
18import Data.Singletons
12
19
13−import Grenade.Core.Network
20+import Grenade.Core
14−import Grenade.Core.Shape
15
21
16-- | A Tanh layer.
22-- | A Tanh layer.
17−-- A layer which can act between any shape of the same dimension, perfoming a tanh function.
23+-- A layer which can act between any shape of the same dimension, performing a tanh function.
18data Tanh = Tanh
24data Tanh = Tanh
19 deriving Show
25 deriving Show
20
26
@@ -27,7 +33,7 @@ instance Serialize Tanh where
27 put _ = return ()
33 put _ = return ()
28 get = return Tanh
34 get = return Tanh
29
35
30−instance (a ~ b, SingI a, SingI b) => Layer Tanh a b where
36+instance (a ~ b, SingI a) => Layer Tanh a b where
31 type Tape Tanh a b = S a
37 type Tape Tanh a b = S a
32 runForwards _ a = (a, tanh a)
38 runForwards _ a = (a, tanh a)
33 runBackwards _ a g = ((), tanh' a * g)
39 runBackwards _ a g = ((), tanh' a * g)
src/Grenade/Recurrent/Layers/Trivial.hs → src/Grenade/Layers/Trivial.hsrenamed+14−4
@@ -3,15 +3,25 @@
3{-# LANGUAGE TypeFamilies #-}
3{-# LANGUAGE TypeFamilies #-}
4{-# LANGUAGE MultiParamTypeClasses #-}
4{-# LANGUAGE MultiParamTypeClasses #-}
5{-# LANGUAGE FlexibleInstances #-}
5{-# LANGUAGE FlexibleInstances #-}
6+{-|
7+Module : Grenade.Core.Trivial
8+Description : Trivial layer which perfoms no operations on the data
9+Copyright : (c) Huw Campbell, 2016-2017
10+License : BSD2
11+Stability : experimental
12+-}
6−module Grenade.Recurrent.Layers.Trivial (
13+module Grenade.Layers.Trivial (
7 Trivial (..)
14 Trivial (..)
8 ) where
15 ) where
9
16
10import Data.Serialize
17import Data.Serialize
11
18
12−import Grenade.Core.Network
19+import Grenade.Core
13
20
14−-- | A trivial layer.
21+-- | A Trivial layer.
22+--
23+-- This can be used to pass an unchanged value up one side of a
24+-- graph, for a Residual network for example.
15data Trivial = Trivial
25data Trivial = Trivial
16 deriving Show
26 deriving Show
17
27
@@ -26,5 +36,5 @@ instance UpdateLayer Trivial where
26
36
27instance (a ~ b) => Layer Trivial a b where
37instance (a ~ b) => Layer Trivial a b where
28 type Tape Trivial a b = ()
38 type Tape Trivial a b = ()
29− runForwards _ x = ((), x)
39+ runForwards _ a = ((), a)
30 runBackwards _ _ y = ((), y)
40 runBackwards _ _ y = ((), y)
src/Grenade/Recurrent.hs+26−6
@@ -1,9 +1,29 @@
1module Grenade.Recurrent (
1module Grenade.Recurrent (
2− module X
2+ -- | This is an empty module which simply re-exports public definitions
3+ -- for recurrent networks in Grenade.
4+
5+ -- * Exported modules
6+ --
7+ -- | The core types and runners for Recurrent Networks.
8+ module Grenade.Recurrent.Core
9+
10+ -- | The recurrent neural network layer zoo
11+ , module Grenade.Recurrent.Layers
12+
13+ -- * Overview of recurrent Networks
14+ -- $recurrent
15+
3 ) where
16 ) where
4
17
5−import Grenade.Recurrent.Core.Network as X
18+import Grenade.Recurrent.Core
6−import Grenade.Recurrent.Core.Runner as X
19+import Grenade.Recurrent.Layers
7−import Grenade.Recurrent.Layers.BasicRecurrent as X
20+
8−import Grenade.Recurrent.Layers.LSTM as X
21+{- $recurrent
9−import Grenade.Recurrent.Layers.Trivial as X
22+There are two ways in which deep learning libraries choose to represent
23+recurrent Neural Networks, as an unrolled graph, or at a first class
24+level. Grenade chooses the latter representation, and provides a network
25+type which is specifically suited for recurrent neural networks.
26+
27+Currently grenade supports two layers, a basic recurrent layer, and an
28+LSTM layer.
29+-}
src/Grenade/Recurrent/Core.hsnew file+9
@@ -0,0 +1,9 @@
1+module Grenade.Recurrent.Core (
2+ module Grenade.Recurrent.Core.Layer
3+ , module Grenade.Recurrent.Core.Network
4+ , module Grenade.Recurrent.Core.Runner
5+ ) where
6+
7+import Grenade.Recurrent.Core.Layer
8+import Grenade.Recurrent.Core.Network
9+import Grenade.Recurrent.Core.Runner
src/Grenade/Recurrent/Core/Layer.hsnew file+35
@@ -0,0 +1,35 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE DataKinds #-}
3+{-# LANGUAGE TypeFamilies #-}
4+{-# LANGUAGE MultiParamTypeClasses #-}
5+{-# LANGUAGE FlexibleContexts #-}
6+{-# LANGUAGE FlexibleInstances #-}
7+module Grenade.Recurrent.Core.Layer (
8+ RecurrentLayer (..)
9+ , RecurrentUpdateLayer (..)
10+ ) where
11+
12+#if MIN_VERSION_base(4,9,0)
13+import Data.Kind (Type)
14+#endif
15+
16+import Grenade.Core
17+
18+-- | Class for a recurrent layer.
19+-- It's quite similar to a normal layer but for the input and output
20+-- of an extra recurrent data shape.
21+class UpdateLayer x => RecurrentUpdateLayer x where
22+ -- | Shape of data that is passed between each subsequent run of the layer
23+ type RecurrentShape x :: Type
24+
25+class (RecurrentUpdateLayer x, Num (RecurrentShape x)) => RecurrentLayer x (i :: Shape) (o :: Shape) where
26+ -- | Wengert Tape
27+ type RecTape x i o :: Type
28+ -- | Used in training and scoring. Take the input from the previous
29+ -- layer, and give the output from this layer.
30+ runRecurrentForwards :: x -> RecurrentShape x -> S i -> (RecTape x i o, RecurrentShape x, S o)
31+ -- | Back propagate a step. Takes the current layer, the input that the
32+ -- layer gave from the input and the back propagated derivatives from
33+ -- the layer above.
34+ -- Returns the gradient layer and the derivatives to push back further.
35+ runRecurrentBackwards :: x -> RecTape x i o -> RecurrentShape x -> S o -> (Gradient x, RecurrentShape x, S i)
src/Grenade/Recurrent/Core/Network.hs+272−82
@@ -1,3 +1,4 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE GADTs #-}
3{-# LANGUAGE GADTs #-}
3{-# LANGUAGE TypeOperators #-}
4{-# LANGUAGE TypeOperators #-}
@@ -6,133 +7,322 @@
6{-# LANGUAGE FlexibleContexts #-}
7{-# LANGUAGE FlexibleContexts #-}
7{-# LANGUAGE FlexibleInstances #-}
8{-# LANGUAGE FlexibleInstances #-}
8{-# LANGUAGE EmptyDataDecls #-}
9{-# LANGUAGE EmptyDataDecls #-}
10+{-# LANGUAGE RankNTypes #-}
11+{-# LANGUAGE BangPatterns #-}
12+{-# LANGUAGE ScopedTypeVariables #-}
13+{-# LANGUAGE UndecidableInstances #-}
14+
9module Grenade.Recurrent.Core.Network (
15module Grenade.Recurrent.Core.Network (
10 Recurrent
16 Recurrent
11 , FeedForward
17 , FeedForward
12− , RecurrentLayer (..)
18+
13− , RecurrentUpdateLayer (..)
14 , RecurrentNetwork (..)
19 , RecurrentNetwork (..)
15 , RecurrentInputs (..)
20 , RecurrentInputs (..)
16− , CreatableRecurrent (..)
21+ , RecurrentTape (..)
22+ , RecurrentGradient (..)
23+
24+ , randomRecurrent
25+ , runRecurrent
26+ , runRecurrent'
27+ , applyRecurrentUpdate
17 ) where
28 ) where
18
29
19
30
20import Control.Monad.Random ( MonadRandom )
31import Control.Monad.Random ( MonadRandom )
21−import Data.Singletons ( SingI )
22import Data.Serialize
32import Data.Serialize
23−import qualified Data.Vector.Storable as V
24
33
25−import Grenade.Core.Shape
34+import Data.Kind (Type)
26−import Grenade.Core.Network
27−
28−import qualified Numeric.LinearAlgebra as LA
29−import qualified Numeric.LinearAlgebra.Static as LAS
30
35
36+import Grenade.Core
37+import Grenade.Recurrent.Core.Layer
38+import Prelude.Singletons
31
39
32-- | Witness type to say indicate we're building up with a normal feed
40-- | Witness type to say indicate we're building up with a normal feed
33-- forward layer.
41-- forward layer.
34−data FeedForward :: * -> *
42+data FeedForward :: Type -> Type
35-- | Witness type to say indicate we're building up with a recurrent layer.
43-- | Witness type to say indicate we're building up with a recurrent layer.
36−data Recurrent :: * -> *
44+data Recurrent :: Type -> Type
37−
45+
38−-- | Class for a recurrent layer.
46+-- | Type of a recurrent neural network.
39−-- It's quite similar to a normal layer but for the input and output
47+--
40−-- of an extra recurrent data shape.
48+-- The [Type] type specifies the types of the layers.
41−class UpdateLayer x => RecurrentUpdateLayer x where
49+--
42− -- | Shape of data that is passed between each subsequent run of the layer
50+-- The [Shape] type specifies the shapes of data passed between the layers.
43− type RecurrentShape x :: Shape
51+--
44−
52+-- The definition is similar to a Network, but every layer in the
45−class (RecurrentUpdateLayer x, SingI (RecurrentShape x)) => RecurrentLayer x (i :: Shape) (o :: Shape) where
53+-- type is tagged by whether it's a FeedForward Layer of a Recurrent layer.
46− -- | Used in training and scoring. Take the input from the previous
54+--
47− -- layer, and give the output from this layer.
55+-- Often, to make the definitions more concise, one will use a type alias
48− runRecurrentForwards :: x -> S (RecurrentShape x) -> S i -> (S (RecurrentShape x), S o)
56+-- for these empty data types.
49− -- | Back propagate a step. Takes the current layer, the input that the
50− -- layer gave from the input and the back propagated derivatives from
51− -- the layer above.
52− -- Returns the gradient layer and the derivatives to push back further.
53− runRecurrentBackwards :: x -> S (RecurrentShape x) -> S i -> S (RecurrentShape x) -> S o -> (Gradient x, S (RecurrentShape x), S i)
54−
55−data RecurrentNetwork :: [*] -> [Shape] -> * where
57+data RecurrentNetwork :: [Type] -> [Shape] -> Type where
58+ RNil :: SingI i
56− OR :: (SingI i, SingI o, Layer x i o) => !x -> RecurrentNetwork '[FeedForward x] '[i, o]
59+ => RecurrentNetwork '[] '[i]
60+
61+ (:~~>) :: (SingI i, Layer x i h)
62+ => !x
63+ -> !(RecurrentNetwork xs (h ': hs))
57− (:~~>) :: (SingI i, Layer x i h) => !x -> !(RecurrentNetwork xs (h ': hs)) -> RecurrentNetwork (FeedForward x ': xs) (i ': h ': hs)
64+ -> RecurrentNetwork (FeedForward x ': xs) (i ': h ': hs)
65+
66+ (:~@>) :: (SingI i, RecurrentLayer x i h)
67+ => !x
68+ -> !(RecurrentNetwork xs (h ': hs))
58− (:~@>) :: (SingI i, RecurrentLayer x i h) => !x -> !(RecurrentNetwork xs (h ': hs)) -> RecurrentNetwork (Recurrent x ': xs) (i ': h ': hs)
69+ -> RecurrentNetwork (Recurrent x ': xs) (i ': h ': hs)
59infixr 5 :~~>
70infixr 5 :~~>
60infixr 5 :~@>
71infixr 5 :~@>
61
72
62−instance Show (RecurrentNetwork l h) where
73+-- | Gradient of a network.
63− show (OR a) = "OR " ++ show a
74+--
64− show (i :~~> o) = show i ++ "\n:~~>\n" ++ show o
75+-- Parameterised on the layers of the network.
65− show (i :~@> o) = show i ++ "\n:~@>\n" ++ show o
76+data RecurrentGradient :: [Type] -> Type where
77+ RGNil :: RecurrentGradient '[]
66
78
79+ (://>) :: UpdateLayer x
80+ => Gradient x
81+ -> RecurrentGradient xs
82+ -> RecurrentGradient (phantom x ': xs)
67
83
68-- | Recurrent inputs (sideways shapes on an imaginary unrolled graph)
84-- | Recurrent inputs (sideways shapes on an imaginary unrolled graph)
69-- Parameterised on the layers of a Network.
85-- Parameterised on the layers of a Network.
70−data RecurrentInputs :: [*] -> * where
86+data RecurrentInputs :: [Type] -> Type where
71− ORS :: UpdateLayer x
87+ RINil :: RecurrentInputs '[]
72− => () -> RecurrentInputs '[FeedForward x]
88+
73− (:~~+>) :: UpdateLayer x
89+ (:~~+>) :: (UpdateLayer x, Fractional (RecurrentInputs xs))
74− => () -> !(RecurrentInputs xs) -> RecurrentInputs (FeedForward x ': xs)
90+ => () -> !(RecurrentInputs xs) -> RecurrentInputs (FeedForward x ': xs)
75− (:~@+>) :: (SingI (RecurrentShape x), RecurrentUpdateLayer x)
91+
76− => !(S (RecurrentShape x)) -> !(RecurrentInputs xs) -> RecurrentInputs (Recurrent x ': xs)
92+ (:~@+>) :: (Fractional (RecurrentShape x), Fractional (RecurrentInputs xs), RecurrentUpdateLayer x)
77−infixr 5 :~~+>
93+ => !(RecurrentShape x) -> !(RecurrentInputs xs) -> RecurrentInputs (Recurrent x ': xs)
78−infixr 5 :~@+>
94+
95+-- | All the information required to backpropogate
96+-- through time safely.
97+--
98+-- We index on the time step length as well, to ensure
99+-- that that all Tape lengths are the same.
100+data RecurrentTape :: [Type] -> [Shape] -> Type where
101+ TRNil :: SingI i
102+ => RecurrentTape '[] '[i]
103+
104+ (:\~>) :: Tape x i h
105+ -> !(RecurrentTape xs (h ': hs))
106+ -> RecurrentTape (FeedForward x ': xs) (i ': h ': hs)
107+
108+ (:\@>) :: RecTape x i h
109+ -> !(RecurrentTape xs (h ': hs))
110+ -> RecurrentTape (Recurrent x ': xs) (i ': h ': hs)
111+
112+
113+runRecurrent :: forall shapes layers.
114+ RecurrentNetwork layers shapes
115+ -> RecurrentInputs layers
116+ -> S (Head shapes)
117+ -> (RecurrentTape layers shapes, RecurrentInputs layers, S (Last shapes))
118+runRecurrent =
119+ go
120+ where
121+ go :: forall js sublayers. (Last js ~ Last shapes)
122+ => RecurrentNetwork sublayers js
123+ -> RecurrentInputs sublayers
124+ -> S (Head js)
125+ -> (RecurrentTape sublayers js, RecurrentInputs sublayers, S (Last js))
126+ go (!layer :~~> n) (() :~~+> nIn) !x
127+ = let (!tape, !forwards) = runForwards layer x
128+
129+ -- recursively run the rest of the network, and get the gradients from above.
130+ (!newFN, !ig, !answer) = go n nIn forwards
131+ in (tape :\~> newFN, () :~~+> ig, answer)
132+
133+ -- This is a recurrent layer, so we need to do a scan, first input to last, providing
134+ -- the recurrent shape output to the next layer.
135+ go (layer :~@> n) (recIn :~@+> nIn) !x
136+ = let (tape, shape, forwards) = runRecurrentForwards layer recIn x
137+ (newFN, ig, answer) = go n nIn forwards
138+ in (tape :\@> newFN, shape :~@+> ig, answer)
139+
140+ -- Handle the output layer, bouncing the derivatives back down.
141+ -- We may not have a target for each example, so when we don't use 0 gradient.
142+ go RNil RINil !x
143+ = (TRNil, RINil, x)
144+
145+runRecurrent' :: forall layers shapes.
146+ RecurrentNetwork layers shapes
147+ -> RecurrentTape layers shapes
148+ -> RecurrentInputs layers
149+ -> S (Last shapes)
150+ -> (RecurrentGradient layers, RecurrentInputs layers, S (Head shapes))
151+runRecurrent' net tapes r o =
152+ go net tapes r
153+ where
154+ -- We have to be careful regarding the direction of the lists
155+ -- Inputs come in forwards, but our return value is backwards
156+ -- through time.
157+ go :: forall js ss. (Last js ~ Last shapes)
158+ => RecurrentNetwork ss js
159+ -> RecurrentTape ss js
160+ -> RecurrentInputs ss
161+ -> (RecurrentGradient ss, RecurrentInputs ss, S (Head js))
162+ -- This is a simple non-recurrent layer
163+ -- Run the rest of the network, update with the tapes and gradients
164+ go (!layer :~~> n) (!tape :\~> nTapes) (() :~~+> nRecs) =
165+ let (!gradients, !rins, !feed) = go n nTapes nRecs
166+ (!grad, !back) = runBackwards layer tape feed
167+ in (grad ://> gradients, () :~~+> rins, back)
168+
169+ -- This is a recurrent layer
170+ -- Run the rest of the network, scan over the tapes in reverse
171+ go (!layer :~@> n) (!tape :\@> nTapes) (!recGrad :~@+> nRecs) =
172+ let (!gradients, !rins, !feed) = go n nTapes nRecs
173+ (!grad, !sidegrad, !back) = runRecurrentBackwards layer tape recGrad feed
174+ in (grad ://> gradients, sidegrad :~@+> rins, back)
175+
176+ -- End of the road, so we reflect the given gradients backwards.
177+ -- Crucially, we reverse the list, so it's backwards in time as
178+ -- well.
179+ go !RNil !TRNil !RINil
180+ = (RGNil, RINil, o)
181+
182+-- | Apply a batch of gradients to the network
183+-- Uses runUpdates which can be specialised for
184+-- a layer.
185+applyRecurrentUpdate :: LearningParameters
186+ -> RecurrentNetwork layers shapes
187+ -> RecurrentGradient layers
188+ -> RecurrentNetwork layers shapes
189+applyRecurrentUpdate rate (layer :~~> rest) (gradient ://> grest)
190+ = runUpdate rate layer gradient :~~> applyRecurrentUpdate rate rest grest
191+
192+applyRecurrentUpdate rate (layer :~@> rest) (gradient ://> grest)
193+ = runUpdate rate layer gradient :~@> applyRecurrentUpdate rate rest grest
194+
195+applyRecurrentUpdate _ RNil RGNil
196+ = RNil
197+
198+
199+instance Show (RecurrentNetwork '[] '[i]) where
200+ show RNil = "NNil"
201+instance (Show x, Show (RecurrentNetwork xs rs)) => Show (RecurrentNetwork (FeedForward x ': xs) (i ': rs)) where
202+ show (x :~~> xs) = show x ++ "\n~~>\n" ++ show xs
203+instance (Show x, Show (RecurrentNetwork xs rs)) => Show (RecurrentNetwork (Recurrent x ': xs) (i ': rs)) where
204+ show (x :~@> xs) = show x ++ "\n~~>\n" ++ show xs
205+
79
206
80-- | A network can easily be created by hand with (:~~>) and (:~@>), but an easy way to initialise a random
207-- | A network can easily be created by hand with (:~~>) and (:~@>), but an easy way to initialise a random
81-- recurrent network and a set of random inputs for it is with the randomRecurrent.
208-- recurrent network and a set of random inputs for it is with the randomRecurrent.
82−class CreatableRecurrent (xs :: [*]) (ss :: [Shape]) where
209+class CreatableRecurrent (xs :: [Type]) (ss :: [Shape]) where
83 -- | Create a network of the types requested
210 -- | Create a network of the types requested
84− randomRecurrent :: MonadRandom m => m (RecurrentNetwork xs ss, RecurrentInputs xs)
211+ randomRecurrent :: MonadRandom m => m (RecurrentNetwork xs ss)
85
212
86−instance (SingI i, SingI o, Layer x i o) => CreatableRecurrent (FeedForward x ': '[]) (i ': o ': '[]) where
213+instance SingI i => CreatableRecurrent '[] '[i] where
87− randomRecurrent = do
214+ randomRecurrent =
88− thisLayer <- createRandom
89− return (OR thisLayer, ORS ())
215+ return RNil
90
216
91−instance (SingI i, Layer x i o, CreatableRecurrent xs (o ': r ': rs)) => CreatableRecurrent (FeedForward x ': xs) (i ': o ': r ': rs) where
217+instance (SingI i, Layer x i o, CreatableRecurrent xs (o ': rs)) => CreatableRecurrent (FeedForward x ': xs) (i ': o ': rs) where
92 randomRecurrent = do
218 randomRecurrent = do
93 thisLayer <- createRandom
219 thisLayer <- createRandom
94− (rest, resti) <- randomRecurrent
220+ rest <- randomRecurrent
95− return (thisLayer :~~> rest, () :~~+> resti)
221+ return (thisLayer :~~> rest)
96
222
97−instance (SingI i, RecurrentLayer x i o, CreatableRecurrent xs (o ': r ': rs)) => CreatableRecurrent (Recurrent x ': xs) (i ': o ': r ': rs) where
223+instance (SingI i, RecurrentLayer x i o, CreatableRecurrent xs (o ': rs)) => CreatableRecurrent (Recurrent x ': xs) (i ': o ': rs) where
98 randomRecurrent = do
224 randomRecurrent = do
99 thisLayer <- createRandom
225 thisLayer <- createRandom
100− thisShape <- randomOfShape
101− (rest, resti) <- randomRecurrent
226+ rest <- randomRecurrent
102− return (thisLayer :~@> rest, thisShape :~@+> resti)
227+ return (thisLayer :~@> rest)
103
228
104-- | Add very simple serialisation to the recurrent network
229-- | Add very simple serialisation to the recurrent network
105−instance (SingI i, SingI o, Layer x i o, Serialize x) => Serialize (RecurrentNetwork '[FeedForward x] '[i, o]) where
230+instance SingI i => Serialize (RecurrentNetwork '[] '[i]) where
106− put (OR x) = put x
231+ put RNil = pure ()
107− put _ = error "impossible"
232+ get = pure RNil
108− get = OR <$> get
109
233
110−instance (SingI i, Layer x i o, Serialize x, Serialize (RecurrentNetwork xs (o ': r ': rs))) => Serialize (RecurrentNetwork (FeedForward x ': xs) (i ': o ': r ': rs)) where
234+instance (SingI i, Layer x i o, Serialize x, Serialize (RecurrentNetwork xs (o ': rs))) => Serialize (RecurrentNetwork (FeedForward x ': xs) (i ': o ': rs)) where
111 put (x :~~> r) = put x >> put r
235 put (x :~~> r) = put x >> put r
112 get = (:~~>) <$> get <*> get
236 get = (:~~>) <$> get <*> get
113
237
114−instance (SingI i, RecurrentLayer x i o, Serialize x, Serialize (RecurrentNetwork xs (o ': r ': rs))) => Serialize (RecurrentNetwork (Recurrent x ': xs) (i ': o ': r ': rs)) where
238+instance (SingI i, RecurrentLayer x i o, Serialize x, Serialize (RecurrentNetwork xs (o ': rs))) => Serialize (RecurrentNetwork (Recurrent x ': xs) (i ': o ': rs)) where
115 put (x :~@> r) = put x >> put r
239 put (x :~@> r) = put x >> put r
116 get = (:~@>) <$> get <*> get
240 get = (:~@>) <$> get <*> get
117
241
118−instance (UpdateLayer x) => (Serialize (RecurrentInputs '[FeedForward x])) where
242+instance (Serialize (RecurrentInputs '[])) where
119 put _ = return ()
243 put _ = return ()
120− get = return (ORS ())
244+ get = return RINil
121
245
122−instance (UpdateLayer x, Serialize (RecurrentInputs (y ': ys))) => (Serialize (RecurrentInputs (FeedForward x ': y ': ys))) where
246+instance (UpdateLayer x, Serialize (RecurrentInputs ys), Fractional (RecurrentInputs ys)) => (Serialize (RecurrentInputs (FeedForward x ': ys))) where
123 put ( () :~~+> rest) = put rest
247 put ( () :~~+> rest) = put rest
124 get = ( () :~~+> ) <$> get
248 get = ( () :~~+> ) <$> get
125
249
126−instance (SingI (RecurrentShape x), RecurrentUpdateLayer x, Serialize (RecurrentInputs (y ': ys))) => (Serialize (RecurrentInputs (Recurrent x ': y ': ys))) where
250+instance (Serialize (RecurrentShape x), Fractional (RecurrentShape x), RecurrentUpdateLayer x, Serialize (RecurrentInputs ys), Fractional (RecurrentInputs ys)) => (Serialize (RecurrentInputs (Recurrent x ': ys))) where
127− put ( i :~@+> rest ) = do
251+ put ( i :~@+> rest ) = put i >> put rest
128− _ <- (case i of
129− (S1D x) -> putListOf put . LA.toList . LAS.extract $ x
130− (S2D x) -> putListOf put . LA.toList . LA.flatten . LAS.extract $ x
131− (S3D x) -> putListOf put . LA.toList . LA.flatten . LAS.extract $ x
132− ) :: PutM ()
133− put rest
134−
135− get = do
252+ get = (:~@+>) <$> get <*> get
136− Just i <- fromStorable . V.fromList <$> getListOf get
253+
137− rest <- get
254+
138− return ( i :~@+> rest)
255+-- Num instance for `RecurrentInputs layers`
256+-- Not sure if this is really needed, as I only need a `fromInteger 0` at
257+-- the moment for training, to create a null gradient on the recurrent
258+-- edge.
259+--
260+-- It does raise an interesting question though? Is a 0 gradient actually
261+-- the best?
262+--
263+-- I could imaging that weakly push back towards the optimum input could
264+-- help make a more stable generator.
265+instance (Num (RecurrentInputs '[])) where
266+ (+) _ _ = RINil
267+ (-) _ _ = RINil
268+ (*) _ _ = RINil
269+ abs _ = RINil
270+ signum _ = RINil
271+ fromInteger _ = RINil
272+
273+instance (UpdateLayer x, Fractional (RecurrentInputs ys)) => (Num (RecurrentInputs (FeedForward x ': ys))) where
274+ (+) (() :~~+> x) (() :~~+> y) = () :~~+> (x + y)
275+ (-) (() :~~+> x) (() :~~+> y) = () :~~+> (x - y)
276+ (*) (() :~~+> x) (() :~~+> y) = () :~~+> (x * y)
277+ abs (() :~~+> x) = () :~~+> abs x
278+ signum (() :~~+> x) = () :~~+> signum x
279+ fromInteger x = () :~~+> fromInteger x
280+
281+instance (Fractional (RecurrentShape x), RecurrentUpdateLayer x, Fractional (RecurrentInputs ys)) => (Num (RecurrentInputs (Recurrent x ': ys))) where
282+ (+) (x :~@+> x') (y :~@+> y') = (x + y) :~@+> (x' + y')
283+ (-) (x :~@+> x') (y :~@+> y') = (x - y) :~@+> (x' - y')
284+ (*) (x :~@+> x') (y :~@+> y') = (x * y) :~@+> (x' * y')
285+ abs (x :~@+> x') = abs x :~@+> abs x'
286+ signum (x :~@+> x') = signum x :~@+> signum x'
287+ fromInteger x = fromInteger x :~@+> fromInteger x
288+
289+instance (Fractional (RecurrentInputs '[])) where
290+ (/) _ _ = RINil
291+ recip _ = RINil
292+ fromRational _ = RINil
293+
294+instance (UpdateLayer x, Fractional (RecurrentInputs ys)) => (Fractional (RecurrentInputs (FeedForward x ': ys))) where
295+ (/) (() :~~+> x) (() :~~+> y) = () :~~+> (x / y)
296+ recip (() :~~+> x) = () :~~+> recip x
297+ fromRational x = () :~~+> fromRational x
298+
299+instance (Fractional (RecurrentShape x), RecurrentUpdateLayer x, Fractional (RecurrentInputs ys)) => (Fractional (RecurrentInputs (Recurrent x ': ys))) where
300+ (/) (x :~@+> x') (y :~@+> y') = (x / y) :~@+> (x' / y')
301+ recip (x :~@+> x') = recip x :~@+> recip x'
302+ fromRational x = fromRational x :~@+> fromRational x
303+
304+
305+-- | Ultimate composition.
306+--
307+-- This allows a complete network to be treated as a layer in a larger network.
308+instance CreatableRecurrent sublayers subshapes => UpdateLayer (RecurrentNetwork sublayers subshapes) where
309+ type Gradient (RecurrentNetwork sublayers subshapes) = RecurrentGradient sublayers
310+ runUpdate = applyRecurrentUpdate
311+ createRandom = randomRecurrent
312+
313+-- | Ultimate composition.
314+--
315+-- This allows a complete network to be treated as a layer in a larger network.
316+instance CreatableRecurrent sublayers subshapes => RecurrentUpdateLayer (RecurrentNetwork sublayers subshapes) where
317+ type RecurrentShape (RecurrentNetwork sublayers subshapes) = RecurrentInputs sublayers
318+
319+-- | Ultimate composition.
320+--
321+-- This allows a complete network to be treated as a layer in a larger network.
322+instance ( CreatableRecurrent sublayers subshapes
323+ , i ~ (Head subshapes), o ~ (Last subshapes)
324+ , Num (RecurrentShape (RecurrentNetwork sublayers subshapes))
325+ ) => RecurrentLayer (RecurrentNetwork sublayers subshapes) i o where
326+ type RecTape (RecurrentNetwork sublayers subshapes) i o = RecurrentTape sublayers subshapes
327+ runRecurrentForwards = runRecurrent
328+ runRecurrentBackwards = runRecurrent'
src/Grenade/Recurrent/Core/Runner.hs+89−133
@@ -6,149 +6,105 @@
6{-# LANGUAGE TypeFamilies #-}
6{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE FlexibleContexts #-}
7{-# LANGUAGE FlexibleContexts #-}
8{-# LANGUAGE RankNTypes #-}
8{-# LANGUAGE RankNTypes #-}
9−{-# LANGUAGE RecordWildCards #-}
9+
10module Grenade.Recurrent.Core.Runner (
10module Grenade.Recurrent.Core.Runner (
11+ runRecurrentExamples
12+ , runRecurrentBackprop
13+ , backPropagateRecurrent
11− trainRecurrent
14+ , trainRecurrent
12− , runRecurrent
15+
16+ , RecurrentGradients
13 ) where
17 ) where
14
18
15−import Data.Singletons.Prelude
19+import Data.List ( foldl' )
16−import Grenade.Core.Network
17−import Grenade.Core.Shape
20+import Grenade.Core
18
21
19import Grenade.Recurrent.Core.Network
22import Grenade.Recurrent.Core.Network
23+import Prelude.Singletons
24+
25+type RecurrentGradients layers = [RecurrentGradient layers]
26+
27+runRecurrentExamples :: forall shapes layers.
28+ RecurrentNetwork layers shapes
29+ -> RecurrentInputs layers
30+ -> [S (Head shapes)]
31+ -> ([(RecurrentTape layers shapes, S (Last shapes))], RecurrentInputs layers)
32+runRecurrentExamples net =
33+ go
34+ where
35+ go !side [] = ([], side)
36+ go !side (!x : xs) =
37+ let (!tape, !side', !o) = runRecurrent net side x
38+ (!res, !finalSide) = go side' xs
39+ in (( tape, o ) : res, finalSide)
40+
41+runRecurrentBackprop :: forall layers shapes.
42+ RecurrentNetwork layers shapes
43+ -> RecurrentInputs layers
44+ -> [(RecurrentTape layers shapes, S (Last shapes))]
45+ -> ([(RecurrentGradient layers, S (Head shapes))], RecurrentInputs layers)
46+runRecurrentBackprop net =
47+ go
48+ where
49+ go !side [] = ([], side)
50+ go !side ((!tape,!x):xs) =
51+ let (res, !side') = go side xs
52+ (!grad, !finalSide, !o) = runRecurrent' net tape side' x
53+ in (( grad, o ) : res, finalSide)
54+
20
55
21-- | Drive and network and collect its back propogated gradients.
56-- | Drive and network and collect its back propogated gradients.
22−--
57+backPropagateRecurrent :: forall shapes layers. (SingI (Last shapes), Fractional (RecurrentInputs layers))
23−-- TODO: split this nicely into backpropagate and update.
58+ => RecurrentNetwork layers shapes
24−--
59+ -> RecurrentInputs layers
25−-- QUESTION: Should we return a list of gradients or the sum of
60+ -> [(S (Head shapes), Maybe (S (Last shapes)))]
26−-- the gradients? It's different taking into account
61+ -> (RecurrentGradients layers, RecurrentInputs layers)
27−-- momentum and L2.
62+backPropagateRecurrent network recinputs examples =
63+ let (outForwards, _) = runRecurrentExamples network recinputs inputs
64+
65+ backPropagations = zipWith makeError outForwards targets
66+
67+ (outBackwards, input') = runRecurrentBackprop network 0 backPropagations
68+
69+ gradients = fmap fst outBackwards
70+ in (gradients, input')
71+
72+ where
73+
74+ inputs = fst <$> examples
75+ targets = snd <$> examples
76+
77+ makeError :: (x, S (Last shapes)) -> Maybe (S (Last shapes)) -> (x, S (Last shapes))
78+ makeError (x, _) Nothing = (x, 0)
79+ makeError (x, y) (Just t) = (x, y - t)
80+
81+
28−trainRecurrent :: forall shapes layers. SingI (Last shapes)
82+trainRecurrent :: forall shapes layers. (SingI (Last shapes), Fractional (RecurrentInputs layers))
29 => LearningParameters
83 => LearningParameters
30 -> RecurrentNetwork layers shapes
84 -> RecurrentNetwork layers shapes
31 -> RecurrentInputs layers
85 -> RecurrentInputs layers
32 -> [(S (Head shapes), Maybe (S (Last shapes)))]
86 -> [(S (Head shapes), Maybe (S (Last shapes)))]
33 -> (RecurrentNetwork layers shapes, RecurrentInputs layers)
87 -> (RecurrentNetwork layers shapes, RecurrentInputs layers)
34trainRecurrent rate network recinputs examples =
88trainRecurrent rate network recinputs examples =
35− updateBack $ go inputs network recinputs
89+ let (gradients, recinputs') = backPropagateRecurrent network recinputs examples
36− where
90+
37− inputs = fst <$> examples
91+ newInputs = updateRecInputs rate recinputs recinputs'
38− targets = snd <$> examples
92+
39− updateBack (a,recgrad,_) = (a,updateRecInputs rate recinputs recgrad)
93+ newNetwork = foldl' (applyRecurrentUpdate rate) network gradients
40−
94+
41− go :: forall js sublayers. (Last js ~ Last shapes)
95+ in (newNetwork, newInputs)
42− => [S (Head js)] -- ^ input vector
96+
43− -> RecurrentNetwork sublayers js -- ^ network to train
44− -> RecurrentInputs sublayers
45− -> (RecurrentNetwork sublayers js, RecurrentInputs sublayers, [S (Head js)])
46−
47− -- This is a simple non-recurrent layer, just map it forwards
48− -- Note we're doing training here, we could just return a list of gradients
49− -- (and probably will in future).
50− go !xs (layer :~~> n) (() :~~+> nIn)
51− = let tys = runForwards layer <$> xs
52− tapes = fst <$> tys
53− ys = snd <$> tys
54− -- recursively run the rest of the network, and get the gradients from above.
55− (newFN, ig, grads) = go ys n nIn
56− -- calculate the gradient for this layer to pass down,
57− back = uncurry (runBackwards layer) <$> zip (reverse tapes) grads
58− -- the new trained layer.
59− newlayer = runUpdates rate layer (fst <$> back)
60−
61− in (newlayer :~~> newFN, () :~~+> ig, snd <$> back)
62−
63− -- This is a recurrent layer, so we need to do a scan, first input to last, providing
64− -- the recurrent shape output to the next layer.
65− go !xs (layer :~@> n) (g :~@+> nIn)
66− = let ys = scanlFrom layer g xs
67−
68− (newFN, ig, grads) = go (snd <$> ys) n nIn
69−
70− backExamples = zip3 (fst <$> reverse ys) (reverse xs) grads
71−
72− (rg, back) = myscanbackward layer backExamples
73− -- the new trained layer.
74− newlayer = runUpdates rate layer (fst <$> back)
75− in (newlayer :~@> newFN, rg :~@+> ig, snd <$> back)
76−
77− -- Handle the output layer, bouncing the derivatives back down.
78− -- We may not have a target for each example, so when we don't use 0 gradient.
79− go !xs (OR layer) (ORS ())
80− = let tys = runForwards layer <$> xs
81− tapes = fst <$> tys
82− ys = snd <$> tys
83− -- recursively run the rest of the network, and get the gradients from above.
84− back = uncurry (runBackwards layer) <$> zip tapes (zipWith makeError ys targets)
85− -- the new trained layer.
86− newlayer = runUpdates rate layer (reverse $ fst <$> back)
87− in (OR newlayer, ORS (), reverse (snd <$> back))
88−
89− go _ _ _ =
90− error "Impossible for network and recurrent inputs to have different shapes"
91−
92−
93− makeError :: S (Last shapes) -> Maybe (S (Last shapes)) -> S (Last shapes)
94− makeError _ Nothing = 0
95− makeError y (Just t) = y - t
96−
97− updateRecInputs :: forall sublayers.
97+updateRecInputs :: Fractional (RecurrentInputs sublayers)
98− LearningParameters
98+ => LearningParameters
99− -> RecurrentInputs sublayers
99+ -> RecurrentInputs sublayers
100− -> RecurrentInputs sublayers
100+ -> RecurrentInputs sublayers
101− -> RecurrentInputs sublayers
101+ -> RecurrentInputs sublayers
102−
102+
103− updateRecInputs l@LearningParameters {..} (() :~~+> xs) (() :~~+> ys)
103+updateRecInputs lp (() :~~+> xs) (() :~~+> ys)
104− = () :~~+> updateRecInputs l xs ys
104+ = () :~~+> updateRecInputs lp xs ys
105−
105+
106− updateRecInputs l@LearningParameters {..} (x :~@+> xs) (y :~@+> ys)
106+updateRecInputs lp (x :~@+> xs) (y :~@+> ys)
107− = (realToFrac (learningRate * learningRegulariser) * x - realToFrac learningRate * y) :~@+> updateRecInputs l xs ys
107+ = (realToFrac (1 - learningRate lp * learningRegulariser lp) * x - realToFrac (learningRate lp) * y) :~@+> updateRecInputs lp xs ys
108−
108+
109− updateRecInputs _ (ORS ()) (ORS ())
109+updateRecInputs _ RINil RINil
110− = ORS ()
110+ = RINil
111− updateRecInputs _ _ _
112− = error "Impossible for updateRecInputs to have different shapes"
113−
114−scanlFrom :: forall x i o. RecurrentLayer x i o
115− => x -- ^ the layer
116− -> S (RecurrentShape x) -- ^ place to start
117− -> [S i] -- ^ list of inputs to scan through
118− -> [(S (RecurrentShape x), S o)] -- ^ list of scan inputs and outputs
119−scanlFrom !layer !recShape (x:xs) =
120− let (lerec, lepush) = runRecurrentForwards layer recShape x
121− in (recShape, lepush) : scanlFrom layer lerec xs
122−scanlFrom _ _ [] = []
123−
124−myscanbackward :: forall x i o. RecurrentLayer x i o
125− => x -- ^ the layer
126− -> [(S (RecurrentShape x), S i, S o)] -- ^ the list of inputs and output to scan over
127− -> (S (RecurrentShape x), [(Gradient x, S i)]) -- ^ list of gradients to fold and inputs to backprop
128−myscanbackward layer =
129− goX 0
130− where
131− goX :: S (RecurrentShape x) -> [(S (RecurrentShape x), S i, S o)] -> (S (RecurrentShape x), [(Gradient x, S i)])
132− goX !lastback ((recShape, lastin, backgrad):xs) =
133− let (layergrad, recgrad, ingrad) = runRecurrentBackwards layer recShape lastin lastback backgrad
134− (pushedback, ll) = goX recgrad xs
135− in (pushedback, (layergrad, ingrad) : ll)
136− goX !lastback [] = (lastback, [])
137−
138−-- | Just forwards propagation with no training.
139−runRecurrent :: RecurrentNetwork layers shapes
140− -> RecurrentInputs layers -> S (Head shapes)
141− -> (RecurrentInputs layers, S (Last shapes))
142−runRecurrent (layer :~~> n) (() :~~+> nr) !x
143− = let (_, ys) = runForwards layer x
144− (nr', o) = runRecurrent n nr ys
145− in (() :~~+> nr', o)
146−runRecurrent (layer :~@> n) (recin :~@+> nr) !x
147− = let (recin', y) = runRecurrentForwards layer recin x
148− (nr', o) = runRecurrent n nr y
149− in (recin' :~@+> nr', o)
150−runRecurrent (OR layer) (ORS ()) !x
151− = (ORS (), snd $ runForwards layer x)
152−
153−runRecurrent _ _ _
154− = error "Impossible for the gradients of a network to have a different length or shape to the network"
src/Grenade/Recurrent/Layers.hsnew file+9
@@ -0,0 +1,9 @@
1+module Grenade.Recurrent.Layers (
2+ module Grenade.Recurrent.Layers.BasicRecurrent
3+ , module Grenade.Recurrent.Layers.ConcatRecurrent
4+ , module Grenade.Recurrent.Layers.LSTM
5+ ) where
6+
7+import Grenade.Recurrent.Layers.BasicRecurrent
8+import Grenade.Recurrent.Layers.ConcatRecurrent
9+import Grenade.Recurrent.Layers.LSTM
src/Grenade/Recurrent/Layers/BasicRecurrent.hs+16−16
@@ -1,14 +1,12 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE DataKinds #-}
2{-# LANGUAGE GADTs #-}
3{-# LANGUAGE GADTs #-}
3−{-# LANGUAGE RecordWildCards #-}
4{-# LANGUAGE TypeOperators #-}
4{-# LANGUAGE TypeOperators #-}
5{-# LANGUAGE TypeFamilies #-}
5{-# LANGUAGE TypeFamilies #-}
6{-# LANGUAGE MultiParamTypeClasses #-}
6{-# LANGUAGE MultiParamTypeClasses #-}
7{-# LANGUAGE FlexibleContexts #-}
7{-# LANGUAGE FlexibleContexts #-}
8{-# LANGUAGE UndecidableInstances #-}
8{-# LANGUAGE UndecidableInstances #-}
9
9
10−-- GHC 7.10 doesn't see recurrent run functions as total.
11−{-# OPTIONS_GHC -fno-warn-incomplete-patterns #-}
12module Grenade.Recurrent.Layers.BasicRecurrent (
10module Grenade.Recurrent.Layers.BasicRecurrent (
13 BasicRecurrent (..)
11 BasicRecurrent (..)
14 , randomBasicRecurrent
12 , randomBasicRecurrent
@@ -18,19 +16,19 @@ module Grenade.Recurrent.Layers.BasicRecurrent (
18
16
19import Control.Monad.Random ( MonadRandom, getRandom )
17import Control.Monad.Random ( MonadRandom, getRandom )
20
18
19+
21−import Data.Singletons.TypeLits
20+import Data.Kind (Type)
22
21
23import Numeric.LinearAlgebra.Static
22import Numeric.LinearAlgebra.Static
24
23
25import GHC.TypeLits
24import GHC.TypeLits
26
25
27−import Grenade.Core.Network
26+import Grenade.Core
28−import Grenade.Core.Shape
29−import Grenade.Recurrent.Core.Network
27+import Grenade.Recurrent.Core
30
28
31data BasicRecurrent :: Nat -- Input layer size
29data BasicRecurrent :: Nat -- Input layer size
32 -> Nat -- Output layer size
30 -> Nat -- Output layer size
33− -> * where
31+ -> Type where
34 BasicRecurrent :: ( KnownNat input
32 BasicRecurrent :: ( KnownNat input
35 , KnownNat output
33 , KnownNat output
36 , KnownNat matrixCols
34 , KnownNat matrixCols
@@ -43,7 +41,7 @@ data BasicRecurrent :: Nat -- Input layer size
43
41
44data BasicRecurrent' :: Nat -- Input layer size
42data BasicRecurrent' :: Nat -- Input layer size
45 -> Nat -- Output layer size
43 -> Nat -- Output layer size
46− -> * where
44+ -> Type where
47 BasicRecurrent' :: ( KnownNat input
45 BasicRecurrent' :: ( KnownNat input
48 , KnownNat output
46 , KnownNat output
49 , KnownNat matrixCols
47 , KnownNat matrixCols
@@ -58,27 +56,29 @@ instance Show (BasicRecurrent i o) where
58instance (KnownNat i, KnownNat o, KnownNat (i + o)) => UpdateLayer (BasicRecurrent i o) where
56instance (KnownNat i, KnownNat o, KnownNat (i + o)) => UpdateLayer (BasicRecurrent i o) where
59 type Gradient (BasicRecurrent i o) = (BasicRecurrent' i o)
57 type Gradient (BasicRecurrent i o) = (BasicRecurrent' i o)
60
58
61− runUpdate LearningParameters {..} (BasicRecurrent oldBias oldBiasMomentum oldActivations oldMomentum) (BasicRecurrent' biasGradient activationGradient) =
59+ runUpdate lp (BasicRecurrent oldBias oldBiasMomentum oldActivations oldMomentum) (BasicRecurrent' biasGradient activationGradient) =
62− let newBiasMomentum = konst learningMomentum * oldBiasMomentum - konst learningRate * biasGradient
60+ let newBiasMomentum = konst (learningMomentum lp) * oldBiasMomentum - konst (learningRate lp) * biasGradient
63 newBias = oldBias + newBiasMomentum
61 newBias = oldBias + newBiasMomentum
64− newMomentum = konst learningMomentum * oldMomentum - konst learningRate * activationGradient
62+ newMomentum = konst (learningMomentum lp) * oldMomentum - konst (learningRate lp) * activationGradient
65− regulariser = konst (learningRegulariser * learningRate) * oldActivations
63+ regulariser = konst (learningRegulariser lp * learningRate lp) * oldActivations
66 newActivations = oldActivations + newMomentum - regulariser
64 newActivations = oldActivations + newMomentum - regulariser
67 in BasicRecurrent newBias newBiasMomentum newActivations newMomentum
65 in BasicRecurrent newBias newBiasMomentum newActivations newMomentum
68
66
69 createRandom = randomBasicRecurrent
67 createRandom = randomBasicRecurrent
70
68
71instance (KnownNat i, KnownNat o, KnownNat (i + o), i <= (i + o), o ~ ((i + o) - i)) => RecurrentUpdateLayer (BasicRecurrent i o) where
69instance (KnownNat i, KnownNat o, KnownNat (i + o), i <= (i + o), o ~ ((i + o) - i)) => RecurrentUpdateLayer (BasicRecurrent i o) where
72− type RecurrentShape (BasicRecurrent i o) = 'D1 o
70+ type RecurrentShape (BasicRecurrent i o) = S ('D1 o)
73
71
74instance (KnownNat i, KnownNat o, KnownNat (i + o), i <= (i + o), o ~ ((i + o) - i)) => RecurrentLayer (BasicRecurrent i o) ('D1 i) ('D1 o) where
72instance (KnownNat i, KnownNat o, KnownNat (i + o), i <= (i + o), o ~ ((i + o) - i)) => RecurrentLayer (BasicRecurrent i o) ('D1 i) ('D1 o) where
73+
74+ type RecTape (BasicRecurrent i o) ('D1 i) ('D1 o) = (S ('D1 o), S ('D1 i))
75 -- Do a matrix vector multiplication and return the result.
75 -- Do a matrix vector multiplication and return the result.
76 runRecurrentForwards (BasicRecurrent wB _ wN _) (S1D lastOutput) (S1D thisInput) =
76 runRecurrentForwards (BasicRecurrent wB _ wN _) (S1D lastOutput) (S1D thisInput) =
77 let thisOutput = S1D $ wB + wN #> (thisInput # lastOutput)
77 let thisOutput = S1D $ wB + wN #> (thisInput # lastOutput)
78− in (thisOutput, thisOutput)
78+ in ((S1D lastOutput, S1D thisInput), thisOutput, thisOutput)
79
79
80 -- Run a backpropogation step for a full connected layer.
80 -- Run a backpropogation step for a full connected layer.
81− runRecurrentBackwards (BasicRecurrent _ _ wN _) (S1D lastOutput) (S1D thisInput) (S1D dRec) (S1D dEdy) =
81+ runRecurrentBackwards (BasicRecurrent _ _ wN _) (S1D lastOutput, S1D thisInput) (S1D dRec) (S1D dEdy) =
82 let biasGradient = (dRec + dEdy)
82 let biasGradient = (dRec + dEdy)
83 layerGrad = (dRec + dEdy) `outer` (thisInput # lastOutput)
83 layerGrad = (dRec + dEdy) `outer` (thisInput # lastOutput)
84 -- calcluate derivatives for next step
84 -- calcluate derivatives for next step
src/Grenade/Recurrent/Layers/ConcatRecurrent.hsnew file+180
@@ -0,0 +1,180 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE DataKinds #-}
3+{-# LANGUAGE GADTs #-}
4+{-# LANGUAGE TypeOperators #-}
5+{-# LANGUAGE TypeFamilies #-}
6+{-# LANGUAGE MultiParamTypeClasses #-}
7+{-# LANGUAGE FlexibleContexts #-}
8+{-# LANGUAGE RankNTypes #-}
9+{-# LANGUAGE FlexibleInstances #-}
10+{-# LANGUAGE ScopedTypeVariables #-}
11+{-# LANGUAGE StandaloneDeriving #-}
12+{-# LANGUAGE UndecidableInstances #-}
13+{-|
14+Module : Grenade.Layers.Concat
15+Description : Concatenation layer
16+Copyright : (c) Huw Campbell, 2016-2017
17+License : BSD2
18+Stability : experimental
19+
20+This module provides the concatenation layer, which runs two chilld layers in parallel and combines their outputs.
21+-}
22+module Grenade.Recurrent.Layers.ConcatRecurrent (
23+ ConcatRecurrent (..)
24+ ) where
25+
26+import Data.Serialize
27+
28+import Data.Singletons
29+import GHC.TypeLits
30+
31+#if MIN_VERSION_base(4,9,0)
32+import Data.Kind (Type)
33+#endif
34+
35+import Grenade.Core
36+import Grenade.Recurrent.Core
37+
38+import Numeric.LinearAlgebra.Static ( (#), split, R )
39+
40+-- | A Concatentating Layer.
41+--
42+-- This layer shares it's input state between two sublayers, and concatenates their output.
43+--
44+-- With Networks able to be Layers, this allows for very expressive composition of complex Networks.
45+--
46+-- The Concat layer has a few instances, which allow one to flexibly "bash" together the outputs.
47+--
48+-- Two 1D vectors, can go to a 2D shape with 2 rows if their lengths are identical.
49+-- Any 2 1D vectors can also become a longer 1D Vector.
50+--
51+-- 3D images become 3D images with more channels. The sizes must be the same, one can use Pad
52+-- and Crop layers to ensure this is the case.
53+data ConcatRecurrent :: Shape -> Type -> Shape -> Type -> Type where
54+ ConcatRecLeft :: x -> y -> ConcatRecurrent m (Recurrent x) n (FeedForward y)
55+ ConcatRecRight :: x -> y -> ConcatRecurrent m (FeedForward x) n (Recurrent y)
56+ ConcatRecBoth :: x -> y -> ConcatRecurrent m (Recurrent x) n (Recurrent y)
57+
58+instance (Show x, Show y) => Show (ConcatRecurrent m (p x) n (q y)) where
59+ show (ConcatRecLeft x y) = "ConcatRecLeft\n" ++ show x ++ "\n" ++ show y
60+ show (ConcatRecRight x y) = "ConcatRecRight\n" ++ show x ++ "\n" ++ show y
61+ show (ConcatRecBoth x y) = "ConcatRecBoth\n" ++ show x ++ "\n" ++ show y
62+
63+instance (RecurrentUpdateLayer x, UpdateLayer y) => UpdateLayer (ConcatRecurrent m (Recurrent x) n (FeedForward y)) where
64+ type Gradient (ConcatRecurrent m (Recurrent x) n (FeedForward y)) = (Gradient x, Gradient y)
65+ runUpdate lr (ConcatRecLeft x y) (x', y')
66+ = ConcatRecLeft (runUpdate lr x x') (runUpdate lr y y')
67+ createRandom
68+ = ConcatRecLeft <$> createRandom <*> createRandom
69+
70+instance (UpdateLayer x, RecurrentUpdateLayer y) => UpdateLayer (ConcatRecurrent m (FeedForward x) n (Recurrent y)) where
71+ type Gradient (ConcatRecurrent m (FeedForward x) n (Recurrent y)) = (Gradient x, Gradient y)
72+ runUpdate lr (ConcatRecRight x y) (x', y')
73+ = ConcatRecRight (runUpdate lr x x') (runUpdate lr y y')
74+ createRandom
75+ = ConcatRecRight <$> createRandom <*> createRandom
76+
77+instance (RecurrentUpdateLayer x, RecurrentUpdateLayer y) => UpdateLayer (ConcatRecurrent m (Recurrent x) n (Recurrent y)) where
78+ type Gradient (ConcatRecurrent m (Recurrent x) n (Recurrent y)) = (Gradient x, Gradient y)
79+ runUpdate lr (ConcatRecBoth x y) (x', y')
80+ = ConcatRecBoth (runUpdate lr x x') (runUpdate lr y y')
81+ createRandom
82+ = ConcatRecBoth <$> createRandom <*> createRandom
83+
84+instance (RecurrentUpdateLayer x, UpdateLayer y) => RecurrentUpdateLayer (ConcatRecurrent m (Recurrent x) n (FeedForward y)) where
85+ type RecurrentShape (ConcatRecurrent m (Recurrent x) n (FeedForward y)) = RecurrentShape x
86+
87+instance (UpdateLayer x, RecurrentUpdateLayer y) => RecurrentUpdateLayer (ConcatRecurrent m (FeedForward x) n (Recurrent y)) where
88+ type RecurrentShape (ConcatRecurrent m (FeedForward x) n (Recurrent y)) = RecurrentShape y
89+
90+instance (RecurrentUpdateLayer x, RecurrentUpdateLayer y) => RecurrentUpdateLayer (ConcatRecurrent m (Recurrent x) n (Recurrent y)) where
91+ type RecurrentShape (ConcatRecurrent m (Recurrent x) n (Recurrent y)) = RecurrentInputs '[ Recurrent x, Recurrent y ]
92+
93+instance ( SingI i
94+ , Layer x i ('D1 m)
95+ , RecurrentLayer y i ('D1 n)
96+ , KnownNat o
97+ , KnownNat m
98+ , KnownNat n
99+ , o ~ (m + n)
100+ , n ~ (o - m)
101+ , (m <=? o) ~ 'True
102+ ) => RecurrentLayer (ConcatRecurrent ('D1 m) (FeedForward x) ('D1 n) (Recurrent y)) i ('D1 o) where
103+ type RecTape (ConcatRecurrent ('D1 m) (FeedForward x) ('D1 n) (Recurrent y)) i ('D1 o) = (Tape x i ('D1 m), RecTape y i ('D1 n))
104+
105+ runRecurrentForwards (ConcatRecRight x y) s input =
106+ let (xT, xOut :: S ('D1 m)) = runForwards x input
107+ (yT, side, yOut :: S ('D1 n)) = runRecurrentForwards y s input
108+ in case (xOut, yOut) of
109+ (S1D xOut', S1D yOut') ->
110+ ((xT, yT), side, S1D (xOut' # yOut'))
111+
112+ runRecurrentBackwards (ConcatRecRight x y) (xTape, yTape) s (S1D o) =
113+ let (ox :: R m , oy :: R n) = split o
114+ (x', xB :: S i) = runBackwards x xTape (S1D ox)
115+ (y', side, yB :: S i) = runRecurrentBackwards y yTape s (S1D oy)
116+ in ((x', y'), side, xB + yB)
117+
118+instance ( SingI i
119+ , RecurrentLayer x i ('D1 m)
120+ , Layer y i ('D1 n)
121+ , KnownNat o
122+ , KnownNat m
123+ , KnownNat n
124+ , o ~ (m + n)
125+ , n ~ (o - m)
126+ , (m <=? o) ~ 'True
127+ ) => RecurrentLayer (ConcatRecurrent ('D1 m) (Recurrent x) ('D1 n) (FeedForward y)) i ('D1 o) where
128+ type RecTape (ConcatRecurrent ('D1 m) (Recurrent x) ('D1 n) (FeedForward y)) i ('D1 o) = (RecTape x i ('D1 m), Tape y i ('D1 n))
129+
130+ runRecurrentForwards (ConcatRecLeft x y) s input =
131+ let (xT, side, xOut :: S ('D1 m)) = runRecurrentForwards x s input
132+ (yT, yOut :: S ('D1 n)) = runForwards y input
133+ in case (xOut, yOut) of
134+ (S1D xOut', S1D yOut') ->
135+ ((xT, yT), side, S1D (xOut' # yOut'))
136+
137+ runRecurrentBackwards (ConcatRecLeft x y) (xTape, yTape) s (S1D o) =
138+ let (ox :: R m , oy :: R n) = split o
139+ (x', side, xB :: S i) = runRecurrentBackwards x xTape s (S1D ox)
140+ (y', yB :: S i) = runBackwards y yTape (S1D oy)
141+ in ((x', y'), side, xB + yB)
142+
143+instance ( SingI i
144+ , RecurrentLayer x i ('D1 m)
145+ , RecurrentLayer y i ('D1 n)
146+ , Fractional (RecurrentShape x)
147+ , Fractional (RecurrentShape y)
148+ , KnownNat o
149+ , KnownNat m
150+ , KnownNat n
151+ , o ~ (m + n)
152+ , n ~ (o - m)
153+ , (m <=? o) ~ 'True
154+ ) => RecurrentLayer (ConcatRecurrent ('D1 m) (Recurrent x) ('D1 n) (Recurrent y)) i ('D1 o) where
155+ type RecTape (ConcatRecurrent ('D1 m) (Recurrent x) ('D1 n) (Recurrent y)) i ('D1 o) = (RecTape x i ('D1 m), RecTape y i ('D1 n))
156+
157+ runRecurrentForwards (ConcatRecBoth x y) (sx :~@+> (sy :~@+> RINil)) input =
158+ let (xT, s'x, xOut :: S ('D1 m)) = runRecurrentForwards x sx input
159+ (yT, s'y, yOut :: S ('D1 n)) = runRecurrentForwards y sy input
160+ in case (xOut, yOut) of
161+ (S1D xOut', S1D yOut') ->
162+ ((xT, yT), (s'x :~@+> (s'y :~@+> RINil)), S1D (xOut' # yOut'))
163+
164+ runRecurrentBackwards (ConcatRecBoth x y) (xTape, yTape) (sx :~@+> (sy :~@+> RINil)) (S1D o) =
165+ let (ox :: R m , oy :: R n) = split o
166+ (x', s'x, xB :: S i) = runRecurrentBackwards x xTape sx (S1D ox)
167+ (y', s'y, yB :: S i) = runRecurrentBackwards y yTape sy (S1D oy)
168+ in ((x', y'), (s'x :~@+> (s'y :~@+> RINil)), xB + yB)
169+
170+instance (Serialize a, Serialize b) => Serialize (ConcatRecurrent sa (FeedForward a) sb (Recurrent b)) where
171+ put (ConcatRecRight a b) = put a *> put b
172+ get = ConcatRecRight <$> get <*> get
173+
174+instance (Serialize a, Serialize b) => Serialize (ConcatRecurrent sa (Recurrent a) sb (Recurrent b)) where
175+ put (ConcatRecBoth a b) = put a *> put b
176+ get = ConcatRecBoth <$> get <*> get
177+
178+instance (Serialize a, Serialize b) => Serialize (ConcatRecurrent sa (Recurrent a) sb (FeedForward b)) where
179+ put (ConcatRecLeft a b) = put a *> put b
180+ get = ConcatRecLeft <$> get <*> get
src/Grenade/Recurrent/Layers/LSTM.hs+60−86
@@ -1,8 +1,8 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE BangPatterns #-}
2{-# LANGUAGE BangPatterns #-}
2{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE GADTs #-}
4{-# LANGUAGE GADTs #-}
4{-# LANGUAGE RankNTypes #-}
5{-# LANGUAGE RankNTypes #-}
5−{-# LANGUAGE RecordWildCards #-}
6{-# LANGUAGE TypeOperators #-}
6{-# LANGUAGE TypeOperators #-}
7{-# LANGUAGE TypeFamilies #-}
7{-# LANGUAGE TypeFamilies #-}
8{-# LANGUAGE MultiParamTypeClasses #-}
8{-# LANGUAGE MultiParamTypeClasses #-}
@@ -10,8 +10,6 @@
10{-# LANGUAGE ViewPatterns #-}
10{-# LANGUAGE ViewPatterns #-}
11{-# LANGUAGE ScopedTypeVariables #-}
11{-# LANGUAGE ScopedTypeVariables #-}
12
12
13−-- GHC 7.10 doesn't see recurrent run functions as total.
14−{-# OPTIONS_GHC -fno-warn-incomplete-patterns #-}
15module Grenade.Recurrent.Layers.LSTM (
13module Grenade.Recurrent.Layers.LSTM (
16 LSTM (..)
14 LSTM (..)
17 , LSTMWeights (..)
15 , LSTMWeights (..)
@@ -20,35 +18,33 @@ module Grenade.Recurrent.Layers.LSTM (
20
18
21import Control.Monad.Random ( MonadRandom, getRandom )
19import Control.Monad.Random ( MonadRandom, getRandom )
22
20
23−
24-- import Data.List ( foldl1' )
21-- import Data.List ( foldl1' )
25import Data.Proxy
22import Data.Proxy
26import Data.Serialize
23import Data.Serialize
27−import Data.Singletons.TypeLits
24+import Data.Kind (Type)
28
25
29import qualified Numeric.LinearAlgebra as LA
26import qualified Numeric.LinearAlgebra as LA
30import Numeric.LinearAlgebra.Static
27import Numeric.LinearAlgebra.Static
31
28
32−import Grenade.Core.Network
29+import Grenade.Core
33−import Grenade.Core.Shape
30+import Grenade.Recurrent.Core
34−
35import Grenade.Layers.Internal.Update
31import Grenade.Layers.Internal.Update
32+import GHC.TypeLits
36
33
37−import Grenade.Recurrent.Core.Network
38
34
39-- | Long Short Term Memory Recurrent unit
35-- | Long Short Term Memory Recurrent unit
40--
36--
41-- This is a Peephole formulation, so the recurrent shape is
37-- This is a Peephole formulation, so the recurrent shape is
42-- just the cell state, the previous output is not held or used
38-- just the cell state, the previous output is not held or used
43-- at all.
39-- at all.
44−data LSTM :: Nat -> Nat -> * where
40+data LSTM :: Nat -> Nat -> Type where
45 LSTM :: ( KnownNat input
41 LSTM :: ( KnownNat input
46 , KnownNat output
42 , KnownNat output
47 ) => !(LSTMWeights input output) -- Weights
43 ) => !(LSTMWeights input output) -- Weights
48 -> !(LSTMWeights input output) -- Momentums
44 -> !(LSTMWeights input output) -- Momentums
49 -> LSTM input output
45 -> LSTM input output
50
46
51−data LSTMWeights :: Nat -> Nat -> * where
47+data LSTMWeights :: Nat -> Nat -> Type where
52 LSTMWeights :: ( KnownNat input
48 LSTMWeights :: ( KnownNat input
53 , KnownNat output
49 , KnownNat output
54 ) => {
50 ) => {
@@ -75,7 +71,7 @@ instance (KnownNat i, KnownNat o) => UpdateLayer (LSTM i o) where
75
71
76 -- Run the update function for each group matrix/vector of weights, momentums and gradients.
72 -- Run the update function for each group matrix/vector of weights, momentums and gradients.
77 -- Hmm, maybe the function should be used instead of passing in the learning parameters.
73 -- Hmm, maybe the function should be used instead of passing in the learning parameters.
78− runUpdate LearningParameters {..} (LSTM w m) g =
74+ runUpdate lp (LSTM w m) g =
79 let (wf, wf') = u lstmWf w m g
75 let (wf, wf') = u lstmWf w m g
80 (uf, uf') = u lstmUf w m g
76 (uf, uf') = u lstmUf w m g
81 (bf, bf') = v lstmBf w m g
77 (bf, bf') = v lstmBf w m g
@@ -92,11 +88,11 @@ instance (KnownNat i, KnownNat o) => UpdateLayer (LSTM i o) where
92 -- Utility function for updating with the momentum, gradients, and weights.
88 -- Utility function for updating with the momentum, gradients, and weights.
93 u :: forall x ix out. (KnownNat ix, KnownNat out) => (x -> (L out ix)) -> x -> x -> x -> ((L out ix), (L out ix))
89 u :: forall x ix out. (KnownNat ix, KnownNat out) => (x -> (L out ix)) -> x -> x -> x -> ((L out ix), (L out ix))
94 u e (e -> weights) (e -> momentum) (e -> gradient) =
90 u e (e -> weights) (e -> momentum) (e -> gradient) =
95− decendMatrix learningRate learningMomentum learningRegulariser weights gradient momentum
91+ descendMatrix (learningRate lp) (learningMomentum lp) (learningRegulariser lp) weights gradient momentum
96
92
97 v :: forall x ix. (KnownNat ix) => (x -> (R ix)) -> x -> x -> x -> ((R ix), (R ix))
93 v :: forall x ix. (KnownNat ix) => (x -> (R ix)) -> x -> x -> x -> ((R ix), (R ix))
98 v e (e -> weights) (e -> momentum) (e -> gradient) =
94 v e (e -> weights) (e -> momentum) (e -> gradient) =
99− decendVector learningRate learningMomentum learningRegulariser weights gradient momentum
95+ descendVector (learningRate lp) (learningMomentum lp) (learningRegulariser lp) weights gradient momentum
100
96
101 -- There's a lot of updates here, so to try and minimise the number of data copies
97 -- There's a lot of updates here, so to try and minimise the number of data copies
102 -- we'll create a mutable bucket for each.
98 -- we'll create a mutable bucket for each.
@@ -128,25 +124,33 @@ instance (KnownNat i, KnownNat o) => UpdateLayer (LSTM i o) where
128instance (KnownNat i, KnownNat o) => RecurrentUpdateLayer (LSTM i o) where
124instance (KnownNat i, KnownNat o) => RecurrentUpdateLayer (LSTM i o) where
129 -- The recurrent shape is the same size as the output.
125 -- The recurrent shape is the same size as the output.
130 -- It's actually the cell state however, as this is a peephole variety LSTM.
126 -- It's actually the cell state however, as this is a peephole variety LSTM.
131− type RecurrentShape (LSTM i o) = 'D1 o
127+ type RecurrentShape (LSTM i o) = S ('D1 o)
132
128
133instance (KnownNat i, KnownNat o) => RecurrentLayer (LSTM i o) ('D1 i) ('D1 o) where
129instance (KnownNat i, KnownNat o) => RecurrentLayer (LSTM i o) ('D1 i) ('D1 o) where
130+
131+ -- The tape stores essentially every variable we calculate,
132+ -- so we don't have to run any forwards component again.
133+ type RecTape (LSTM i o) ('D1 i) ('D1 o) = (R o, R i, R o, R o, R o, R o, R o, R o, R o, R o, R o)
134 -- Forward propagation for the LSTM layer.
134 -- Forward propagation for the LSTM layer.
135 -- The size of the cell state is also the size of the output.
135 -- The size of the cell state is also the size of the output.
136− runRecurrentForwards (LSTM (LSTMWeights {..}) _) (S1D cell) (S1D input) =
136+ runRecurrentForwards (LSTM lw _) (S1D cell) (S1D input) =
137 let -- Forget state vector
137 let -- Forget state vector
138− f_t = sigmoid $ lstmBf + lstmWf #> input + lstmUf #> cell
138+ f_s = lstmBf lw + lstmWf lw #> input + lstmUf lw #> cell
139+ f_t = sigmoid f_s
139 -- Input state vector
140 -- Input state vector
140− i_t = sigmoid $ lstmBi + lstmWi #> input + lstmUi #> cell
141+ i_s = lstmBi lw + lstmWi lw #> input + lstmUi lw #> cell
142+ i_t = sigmoid i_s
141 -- Output state vector
143 -- Output state vector
142− o_t = sigmoid $ lstmBo + lstmWo #> input + lstmUo #> cell
144+ o_s = lstmBo lw + lstmWo lw #> input + lstmUo lw #> cell
145+ o_t = sigmoid o_s
143 -- Cell input state vector
146 -- Cell input state vector
144− c_x = tanh $ lstmBc + lstmWc #> input
147+ c_s = lstmBc lw + lstmWc lw #> input
148+ c_x = tanh c_s
145 -- Cell state
149 -- Cell state
146 c_t = f_t * cell + i_t * c_x
150 c_t = f_t * cell + i_t * c_x
147 -- Output (it's sometimes recommended to use tanh c_t)
151 -- Output (it's sometimes recommended to use tanh c_t)
148 h_t = o_t * c_t
152 h_t = o_t * c_t
149− in (S1D c_t, S1D h_t)
153+ in ((cell, input, f_s, f_t, i_s, i_t, o_s, o_t, c_s, c_x, c_t), S1D c_t, S1D h_t)
150
154
151 -- Run a backpropogation step for an LSTM layer.
155 -- Run a backpropogation step for an LSTM layer.
152 -- We're doing all the derivatives by hand here, so one should
156 -- We're doing all the derivatives by hand here, so one should
@@ -154,29 +158,8 @@ instance (KnownNat i, KnownNat o) => RecurrentLayer (LSTM i o) ('D1 i) ('D1 o) w
154 --
158 --
155 -- There's a test version using the AD library without hmatrix in the test
159 -- There's a test version using the AD library without hmatrix in the test
156 -- suite. These should match always.
160 -- suite. These should match always.
157− runRecurrentBackwards (LSTM (LSTMWeights {..}) _) (S1D cell) (S1D input) (S1D cellGrad) (S1D h_t') =
161+ runRecurrentBackwards (LSTM lw _) (cell, input, f_s, f_t, i_s, i_t, o_s, o_t, c_s, c_x, c_t) (S1D cellGrad) (S1D h_t') =
158− -- We're not keeping the Wengert tape during the forward pass,
159− -- so we're duplicating some work here.
160− --
161− -- If I was being generous, I'd call it checkpointing.
162− --
163− -- Maybe think about better ways to store some intermediate states.
164− let -- Forget state vector
165− f_s = lstmBf + lstmWf #> input + lstmUf #> cell
166− f_t = sigmoid f_s
167− -- Input state vector
168− i_s = lstmBi + lstmWi #> input + lstmUi #> cell
169− i_t = sigmoid i_s
170− -- Output state vector
171− o_s = lstmBo + lstmWo #> input + lstmUo #> cell
172− o_t = sigmoid o_s
173− -- Cell input state vector
174− c_s = lstmBc + lstmWc #> input
175− c_x = tanh c_s
176− -- Cell state
177− c_t = f_t * cell + i_t * c_x
178−
179− -- Reverse Mode AD Derivitives
162+ let -- Reverse Mode AD Derivitives
180 c_t' = h_t' * o_t + cellGrad
163 c_t' = h_t' * o_t + cellGrad
181
164
182 f_t' = c_t' * cell
165 f_t' = c_t' * cell
@@ -192,8 +175,8 @@ instance (KnownNat i, KnownNat o) => RecurrentLayer (LSTM i o) ('D1 i) ('D1 o) w
192 c_s' = tanh' c_s * c_x'
175 c_s' = tanh' c_s * c_x'
193
176
194 -- The derivatives to pass sideways (recurrent) and downwards
177 -- The derivatives to pass sideways (recurrent) and downwards
195− cell' = tr lstmUf #> f_s' + tr lstmUo #> o_s' + tr lstmUi #> i_s' + c_t' * f_t
178+ cell' = tr (lstmUf lw) #> f_s' + tr (lstmUo lw) #> o_s' + tr (lstmUi lw) #> i_s' + c_t' * f_t
196− input' = tr lstmWf #> f_s' + tr lstmWo #> o_s' + tr lstmWi #> i_s' + tr lstmWc #> c_s'
179+ input' = tr (lstmWf lw) #> f_s' + tr (lstmWo lw) #> o_s' + tr (lstmWi lw) #> i_s' + tr (lstmWc lw) #> c_s'
197
180
198 -- Calculate the gradient Matricies for the input
181 -- Calculate the gradient Matricies for the input
199 lstmWf' = f_s' `outer` input
182 lstmWf' = f_s' `outer` input
@@ -238,7 +221,8 @@ randomLSTM = do
238
221
239-- | Maths
222-- | Maths
240--
223--
241−-- TODO: move to not here
224+-- TODO: Move to not here
225+-- Optimise backwards derivative
242sigmoid :: Floating a => a -> a
226sigmoid :: Floating a => a -> a
243sigmoid x = 1 / (1 + exp (-x))
227sigmoid x = 1 / (1 + exp (-x))
244
228
@@ -251,44 +235,34 @@ tanh' :: (Floating a) => a -> a
251tanh' t = 1 - s ^ (2 :: Int) where s = tanh t
235tanh' t = 1 - s ^ (2 :: Int) where s = tanh t
252
236
253instance (KnownNat i, KnownNat o) => Serialize (LSTM i o) where
237instance (KnownNat i, KnownNat o) => Serialize (LSTM i o) where
254− put (LSTM LSTMWeights {..} _) = do
238+ put (LSTM lw _) = do
255− u lstmWf
239+ u (lstmWf lw)
256− u lstmUf
240+ u (lstmUf lw)
257− v lstmBf
241+ v (lstmBf lw)
258− u lstmWi
242+ u (lstmWi lw)
259− u lstmUi
243+ u (lstmUi lw)
260− v lstmBi
244+ v (lstmBi lw)
261− u lstmWo
245+ u (lstmWo lw)
262− u lstmUo
246+ u (lstmUo lw)
263− v lstmBo
247+ v (lstmBo lw)
264− u lstmWc
248+ u (lstmWc lw)
265− v lstmBc
249+ v (lstmBc lw)
266− where
250+ where
267− u :: forall a b. (KnownNat a, KnownNat b) => Putter (L b a)
251+ u :: forall a b. (KnownNat a, KnownNat b) => Putter (L b a)
268− u = putListOf put . LA.toList . LA.flatten . extract
252+ u = putListOf put . LA.toList . LA.flatten . extract
269− v :: forall a. (KnownNat a) => Putter (R a)
253+ v :: forall a. (KnownNat a) => Putter (R a)
270− v = putListOf put . LA.toList . extract
254+ v = putListOf put . LA.toList . extract
271
255
272 get = do
256 get = do
273− lstmWf <- u
257+ w <- LSTMWeights <$> u <*> u <*> v <*> u <*> u <*> v <*> u <*> u <*> v <*> u <*> v
274− lstmUf <- u
275− lstmBf <- v
276− lstmWi <- u
277− lstmUi <- u
278− lstmBi <- v
279− lstmWo <- u
280− lstmUo <- u
281− lstmBo <- v
282− lstmWc <- u
283− lstmBc <- v
284− return $ LSTM (LSTMWeights {..}) (LSTMWeights w0 u0 v0 w0 u0 v0 w0 u0 v0 w0 v0)
258+ return $ LSTM w (LSTMWeights w0 u0 v0 w0 u0 v0 w0 u0 v0 w0 v0)
285− where
259+ where
286− u :: forall a b. (KnownNat a, KnownNat b) => Get (L b a)
260+ u :: forall a b. (KnownNat a, KnownNat b) => Get (L b a)
287− u = let f = fromIntegral $ natVal (Proxy :: Proxy a)
261+ u = let f = fromIntegral $ natVal (Proxy :: Proxy a)
288− in maybe (fail "Vector of incorrect size") return . create . LA.reshape f . LA.fromList =<< getListOf get
262+ in maybe (fail "Vector of incorrect size") return . create . LA.reshape f . LA.fromList =<< getListOf get
289− v :: forall a. (KnownNat a) => Get (R a)
263+ v :: forall a. (KnownNat a) => Get (R a)
290− v = maybe (fail "Vector of incorrect size") return . create . LA.fromList =<< getListOf get
264+ v = maybe (fail "Vector of incorrect size") return . create . LA.fromList =<< getListOf get
291−
265+
292− w0 = konst 0
266+ w0 = konst 0
293− u0 = konst 0
267+ u0 = konst 0
294− v0 = konst 0
268+ v0 = konst 0
src/Grenade/Utils/OneHot.hs+6−5
@@ -22,7 +22,6 @@ import Data.Map ( Map )
22import qualified Data.Map as M
22import qualified Data.Map as M
23
23
24import Data.Proxy
24import Data.Proxy
25−import Data.Singletons.TypeLits
26
25
27import Data.Vector ( Vector )
26import Data.Vector ( Vector )
28import qualified Data.Vector as V
27import qualified Data.Vector as V
@@ -31,6 +30,7 @@ import qualified Data.Vector.Storable as VS
31import Numeric.LinearAlgebra ( maxIndex )
30import Numeric.LinearAlgebra ( maxIndex )
32import Numeric.LinearAlgebra.Devel
31import Numeric.LinearAlgebra.Devel
33import Numeric.LinearAlgebra.Static
32import Numeric.LinearAlgebra.Static
33+import GHC.TypeLits
34
34
35import Grenade.Core.Shape
35import Grenade.Core.Shape
36
36
@@ -52,15 +52,16 @@ oneHot hot =
52
52
53-- | Create a one hot map from any enumerable.
53-- | Create a one hot map from any enumerable.
54-- Returns a map, and the ordered list for the reverse transformation
54-- Returns a map, and the ordered list for the reverse transformation
55−hotMap :: (Ord a, KnownNat n) => Proxy n -> [a] -> Maybe (Map a Int, Vector a)
55+hotMap :: (Ord a, KnownNat n) => Proxy n -> [a] -> Either String (Map a Int, Vector a)
56hotMap n as =
56hotMap n as =
57 let len = fromIntegral $ natVal n
57 let len = fromIntegral $ natVal n
58 uniq = [ c | (c:_) <- group $ sort as]
58 uniq = [ c | (c:_) <- group $ sort as]
59 hotl = length uniq
59 hotl = length uniq
60− in if hotl <= len
60+ in if hotl == len
61 then
61 then
62− Just (M.fromList $ zip uniq [0..], V.fromList uniq)
62+ Right (M.fromList $ zip uniq [0..], V.fromList uniq)
63− else Nothing
63+ else
64+ Left ("Couldn't create hotMap of size " ++ show len ++ " from vector with " ++ show hotl ++ " unique characters")
64
65
65-- | From a map and value, create a 1D Shape
66-- | From a map and value, create a 1D Shape
66-- with one index hot (1) with the rest 0.
67-- with one index hot (1) with the rest 0.
stack.yamlnew file+78
@@ -0,0 +1,78 @@
1+# This file was automatically generated by 'stack init'
2+#
3+# Some commonly used options have been documented as comments in this file.
4+# For advanced use and comprehensive documentation of the format, please see:
5+# https://docs.haskellstack.org/en/stable/yaml_configuration/
6+
7+# A warning or info to be displayed to the user on config load.
8+# user-message: |
9+
10+# Resolver to choose a 'specific' stackage snapshot or a compiler version.
11+# A snapshot resolver dictates the compiler version and the set of packages
12+# to be used for project dependencies. For example:
13+#
14+# resolver: lts-3.5
15+# resolver: nightly-2015-09-21
16+# resolver: ghc-7.10.2
17+#
18+# The location of a snapshot can be provided as a file or url. Stack assumes
19+# a snapshot provided as a file might change, whereas a url resource does not.
20+#
21+# resolver: ./custom-snapshot.yaml
22+# resolver: https://example.com/snapshots/2018-01-01.yaml
23+resolver: lts-20.18
24+
25+# User packages to be built.
26+# Various formats can be used as shown in the example below.
27+#
28+# packages:
29+# - some-directory
30+# - https://example.com/foo/bar/baz-0.0.2.tar.gz
31+# subdirs:
32+# - auto-update
33+# - wai
34+packages:
35+ - .
36+ - examples
37+
38+# Dependency packages to be pulled from upstream that are not in the resolver.
39+# These entries can reference officially published versions as well as
40+# forks / in-progress versions pinned to a git hash. For example:
41+#
42+# extra-deps:
43+# - acme-missiles-0.3
44+# - git: https://github.com/commercialhaskell/stack.git
45+# commit: e7b331f14bcffb8367cd58fbfc8b40ec7642100a
46+#
47+extra-deps:
48+ - typelits-witnesses-0.3.0.3
49+
50+# Override default flag values for local packages and extra-deps
51+# flags: {}
52+
53+# Extra package databases containing global packages
54+# extra-package-dbs: []
55+
56+# Control whether we use the GHC we find on the path
57+# system-ghc: true
58+#
59+# Require a specific version of stack, using version ranges
60+# require-stack-version: -any # Default
61+# require-stack-version: ">=2.1"
62+#
63+# Override the architecture used by stack, especially useful on Windows
64+# arch: i386
65+# arch: x86_64
66+#
67+# Extra directories used by stack for building
68+# extra-include-dirs: [/path/to/dir]
69+# extra-lib-dirs: [/path/to/dir]
70+#
71+# Allow a newer minor version of GHC than the snapshot specifies
72+# compiler-check: newer-minor
73+
74+nix:
75+ enable: true
76+ packages:
77+ - blas
78+ - lapack
stack.yaml.locknew file+19
@@ -0,0 +1,19 @@
1+# This file was autogenerated by Stack.
2+# You should not edit this file by hand.
3+# For more information, please see the documentation at:
4+# https://docs.haskellstack.org/en/stable/lock_files
5+
6+packages:
7+- completed:
8+ hackage: typelits-witnesses-0.3.0.3@sha256:2d9df4ac6ff3077bfd2bf659e4b495e157723ac5b45c519762853f55df5c16db,2738
9+ pantry-tree:
10+ sha256: 6a42a462f98e94933b6e9721acd912c6c6b6a4743635efd15cc5871908c816a0
11+ size: 469
12+ original:
13+ hackage: typelits-witnesses-0.3.0.3
14+snapshots:
15+- completed:
16+ sha256: 9fa4bece7acfac1fc7930c5d6e24606004b09e80aa0e52e9f68b148201008db9
17+ size: 649606
18+ url: https://raw.githubusercontent.com/commercialhaskell/stackage-snapshots/master/lts/20/18.yaml
19+ original: lts-20.18
test/Test/Grenade/Layers/Convolution.hs+32−28
@@ -1,3 +1,4 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE TemplateHaskell #-}
2{-# LANGUAGE TemplateHaskell #-}
2{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE GADTs #-}
4{-# LANGUAGE GADTs #-}
@@ -15,16 +16,21 @@ import Data.Singletons ()
15import GHC.TypeLits
16import GHC.TypeLits
16import GHC.TypeLits.Witnesses
17import GHC.TypeLits.Witnesses
17
18
18−import Grenade.Core.Shape
19+#if MIN_VERSION_base(4,9,0)
20+import Data.Kind (Type)
21+#endif
22+
19−import Grenade.Core.Network
23+import Grenade.Core
20import Grenade.Layers.Convolution
24import Grenade.Layers.Convolution
21
25
22−import Disorder.Jack
26+import Hedgehog
27+import qualified Hedgehog.Gen as Gen
23
28
24−import Test.Jack.Hmatrix
29+import Test.Hedgehog.Hmatrix
25−import Test.Jack.TypeLits
30+import Test.Hedgehog.TypeLits
31+import Test.Hedgehog.Compat
26
32
27−data OpaqueConvolution :: * where
33+data OpaqueConvolution :: Type where
28 OpaqueConvolution :: Convolution channels filters kernelRows kernelColumns strideRows strideColumns -> OpaqueConvolution
34 OpaqueConvolution :: Convolution channels filters kernelRows kernelColumns strideRows strideColumns -> OpaqueConvolution
29
35
30instance Show OpaqueConvolution where
36instance Show OpaqueConvolution where
@@ -38,10 +44,10 @@ genConvolution :: ( KnownNat channels
38 , KnownNat strideColumns
44 , KnownNat strideColumns
39 , KnownNat kernelFlattened
45 , KnownNat kernelFlattened
40 , kernelFlattened ~ (kernelRows * kernelColumns * channels)
46 , kernelFlattened ~ (kernelRows * kernelColumns * channels)
41− ) => Jack (Convolution channels filters kernelRows kernelColumns strideRows strideColumns)
47+ ) => Gen (Convolution channels filters kernelRows kernelColumns strideRows strideColumns)
42genConvolution = Convolution <$> uniformSample <*> uniformSample
48genConvolution = Convolution <$> uniformSample <*> uniformSample
43
49
44−genOpaqueOpaqueConvolution :: Jack OpaqueConvolution
50+genOpaqueOpaqueConvolution :: Gen OpaqueConvolution
45genOpaqueOpaqueConvolution = do
51genOpaqueOpaqueConvolution = do
46 channels <- genNat
52 channels <- genNat
47 filters <- genNat
53 filters <- genNat
@@ -57,17 +63,17 @@ genOpaqueOpaqueConvolution = do
57 p2 = natDict pkc
63 p2 = natDict pkc
58 p3 = natDict pch
64 p3 = natDict pch
59 in case p1 %* p2 %* p3 of
65 in case p1 %* p2 %* p3 of
60− Dict -> OpaqueConvolution <$> (genConvolution :: Jack (Convolution ch fl kr kc sr sc))
66+ Dict -> OpaqueConvolution <$> (genConvolution :: Gen (Convolution ch fl kr kc sr sc))
67+
68+prop_conv_net_witness = property $
69+ blindForAll genOpaqueOpaqueConvolution >>= \onet ->
70+ case onet of
71+ (OpaqueConvolution ((Convolution _ _) :: Convolution channels filters kernelRows kernelCols strideRows strideCols)) -> success
61
72
62−prop_conv_net_witness =
63− gamble genOpaqueOpaqueConvolution $ \onet ->
64− (case onet of
65− (OpaqueConvolution ((Convolution _ _) :: Convolution channels filters kernelRows kernelCols strideRows strideCols)) -> True
66− ) :: Bool
67
73
68−prop_conv_net =
74+prop_conv_net = property $
69− gamble genOpaqueOpaqueConvolution $ \onet ->
75+ blindForAll genOpaqueOpaqueConvolution >>= \onet ->
70− (case onet of
76+ case onet of
71 (OpaqueConvolution (convLayer@(Convolution _ _) :: Convolution channels filters kernelRows kernelCols strideRows strideCols)) ->
77 (OpaqueConvolution (convLayer@(Convolution _ _) :: Convolution channels filters kernelRows kernelCols strideRows strideCols)) ->
72 let ok stride kernel = [extent | extent <- [(kernel + 1) .. 30 ], (extent - kernel) `mod` stride == 0]
78 let ok stride kernel = [extent | extent <- [(kernel + 1) .. 30 ], (extent - kernel) `mod` stride == 0]
73 kr = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
79 kr = fromIntegral $ natVal (Proxy :: Proxy kernelRows)
@@ -75,15 +81,15 @@ prop_conv_net =
75 sr = fromIntegral $ natVal (Proxy :: Proxy strideRows)
81 sr = fromIntegral $ natVal (Proxy :: Proxy strideRows)
76 sc = fromIntegral $ natVal (Proxy :: Proxy strideCols)
82 sc = fromIntegral $ natVal (Proxy :: Proxy strideCols)
77
83
78− in gamble (elements (ok sr kr)) $ \er ->
84+ in forAll (Gen.element (ok sr kr)) >>= \er ->
79− gamble (elements (ok sc kc)) $ \ec ->
85+ forAll (Gen.element (ok sc kc)) >>= \ec ->
80 let rr = ((er - kr) `div` sr) + 1
86 let rr = ((er - kr) `div` sr) + 1
81 rc = ((ec - kc) `div` sc) + 1
87 rc = ((ec - kc) `div` sc) + 1
82 Just er' = someNatVal er
88 Just er' = someNatVal er
83 Just ec' = someNatVal ec
89 Just ec' = someNatVal ec
84 Just rr' = someNatVal rr
90 Just rr' = someNatVal rr
85 Just rc' = someNatVal rc
91 Just rc' = someNatVal rc
86− in (case (er', ec', rr', rc') of
92+ in case (er', ec', rr', rc') of
87 ( SomeNat (pinr :: Proxy inRows), SomeNat (_ :: Proxy inCols), SomeNat (pour :: Proxy outRows), SomeNat (_ :: Proxy outCols)) ->
93 ( SomeNat (pinr :: Proxy inRows), SomeNat (_ :: Proxy inCols), SomeNat (pour :: Proxy outRows), SomeNat (_ :: Proxy outCols)) ->
88 case ( natDict pinr %* natDict (Proxy :: Proxy channels)
94 case ( natDict pinr %* natDict (Proxy :: Proxy channels)
89 , natDict pour %* natDict (Proxy :: Proxy filters)
95 , natDict pour %* natDict (Proxy :: Proxy filters)
@@ -91,14 +97,12 @@ prop_conv_net =
91 , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outRows - 1) * strideRows) ~ (inRows - kernelRows)))
97 , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outRows - 1) * strideRows) ~ (inRows - kernelRows)))
92 , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outCols - 1) * strideCols) ~ (inCols - kernelCols)))) of
98 , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outCols - 1) * strideCols) ~ (inCols - kernelCols)))) of
93 (Dict, Dict, Dict, Dict) ->
99 (Dict, Dict, Dict, Dict) ->
94− gamble (S3D <$> uniformSample) $ \(input :: S ('D3 inRows inCols channels)) ->
100+ blindForAll (S3D <$> uniformSample) >>= \(input :: S ('D3 inRows inCols channels)) ->
95− let output :: S ('D3 outRows outCols filters) = runForwards convLayer input
101+ let (tape, output :: S ('D3 outRows outCols filters)) = runForwards convLayer input
96 backed :: (Gradient (Convolution channels filters kernelRows kernelCols strideRows strideCols), S ('D3 inRows inCols channels))
102 backed :: (Gradient (Convolution channels filters kernelRows kernelCols strideRows strideCols), S ('D3 inRows inCols channels))
97− = runBackwards convLayer input output
103+ = runBackwards convLayer tape output
98− in backed `seq` True
104+ in backed `seq` success
99− ) :: Property
105+
100− ) :: Property
101
106
102−return []
103tests :: IO Bool
107tests :: IO Bool
104−tests = $quickCheckAll
108+tests = checkParallel $$(discover)
test/Test/Grenade/Layers/FullyConnected.hs+23−20
@@ -1,3 +1,4 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE BangPatterns #-}
2{-# LANGUAGE BangPatterns #-}
2{-# LANGUAGE TemplateHaskell #-}
3{-# LANGUAGE TemplateHaskell #-}
3{-# LANGUAGE DataKinds #-}
4{-# LANGUAGE DataKinds #-}
@@ -12,25 +13,28 @@ import Data.Singletons ()
12
13
13import GHC.TypeLits
14import GHC.TypeLits
14
15
15−import Grenade.Core.Shape
16+#if MIN_VERSION_base(4,9,0)
16−import Grenade.Core.Network
17+import Data.Kind (Type)
17−import Grenade.Layers.FullyConnected
18+#endif
18
19
19−import Disorder.Jack
20+import Grenade.Core
21+import Grenade.Layers.FullyConnected
20
22
21−import Test.Jack.Hmatrix
23+import Hedgehog
22
24
25+import Test.Hedgehog.Compat
26+import Test.Hedgehog.Hmatrix
23
27
24−data OpaqueFullyConnected :: * where
28+data OpaqueFullyConnected :: Type where
25 OpaqueFullyConnected :: (KnownNat i, KnownNat o) => FullyConnected i o -> OpaqueFullyConnected
29 OpaqueFullyConnected :: (KnownNat i, KnownNat o) => FullyConnected i o -> OpaqueFullyConnected
26
30
27instance Show OpaqueFullyConnected where
31instance Show OpaqueFullyConnected where
28 show (OpaqueFullyConnected n) = show n
32 show (OpaqueFullyConnected n) = show n
29
33
30−genOpaqueFullyConnected :: Jack OpaqueFullyConnected
34+genOpaqueFullyConnected :: Gen OpaqueFullyConnected
31genOpaqueFullyConnected = do
35genOpaqueFullyConnected = do
32− input :: Integer <- choose (2, 100)
36+ input :: Integer <- choose 2 100
33− output :: Integer <- choose (1, 100)
37+ output :: Integer <- choose 1 100
34 let Just input' = someNatVal input
38 let Just input' = someNatVal input
35 let Just output' = someNatVal output
39 let Just output' = someNatVal output
36 case (input', output') of
40 case (input', output') of
@@ -39,17 +43,16 @@ genOpaqueFullyConnected = do
39 bM <- randomVector
43 bM <- randomVector
40 wN <- uniformSample
44 wN <- uniformSample
41 kM <- uniformSample
45 kM <- uniformSample
42− return . OpaqueFullyConnected $ (FullyConnected wB bM wN kM :: FullyConnected i' o')
46+ return . OpaqueFullyConnected $ (FullyConnected (FullyConnected' wB wN) (FullyConnected' bM kM) :: FullyConnected i' o')
43
47
44prop_fully_connected_forwards :: Property
48prop_fully_connected_forwards :: Property
45−prop_fully_connected_forwards =
49+prop_fully_connected_forwards = property $ do
46− gamble genOpaqueFullyConnected $ \(OpaqueFullyConnected (fclayer :: FullyConnected i o)) ->
50+ OpaqueFullyConnected (fclayer :: FullyConnected i o) <- blindForAll genOpaqueFullyConnected
47− gamble (S1D <$> randomVector) $ \(input :: S ('D1 i)) ->
51+ input :: S ('D1 i) <- blindForAll (S1D <$> randomVector)
48− let output :: S ('D1 o) = runForwards fclayer input
52+ let (tape, output :: S ('D1 o)) = runForwards fclayer input
49− backed :: (Gradient (FullyConnected i o), S ('D1 i))
53+ backed :: (Gradient (FullyConnected i o), S ('D1 i))
50− = runBackwards fclayer input output
54+ = runBackwards fclayer tape output
51− in backed `seq` True
55+ backed `seq` success
52−
56+
53−return []
54tests :: IO Bool
57tests :: IO Bool
55−tests = $quickCheckAll
58+tests = checkParallel $$(discover)
test/Test/Grenade/Layers/Internal/Convolution.hs+34−29
@@ -12,41 +12,46 @@ import Grenade.Layers.Internal.Convolution
12
12
13import Numeric.LinearAlgebra hiding (uniformSample, konst, (===))
13import Numeric.LinearAlgebra hiding (uniformSample, konst, (===))
14
14
15−import Disorder.Jack
15+import Hedgehog
16+import qualified Hedgehog.Gen as Gen
17+import qualified Hedgehog.Range as Range
16
18
17import qualified Test.Grenade.Layers.Internal.Reference as Reference
19import qualified Test.Grenade.Layers.Internal.Reference as Reference
20+import Test.Hedgehog.Compat
18
21
19prop_im2col_col2im_symmetrical_with_kernel_stride =
22prop_im2col_col2im_symmetrical_with_kernel_stride =
20 let factors n = [x | x <- [1..n], n `mod` x == 0]
23 let factors n = [x | x <- [1..n], n `mod` x == 0]
21− in gamble (choose (2, 100)) $ \height ->
24+ in property $ do
22− gamble (choose (2, 100)) $ \width ->
25+ height <- forAll $ choose 2 100
23− gamble ((height `div`) <$> elements (factors height)) $ \kernel_h ->
26+ width <- forAll $ choose 2 100
24− gamble ((width `div`) <$> elements (factors width)) $ \kernel_w ->
27+ kernel_h <- forAll $ (height `div`) <$> Gen.element (factors height)
25− gamble (listOfN (height * width) (height * width) sizedRealFrac) $ \input ->
28+ kernel_w <- forAll $ (width `div`) <$> Gen.element (factors width)
26− let input' = (height >< width) input
29+ input <- forAll $ (height >< width) <$> Gen.list (Range.singleton $ height * width) (Gen.realFloat $ Range.linearFracFrom 0 (-100) 100)
30+
27− stride_h = kernel_h
31+ let stride_h = kernel_h
28− stride_w = kernel_w
32+ let stride_w = kernel_w
29− out = col2im kernel_h kernel_w stride_h stride_w height width . im2col kernel_h kernel_w stride_h stride_w $ input'
33+ let out = col2im kernel_h kernel_w stride_h stride_w height width . im2col kernel_h kernel_w stride_h stride_w $ input
30− in input' === out
34+ input === out
31
35
32prop_im2col_col2im_behaves_as_reference =
36prop_im2col_col2im_behaves_as_reference =
33 let ok extent kernel = [stride | stride <- [1..extent], (extent - kernel) `mod` stride == 0]
37 let ok extent kernel = [stride | stride <- [1..extent], (extent - kernel) `mod` stride == 0]
38+ in property $ do
34− in gamble (choose (2, 100)) $ \height ->
39+ height <- forAll (choose 2 100)
35− gamble (choose (2, 100)) $ \width ->
40+ width <- forAll (choose 2 100)
36− gamble (choose (2, height - 1)) $ \kernel_h ->
41+ kernel_h <- forAll (choose 2 (height - 1))
37− gamble (choose (2, width - 1)) $ \kernel_w ->
42+ kernel_w <- forAll (choose 2 (width - 1))
38− gamble (elements (ok height kernel_h)) $ \stride_h ->
43+ stride_h <- forAll (Gen.element (ok height kernel_h))
39− gamble (elements (ok width kernel_w)) $ \stride_w ->
44+ stride_w <- forAll (Gen.element (ok width kernel_w))
40− gamble (listOfN (height * width) (height * width) sizedRealFrac) $ \input ->
45+ input <- forAll ((height >< width) <$> Gen.list (Range.singleton $ height * width) (Gen.realFloat $ Range.linearFracFrom 0 (-100) 100))
41− let input' = (height >< width) input
46+
42− outFast = im2col kernel_h kernel_w stride_h stride_w input'
47+ let outFast = im2col kernel_h kernel_w stride_h stride_w input
43− retFast = col2im kernel_h kernel_w stride_h stride_w height width outFast
48+ let retFast = col2im kernel_h kernel_w stride_h stride_w height width outFast
44−
49+
45− outReference = Reference.im2col kernel_h kernel_w stride_h stride_w input'
50+ let outReference = Reference.im2col kernel_h kernel_w stride_h stride_w input
46− retReference = Reference.col2im kernel_h kernel_w stride_h stride_w height width outReference
51+ let retReference = Reference.col2im kernel_h kernel_w stride_h stride_w height width outReference
52+
47− in outFast === outReference .&&. retFast === retReference
53+ outFast === outReference
48−
54+ retFast === retReference
49−
55+
50−return []
51tests :: IO Bool
56tests :: IO Bool
52−tests = $quickCheckAll
57+tests = checkParallel $$(discover)
test/Test/Grenade/Layers/Internal/Pooling.hs+24−18
@@ -12,28 +12,34 @@ import Grenade.Layers.Internal.Pooling
12
12
13import Numeric.LinearAlgebra hiding (uniformSample, konst, (===))
13import Numeric.LinearAlgebra hiding (uniformSample, konst, (===))
14
14
15−import Disorder.Jack
15+import Hedgehog
16+import qualified Hedgehog.Gen as Gen
17+import qualified Hedgehog.Range as Range
16
18
17import qualified Test.Grenade.Layers.Internal.Reference as Reference
19import qualified Test.Grenade.Layers.Internal.Reference as Reference
20+import Test.Hedgehog.Compat
18
21
19prop_poolForwards_poolBackwards_behaves_as_reference =
22prop_poolForwards_poolBackwards_behaves_as_reference =
20 let ok extent kernel = [stride | stride <- [1..extent], (extent - kernel) `mod` stride == 0]
23 let ok extent kernel = [stride | stride <- [1..extent], (extent - kernel) `mod` stride == 0]
21 output extent kernel stride = (extent - kernel) `div` stride + 1
24 output extent kernel stride = (extent - kernel) `div` stride + 1
22− in gamble (choose (2, 100)) $ \height ->
25+ in property $ do
26+ height <- forAll $ choose 2 100
23− gamble (choose (2, 100)) $ \width ->
27+ width <- forAll $ choose 2 100
24− gamble (choose (1, height - 1)) $ \kernel_h ->
28+ kernel_h <- forAll $ choose 1 (height - 1)
25− gamble (choose (1, width - 1)) $ \kernel_w ->
29+ kernel_w <- forAll $ choose 1 (width - 1)
26− gamble (elements (ok height kernel_h)) $ \stride_h ->
30+ stride_h <- forAll $ Gen.element (ok height kernel_h)
27− gamble (elements (ok width kernel_w)) $ \stride_w ->
31+ stride_w <- forAll $ Gen.element (ok width kernel_w)
28− gamble (listOfN (height * width) (height * width) sizedRealFrac) $ \input ->
32+ input <- forAll $ (height >< width) <$> Gen.list (Range.singleton $ height * width) (Gen.realFloat $ Range.linearFracFrom 0 (-100) 100)
29− let input' = (height >< width) input
33+
30− outFast = poolForward 1 height width kernel_h kernel_w stride_h stride_w input'
34+ let outFast = poolForward 1 height width kernel_h kernel_w stride_h stride_w input
31− retFast = poolBackward 1 height width kernel_h kernel_w stride_h stride_w input' outFast
35+ let retFast = poolBackward 1 height width kernel_h kernel_w stride_h stride_w input outFast
32−
36+
33− outReference = Reference.poolForward kernel_h kernel_w stride_h stride_w (output height kernel_h stride_h) (output width kernel_w stride_w) input'
37+ let outReference = Reference.poolForward kernel_h kernel_w stride_h stride_w (output height kernel_h stride_h) (output width kernel_w stride_w) input
34− retReference = Reference.poolBackward kernel_h kernel_w stride_h stride_w input' outReference
38+ let retReference = Reference.poolBackward kernel_h kernel_w stride_h stride_w input outReference
39+
35− in outFast === outReference .&&. retFast === retReference
40+ outFast === outReference
36−
41+ retFast === retReference
37−return []
42+
43+
38tests :: IO Bool
44tests :: IO Bool
39−tests = $quickCheckAll
45+tests = checkParallel $$(discover)
test/Test/Grenade/Layers/Nonlinear.hsnew file+54
@@ -0,0 +1,54 @@
1+{-# LANGUAGE BangPatterns #-}
2+{-# LANGUAGE TemplateHaskell #-}
3+{-# LANGUAGE DataKinds #-}
4+{-# LANGUAGE KindSignatures #-}
5+{-# LANGUAGE GADTs #-}
6+{-# LANGUAGE ScopedTypeVariables #-}
7+{-# LANGUAGE LambdaCase #-}
8+{-# OPTIONS_GHC -fno-warn-missing-signatures #-}
9+module Test.Grenade.Layers.Nonlinear where
10+
11+import Data.Singletons
12+
13+import Grenade
14+
15+import Hedgehog
16+
17+import Test.Hedgehog.Compat
18+import Test.Hedgehog.Hmatrix
19+import Test.Hedgehog.TypeLits
20+
21+import Numeric.LinearAlgebra.Static ( norm_Inf )
22+
23+prop_sigmoid_grad :: Property
24+prop_sigmoid_grad = property $
25+ forAllWith rss genShape >>= \case
26+ (SomeSing (r :: Sing s)) ->
27+ withSingI r $
28+ blindForAll genOfShape >>= \(ds :: S s) ->
29+ let (tape, f :: S s) = runForwards Logit ds
30+ ((), ret :: S s) = runBackwards Logit tape (1 :: S s)
31+ (_, numer :: S s) = runForwards Logit (ds + 0.0001)
32+ numericalGradient = (numer - f) * 10000
33+ in assert ((case numericalGradient - ret of
34+ (S1D x) -> norm_Inf x < 0.0001
35+ (S2D x) -> norm_Inf x < 0.0001
36+ (S3D x) -> norm_Inf x < 0.0001) :: Bool)
37+
38+prop_tanh_grad :: Property
39+prop_tanh_grad = property $
40+ forAllWith rss genShape >>= \case
41+ (SomeSing (r :: Sing s)) ->
42+ withSingI r $
43+ blindForAll genOfShape >>= \(ds :: S s) ->
44+ let (tape, f :: S s) = runForwards Tanh ds
45+ ((), ret :: S s) = runBackwards Tanh tape (1 :: S s)
46+ (_, numer :: S s) = runForwards Tanh (ds + 0.0001)
47+ numericalGradient = (numer - f) * 10000
48+ in assert ((case numericalGradient - ret of
49+ (S1D x) -> norm_Inf x < 0.001
50+ (S2D x) -> norm_Inf x < 0.001
51+ (S3D x) -> norm_Inf x < 0.001) :: Bool)
52+
53+tests :: IO Bool
54+tests = checkParallel $$(discover)
test/Test/Grenade/Layers/PadCrop.hsnew file+49
@@ -0,0 +1,49 @@
1+{-# LANGUAGE BangPatterns #-}
2+{-# LANGUAGE CPP #-}
3+{-# LANGUAGE TemplateHaskell #-}
4+{-# LANGUAGE DataKinds #-}
5+{-# LANGUAGE KindSignatures #-}
6+{-# LANGUAGE GADTs #-}
7+{-# LANGUAGE ScopedTypeVariables #-}
8+{-# OPTIONS_GHC -fno-warn-missing-signatures #-}
9+
10+module Test.Grenade.Layers.PadCrop where
11+
12+import Grenade
13+
14+import Hedgehog
15+
16+import Numeric.LinearAlgebra.Static ( norm_Inf )
17+
18+import Test.Hedgehog.Hmatrix
19+
20+prop_pad_crop :: Property
21+prop_pad_crop =
22+ let net :: Network '[Pad 2 3 4 6, Crop 2 3 4 6] '[ 'D3 7 9 5, 'D3 16 15 5, 'D3 7 9 5 ]
23+ net = Pad :~> Crop :~> NNil
24+ in property $
25+ forAll genOfShape >>= \(d :: S ('D3 7 9 5)) ->
26+ let (tapes, res) = runForwards net d
27+ (_ , grad) = runBackwards net tapes d
28+ in do assert $ d ~~~ res
29+ assert $ grad ~~~ d
30+
31+prop_pad_crop_2d :: Property
32+prop_pad_crop_2d =
33+ let net :: Network '[Pad 2 3 4 6, Crop 2 3 4 6] '[ 'D2 7 9, 'D2 16 15, 'D2 7 9 ]
34+ net = Pad :~> Crop :~> NNil
35+ in property $
36+ forAll genOfShape >>= \(d :: S ('D2 7 9)) ->
37+ let (tapes, res) = runForwards net d
38+ (_ , grad) = runBackwards net tapes d
39+ in do assert $ d ~~~ res
40+ assert $ grad ~~~ d
41+
42+(~~~) :: S x -> S x -> Bool
43+(S1D x) ~~~ (S1D y) = norm_Inf (x - y) < 0.00001
44+(S2D x) ~~~ (S2D y) = norm_Inf (x - y) < 0.00001
45+(S3D x) ~~~ (S3D y) = norm_Inf (x - y) < 0.00001
46+
47+
48+tests :: IO Bool
49+tests = checkParallel $$(discover)
test/Test/Grenade/Layers/Pooling.hs+19−14
@@ -1,3 +1,4 @@
1+{-# LANGUAGE CPP #-}
1{-# LANGUAGE TemplateHaskell #-}
2{-# LANGUAGE TemplateHaskell #-}
2{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE DataKinds #-}
3{-# LANGUAGE KindSignatures #-}
4{-# LANGUAGE KindSignatures #-}
@@ -9,36 +10,40 @@ module Test.Grenade.Layers.Pooling where
9import Data.Proxy
10import Data.Proxy
10import Data.Singletons ()
11import Data.Singletons ()
11
12
13+#if MIN_VERSION_base(4,9,0)
14+import Data.Kind (Type)
15+#endif
16+
12import GHC.TypeLits
17import GHC.TypeLits
13import Grenade.Layers.Pooling
18import Grenade.Layers.Pooling
14
19
15−import Disorder.Jack
20+import Hedgehog
16
21
22+import Test.Hedgehog.Compat
17
23
18−data OpaquePooling :: * where
24+data OpaquePooling :: Type where
19 OpaquePooling :: (KnownNat kh, KnownNat kw, KnownNat sh, KnownNat sw) => Pooling kh kw sh sw -> OpaquePooling
25 OpaquePooling :: (KnownNat kh, KnownNat kw, KnownNat sh, KnownNat sw) => Pooling kh kw sh sw -> OpaquePooling
20
26
21instance Show OpaquePooling where
27instance Show OpaquePooling where
22 show (OpaquePooling n) = show n
28 show (OpaquePooling n) = show n
23
29
24−genOpaquePooling :: Jack OpaquePooling
30+genOpaquePooling :: Gen OpaquePooling
25genOpaquePooling = do
31genOpaquePooling = do
26− Just kernelHeight <- someNatVal <$> choose (2, 15)
32+ ~(Just kernelHeight) <- someNatVal <$> choose 2 15
27− Just kernelWidth <- someNatVal <$> choose (2, 15)
33+ ~(Just kernelWidth ) <- someNatVal <$> choose 2 15
28− Just strideHeight <- someNatVal <$> choose (2, 15)
34+ ~(Just strideHeight) <- someNatVal <$> choose 2 15
29− Just strideWidth <- someNatVal <$> choose (2, 15)
35+ ~(Just strideWidth ) <- someNatVal <$> choose 2 15
30
36
31 case (kernelHeight, kernelWidth, strideHeight, strideWidth) of
37 case (kernelHeight, kernelWidth, strideHeight, strideWidth) of
32 (SomeNat (_ :: Proxy kh), SomeNat (_ :: Proxy kw), SomeNat (_ :: Proxy sh), SomeNat (_ :: Proxy sw)) ->
38 (SomeNat (_ :: Proxy kh), SomeNat (_ :: Proxy kw), SomeNat (_ :: Proxy sh), SomeNat (_ :: Proxy sw)) ->
33 return $ OpaquePooling (Pooling :: Pooling kh kw sh sw)
39 return $ OpaquePooling (Pooling :: Pooling kh kw sh sw)
34
40
35prop_pool_layer_witness =
41prop_pool_layer_witness =
42+ property $ do
36− gamble genOpaquePooling $ \onet ->
43+ onet <- forAll genOpaquePooling
37− (case onet of
44+ case onet of
38− (OpaquePooling (Pooling :: Pooling kernelRows kernelCols strideRows strideCols)) -> True
45+ (OpaquePooling (Pooling :: Pooling kernelRows kernelCols strideRows strideCols)) ->
39− ) :: Bool
46+ assert True
40−
41
47
42−return []
43tests :: IO Bool
48tests :: IO Bool
44−tests = $quickCheckAll
49+tests = checkParallel $$(discover)
test/Test/Grenade/Network.hsnew file+504
@@ -0,0 +1,504 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE BangPatterns #-}
3+{-# LANGUAGE RankNTypes #-}
4+{-# LANGUAGE DataKinds #-}
5+{-# LANGUAGE ScopedTypeVariables #-}
6+{-# LANGUAGE PolyKinds #-}
7+{-# LANGUAGE GADTs #-}
8+{-# LANGUAGE TypeOperators #-}
9+{-# LANGUAGE TemplateHaskell #-}
10+{-# LANGUAGE FlexibleContexts #-}
11+module Test.Grenade.Network where
12+
13+import Control.Monad ( guard )
14+import Control.Monad.ST ( runST )
15+
16+import Data.Constraint
17+import qualified Data.Vector.Storable as VS
18+import qualified Data.Vector.Storable.Mutable as VS ( write )
19+import Data.Singletons
20+import Data.Singletons.Prelude.List
21+import Data.Singletons.TypeLits
22+
23+import Hedgehog
24+import qualified Hedgehog.Gen as Gen
25+import Hedgehog.Internal.Source
26+import Hedgehog.Internal.Property ( failWith )
27+
28+import Grenade
29+
30+#if MIN_VERSION_base(4,11,0)
31+import GHC.TypeLits hiding (natVal)
32+#else
33+import GHC.TypeLits
34+#endif
35+#if MIN_VERSION_base(4,9,0)
36+import Data.Kind (Type)
37+#endif
38+
39+import GHC.TypeLits.Witnesses
40+import Test.Hedgehog.Compat
41+import Test.Hedgehog.TypeLits
42+import Test.Hedgehog.Hmatrix
43+import Test.Grenade.Layers.Convolution
44+
45+import Numeric.LinearAlgebra ( flatten )
46+import Numeric.LinearAlgebra.Static ( extract, norm_Inf )
47+import Unsafe.Coerce
48+
49+data SomeNetwork :: Type where
50+ SomeNetwork :: ( SingI shapes, SingI (Head shapes), SingI (Last shapes), Show (Network layers shapes) ) => Network layers shapes -> SomeNetwork
51+
52+instance Show SomeNetwork where
53+ show (SomeNetwork net) = show net
54+
55+-- | Generate a random network of a random type
56+--
57+-- This is slightly insane for a few reasons. Everything must be wrapped up
58+-- in a SomeNetwork.
59+genNetwork :: Gen SomeNetwork
60+genNetwork =
61+ Gen.recursive Gen.choice [
62+ do SomeSing ( r :: Sing final ) <- genShape
63+ withSingI r $
64+ pure (SomeNetwork (NNil :: Network '[] '[ final ] ))
65+ ] [
66+ do SomeNetwork ( rest :: Network layers shapes ) <- genNetwork
67+ case ( sing :: Sing shapes ) of
68+ SNil -> Gen.discard -- Can't occur
69+ SCons ( h :: Sing h ) ( _ :: Sing hs ) ->
70+ withSingI h $
71+ case h of
72+ D1Sing l@SNat ->
73+ Gen.choice [
74+ pure (SomeNetwork (Tanh :~> rest :: Network ( Tanh ': layers ) ( h ': h ': hs )))
75+ , pure (SomeNetwork (Logit :~> rest :: Network ( Logit ': layers ) ( h ': h ': hs )))
76+ , pure (SomeNetwork (Relu :~> rest :: Network ( Relu ': layers ) ( h ': h ': hs )))
77+ , pure (SomeNetwork (Elu :~> rest :: Network ( Elu ': layers ) ( h ': h ': hs )))
78+ , pure (SomeNetwork (Softmax :~> rest :: Network ( Softmax ': layers ) ( h ': h ': hs )))
79+ , do -- Reshape to two dimensions
80+ let divisors :: Integer -> [Integer]
81+ divisors n = 1 : [x | x <- [2..(n-1)], n `rem` x == 0]
82+ let len = fromIntegral $ natVal l
83+ rs <- Gen.element $ divisors len
84+ let cs = len `quot` rs
85+ case ( someNatVal rs, someNatVal cs, someNatVal len ) of
86+ ( Just (SomeNat (rs' :: Proxy inRows)), Just (SomeNat (cs' :: Proxy inCols)), Just (SomeNat (_ :: Proxy outLen ) )) ->
87+ let p1 = natDict rs'
88+ p2 = natDict cs'
89+ in case ( p1 %* p2, unsafeCoerce (Dict :: Dict ()) :: Dict ((inRows * inCols) ~ outLen), unsafeCoerce (Dict :: Dict ()) :: Dict (( 'D1 outLen ) ~ h )) of
90+ ( Dict, Dict, Dict ) ->
91+ pure (SomeNetwork (Reshape :~> rest :: Network ( Reshape ': layers ) ( ('D2 inRows inCols) ': h ': hs )))
92+ _ -> Gen.discard -- Doesn't occur
93+ ]
94+ D2Sing r@SNat c@SNat ->
95+ Gen.choice [
96+ pure (SomeNetwork (Tanh :~> rest :: Network ( Tanh ': layers ) ( h ': h ': hs )))
97+ , pure (SomeNetwork (Logit :~> rest :: Network ( Logit ': layers ) ( h ': h ': hs )))
98+ , pure (SomeNetwork (Relu :~> rest :: Network ( Relu ': layers ) ( h ': h ': hs )))
99+ , pure (SomeNetwork (Elu :~> rest :: Network ( Elu ': layers ) ( h ': h ': hs )))
100+ , do -- Build a convolution layer with one filter output
101+ -- Figure out some kernel sizes which work for this layer
102+ -- There must be a better way than this...
103+ let output_r = fromIntegral $ natVal r
104+ let output_c = fromIntegral $ natVal c
105+
106+ let ok extent kernel = [stride | stride <- [ 1 .. extent ], (extent - kernel) `mod` stride == 0]
107+
108+ -- Get some kernels which will fit
109+ kernel_r <- choose 1 output_r
110+ kernel_c <- choose 1 output_c
111+
112+ -- Build up some strides which also fit
113+ stride_r <- Gen.element $ ok output_r kernel_r
114+ stride_c <- Gen.element $ ok output_c kernel_c
115+
116+ -- Determine the input size
117+ let input_r = (output_r - 1) * stride_r + kernel_r
118+ let input_c = (output_c - 1) * stride_c + kernel_c
119+
120+ guard (input_r < 100)
121+ guard (input_c < 100)
122+
123+ -- Remake types for input
124+ case ( someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal kernel_r, someNatVal kernel_c, someNatVal stride_r, someNatVal stride_c ) of
125+ ( Just (SomeNat (_ :: Proxy inRows)), Just (SomeNat (_ :: Proxy inCols)),
126+ Just (SomeNat (_ :: Proxy outRows)), Just (SomeNat (_ :: Proxy outCols)),
127+ Just (SomeNat (pkr :: Proxy kernelRows)), Just (SomeNat (pkc :: Proxy kernelCols)),
128+ Just (SomeNat (_ :: Proxy strideRows)), Just (SomeNat (_ :: Proxy strideCols))) ->
129+ let p1 = natDict pkr
130+ p2 = natDict pkc
131+ in case ( p1 %* p2
132+ -- Fake it till you make it.
133+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D2 outRows outCols ) ~ h ))
134+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outRows - 1) * strideRows) ~ (inRows - kernelRows)))
135+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outCols - 1) * strideCols) ~ (inCols - kernelCols)))) of
136+ (Dict, Dict, Dict, Dict) -> do
137+ conv <- genConvolution
138+ pure (SomeNetwork (conv :~> rest :: Network ( Convolution 1 1 kernelRows kernelCols strideRows strideCols ': layers ) ( ('D2 inRows inCols) ': h ': hs )))
139+ _ -> Gen.discard -- Can't occur
140+ , do -- Build a convolution layer with one filter output
141+ -- Figure out some kernel sizes which work for this layer
142+ -- There must be a better way than this...
143+ let output_r = fromIntegral $ natVal r
144+ let output_c = fromIntegral $ natVal c
145+
146+ let ok extent kernel = [stride | stride <- [ 1 .. extent ], (extent - kernel) `mod` stride == 0]
147+
148+ -- Get some kernels which will fit
149+ kernel_r <- choose 1 output_r
150+ kernel_c <- choose 1 output_c
151+ channels <- choose 1 10
152+
153+ -- Build up some strides which also fit
154+ stride_r <- Gen.element $ ok output_r kernel_r
155+ stride_c <- Gen.element $ ok output_c kernel_c
156+
157+ -- Determine the input size
158+ let input_r = (output_r - 1) * stride_r + kernel_r
159+ let input_c = (output_c - 1) * stride_c + kernel_c
160+
161+ guard (input_r < 100)
162+ guard (input_c < 100)
163+
164+ -- Remake types for input
165+ case ( someNatVal channels, someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal kernel_r, someNatVal kernel_c, someNatVal stride_r, someNatVal stride_c ) of
166+ ( Just (SomeNat (chan :: Proxy channels)),
167+ Just (SomeNat (pinr :: Proxy inRows)), Just (SomeNat (_ :: Proxy inCols)),
168+ Just (SomeNat (_ :: Proxy outRows)), Just (SomeNat (_ :: Proxy outCols)),
169+ Just (SomeNat (pkr :: Proxy kernelRows)), Just (SomeNat (pkc :: Proxy kernelCols)),
170+ Just (SomeNat (_ :: Proxy strideRows)), Just (SomeNat (_ :: Proxy strideCols))) ->
171+ let p1 = natDict pkr
172+ p2 = natDict pkc
173+ p3 = natDict chan
174+ in case ( p1 %* p2 %* p3
175+ , natDict pinr %* natDict chan
176+ -- Fake it till you make it.
177+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D2 outRows outCols ) ~ h ))
178+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outRows - 1) * strideRows) ~ (inRows - kernelRows)))
179+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outCols - 1) * strideCols) ~ (inCols - kernelCols)))) of
180+ (Dict, Dict, Dict, Dict, Dict) -> do
181+ conv <- genConvolution
182+ pure (SomeNetwork (conv :~> rest :: Network ( Convolution channels 1 kernelRows kernelCols strideRows strideCols ': layers ) ( ('D3 inRows inCols channels) ': h ': hs )))
183+ _ -> Gen.discard -- Can't occur
184+ , do -- Build a Pooling layer
185+ let output_r = fromIntegral $ natVal r
186+ let output_c = fromIntegral $ natVal c
187+
188+ let ok extent kernel = [stride | stride <- [ 1 .. extent ], (extent - kernel) `mod` stride == 0]
189+
190+ -- Get some kernels which will fit
191+ kernel_r <- choose 1 output_r
192+ kernel_c <- choose 1 output_c
193+
194+ -- Build up some strides which also fit
195+ stride_r <- Gen.element $ ok output_r kernel_r
196+ stride_c <- Gen.element $ ok output_c kernel_c
197+
198+ -- Determine the input size
199+ let input_r = (output_r - 1) * stride_r + kernel_r
200+ let input_c = (output_c - 1) * stride_c + kernel_c
201+
202+ guard (input_r < 100)
203+ guard (input_c < 100)
204+
205+ -- Remake types for input
206+ case ( someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal kernel_r, someNatVal kernel_c, someNatVal stride_r, someNatVal stride_c ) of
207+ ( Just (SomeNat (_ :: Proxy inRows)), Just (SomeNat (_ :: Proxy inCols)),
208+ Just (SomeNat (_ :: Proxy outRows)), Just (SomeNat (_ :: Proxy outCols)),
209+ Just (SomeNat (pkr :: Proxy kernelRows)), Just (SomeNat (pkc :: Proxy kernelCols)),
210+ Just (SomeNat (_ :: Proxy strideRows)), Just (SomeNat (_ :: Proxy strideCols))) ->
211+ let p1 = natDict pkr
212+ p2 = natDict pkc
213+ in case ( p1 %* p2
214+ -- Fake it till you make it.
215+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D2 outRows outCols ) ~ h ))
216+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outRows - 1) * strideRows) ~ (inRows - kernelRows)))
217+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outCols - 1) * strideCols) ~ (inCols - kernelCols)))) of
218+ (Dict, Dict, Dict, Dict) ->
219+ pure (SomeNetwork (Pooling :~> rest :: Network ( Pooling kernelRows kernelCols strideRows strideCols ': layers ) ( ('D2 inRows inCols) ': h ': hs )))
220+ _ -> Gen.discard -- Can't occur
221+ , do -- Build a Pad layer
222+ let output_r = fromIntegral $ natVal r
223+ let output_c = fromIntegral $ natVal c
224+
225+ pad_left <- choose 0 (output_r - 1)
226+ pad_right <- choose 0 (output_r - 1 - pad_left)
227+ pad_top <- choose 0 (output_c - 1)
228+ pad_bottom <- choose 0 (output_c - 1 - pad_top)
229+
230+ let input_r = output_r - pad_top - pad_bottom
231+ let input_c = output_c - pad_left - pad_right
232+
233+ guard (input_r > 1)
234+ guard (input_c > 1)
235+
236+ -- Remake types for input
237+ case ( someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal pad_left, someNatVal pad_right, someNatVal pad_top, someNatVal pad_bottom ) of
238+ ( Just (SomeNat (_ :: Proxy inputRows)), Just (SomeNat (_ :: Proxy inputColumns)),
239+ Just (SomeNat (_ :: Proxy outputRows)), Just (SomeNat (_ :: Proxy outputColumns)),
240+ Just (SomeNat (_ :: Proxy padLeft)), Just (SomeNat (_ :: Proxy padRight)),
241+ Just (SomeNat (_ :: Proxy padTop)), Just (SomeNat (_ :: Proxy padBottom))) ->
242+ case ( (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D2 outputRows outputColumns) ~ h ))
243+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (inputRows + padTop + padBottom) ~ outputRows) )
244+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (inputColumns + padLeft + padRight) ~ outputColumns ))) of
245+ ( Dict, Dict, Dict ) ->
246+ pure (SomeNetwork (Pad :~> rest :: Network ( Pad padLeft padTop padRight padBottom ': layers ) ( ('D2 inputRows inputColumns) ': h ': hs )))
247+ _ -> Gen.discard -- Can't occur
248+ , do -- Build a Crop layer
249+ let output_r = fromIntegral $ natVal r
250+ let output_c = fromIntegral $ natVal c
251+
252+ crop_left <- choose 0 10
253+ crop_right <- choose 0 10
254+ crop_top <- choose 0 10
255+ crop_bottom <- choose 0 10
256+
257+ let input_r = output_r + crop_top + crop_bottom
258+ let input_c = output_c + crop_left + crop_right
259+
260+ -- Remake types for input
261+ case ( someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal crop_left, someNatVal crop_right, someNatVal crop_top, someNatVal crop_bottom ) of
262+ ( Just (SomeNat (_ :: Proxy inputRows)), Just (SomeNat (_ :: Proxy inputColumns)),
263+ Just (SomeNat (_ :: Proxy outputRows)), Just (SomeNat (_ :: Proxy outputColumns)),
264+ Just (SomeNat (_ :: Proxy cropLeft)), Just (SomeNat (_ :: Proxy cropRight)),
265+ Just (SomeNat (_ :: Proxy cropTop)), Just (SomeNat (_ :: Proxy cropBottom))) ->
266+ case ( (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D2 outputRows outputColumns) ~ h ))
267+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (outputRows + cropTop + cropBottom) ~ inputRows ) )
268+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (outputColumns + cropLeft + cropRight) ~ inputColumns ))) of
269+ (Dict, Dict, Dict) ->
270+ pure (SomeNetwork (Crop :~> rest :: Network ( Crop cropLeft cropTop cropRight cropBottom ': layers ) ( ('D2 inputRows inputColumns) ': h ': hs )))
271+ _ -> Gen.discard -- Can't occur
272+ ]
273+ D3Sing r@SNat c@SNat f@SNat ->
274+ Gen.choice [
275+ pure (SomeNetwork (Tanh :~> rest :: Network ( Tanh ': layers ) ( h ': h ': hs )))
276+ , pure (SomeNetwork (Logit :~> rest :: Network ( Logit ': layers ) ( h ': h ': hs )))
277+ , pure (SomeNetwork (Relu :~> rest :: Network ( Relu ': layers ) ( h ': h ': hs )))
278+ , pure (SomeNetwork (Elu :~> rest :: Network ( Elu ': layers ) ( h ': h ': hs )))
279+ , do -- Build a convolution layer with one filter output
280+ -- Figure out some kernel sizes which work for this layer
281+ -- There must be a better way than this...
282+ let output_r = fromIntegral $ natVal r
283+ let output_c = fromIntegral $ natVal c
284+ let output_f = fromIntegral $ natVal f
285+
286+ let ok extent kernel = [stride | stride <- [ 1 .. extent ], (extent - kernel) `mod` stride == 0]
287+
288+ -- Get some kernels which will fit
289+ kernel_r <- choose 1 output_r
290+ kernel_c <- choose 1 output_c
291+ channels <- choose 1 10
292+
293+ -- Build up some strides which also fit
294+ stride_r <- Gen.element $ ok output_r kernel_r
295+ stride_c <- Gen.element $ ok output_c kernel_c
296+
297+ -- Determine the input size
298+ let input_r = (output_r - 1) * stride_r + kernel_r
299+ let input_c = (output_c - 1) * stride_c + kernel_c
300+
301+ guard (input_r < 100)
302+ guard (input_c < 100)
303+
304+ -- Remake types for input
305+ case ( someNatVal channels, someNatVal output_f, someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal kernel_r, someNatVal kernel_c, someNatVal stride_r, someNatVal stride_c ) of
306+ ( Just (SomeNat (chan :: Proxy channels)), Just (SomeNat (_ :: Proxy filters)),
307+ Just (SomeNat (pinr :: Proxy inRows)), Just (SomeNat (_ :: Proxy inCols)),
308+ Just (SomeNat (_ :: Proxy outRows)), Just (SomeNat (_ :: Proxy outCols)),
309+ Just (SomeNat (pkr :: Proxy kernelRows)), Just (SomeNat (pkc :: Proxy kernelCols)),
310+ Just (SomeNat (_ :: Proxy strideRows)), Just (SomeNat (_ :: Proxy strideCols))) ->
311+ let p1 = natDict pkr
312+ p2 = natDict pkc
313+ p3 = natDict chan
314+ in case ( p1 %* p2 %* p3
315+ , natDict pinr %* natDict chan
316+ -- Fake it till you make it.
317+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D3 outRows outCols filters) ~ h ))
318+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outRows - 1) * strideRows) ~ (inRows - kernelRows)))
319+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outCols - 1) * strideCols) ~ (inCols - kernelCols)))) of
320+ (Dict, Dict, Dict, Dict, Dict) -> do
321+ conv <- genConvolution
322+ pure (SomeNetwork (conv :~> rest :: Network ( Convolution channels filters kernelRows kernelCols strideRows strideCols ': layers ) ( ('D3 inRows inCols channels) ': h ': hs )))
323+ _ -> Gen.discard -- Can't occur
324+ , do -- Build a Pooling layer
325+ let output_r = fromIntegral $ natVal r
326+ let output_c = fromIntegral $ natVal c
327+ let output_f = fromIntegral $ natVal f
328+
329+ let ok extent kernel = [stride | stride <- [ 1 .. extent ], (extent - kernel) `mod` stride == 0]
330+
331+ -- Get some kernels which will fit
332+ kernel_r <- choose 1 output_r
333+ kernel_c <- choose 1 output_c
334+
335+ -- Build up some strides which also fit
336+ stride_r <- Gen.element $ ok output_r kernel_r
337+ stride_c <- Gen.element $ ok output_c kernel_c
338+
339+ -- Determine the input size
340+ let input_r = (output_r - 1) * stride_r + kernel_r
341+ let input_c = (output_c - 1) * stride_c + kernel_c
342+
343+ guard (input_r < 100)
344+ guard (input_c < 100)
345+
346+ -- Remake types for input
347+ case ( someNatVal output_f, someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal kernel_r, someNatVal kernel_c, someNatVal stride_r, someNatVal stride_c ) of
348+ ( Just (SomeNat (chan :: Proxy filters)),
349+ Just (SomeNat (pinr :: Proxy inRows)), Just (SomeNat (_ :: Proxy inCols)),
350+ Just (SomeNat (_ :: Proxy outRows)), Just (SomeNat (_ :: Proxy outCols)),
351+ Just (SomeNat (pkr :: Proxy kernelRows)), Just (SomeNat (pkc :: Proxy kernelCols)),
352+ Just (SomeNat (_ :: Proxy strideRows)), Just (SomeNat (_ :: Proxy strideCols))) ->
353+ let p1 = natDict pkr
354+ p2 = natDict pkc
355+ p3 = natDict chan
356+ in case ( p1 %* p2 %* p3
357+ , natDict pinr %* natDict chan
358+ -- Fake it till you make it.
359+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D3 outRows outCols filters) ~ h ))
360+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outRows - 1) * strideRows) ~ (inRows - kernelRows)))
361+ , (unsafeCoerce (Dict :: Dict ()) :: Dict (((outCols - 1) * strideCols) ~ (inCols - kernelCols)))) of
362+ (Dict, Dict, Dict, Dict, Dict) ->
363+ pure (SomeNetwork (Pooling :~> rest :: Network ( Pooling kernelRows kernelCols strideRows strideCols ': layers ) ( ('D3 inRows inCols filters) ': h ': hs )))
364+ _ -> Gen.discard -- Can't occur
365+ , do -- Build a Pad layer
366+ let output_r = fromIntegral $ natVal r
367+ let output_c = fromIntegral $ natVal c
368+ let output_f = fromIntegral $ natVal f
369+
370+ pad_left <- choose 0 (output_r - 1)
371+ pad_right <- choose 0 (output_r - 1 - pad_left)
372+ pad_top <- choose 0 (output_c - 1)
373+ pad_bottom <- choose 0 (output_c - 1 - pad_top)
374+
375+ let input_r = output_r - pad_top - pad_bottom
376+ let input_c = output_c - pad_left - pad_right
377+
378+ guard (input_r > 1)
379+ guard (input_c > 1)
380+
381+ -- Remake types for input
382+ case ( someNatVal output_f, someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal pad_left, someNatVal pad_right, someNatVal pad_top, someNatVal pad_bottom ) of
383+ ( Just (SomeNat (chan :: Proxy filters)),
384+ Just (SomeNat (pinr :: Proxy inputRows)), Just (SomeNat (_ :: Proxy inputColumns)),
385+ Just (SomeNat (_ :: Proxy outputRows)), Just (SomeNat (_ :: Proxy outputColumns)),
386+ Just (SomeNat (_ :: Proxy padLeft)), Just (SomeNat (_ :: Proxy padRight)),
387+ Just (SomeNat (_ :: Proxy padTop)), Just (SomeNat (_ :: Proxy padBottom))) ->
388+ case ( natDict pinr %* natDict chan
389+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D3 outputRows outputColumns filters) ~ h ))
390+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (inputRows + padTop + padBottom) ~ outputRows) )
391+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (inputColumns + padLeft + padRight) ~ outputColumns ))) of
392+ (Dict, Dict, Dict, Dict) ->
393+ pure (SomeNetwork (Pad :~> rest :: Network ( Pad padLeft padTop padRight padBottom ': layers ) ( ('D3 inputRows inputColumns filters) ': h ': hs )))
394+ _ -> Gen.discard -- Can't occur
395+ , do -- Build a Crop layer
396+ let output_r = fromIntegral $ natVal r
397+ let output_c = fromIntegral $ natVal c
398+ let output_f = fromIntegral $ natVal f
399+
400+ crop_left <- choose 0 10
401+ crop_right <- choose 0 10
402+ crop_top <- choose 0 10
403+ crop_bottom <- choose 0 10
404+
405+ let input_r = output_r + crop_top + crop_bottom
406+ let input_c = output_c + crop_left + crop_right
407+
408+ -- Remake types for input
409+ case ( someNatVal output_f, someNatVal input_r, someNatVal input_c, someNatVal output_r, someNatVal output_c, someNatVal crop_left, someNatVal crop_right, someNatVal crop_top, someNatVal crop_bottom ) of
410+ ( Just (SomeNat (chan :: Proxy filters)),
411+ Just (SomeNat (pinr :: Proxy inputRows)), Just (SomeNat (_ :: Proxy inputColumns)),
412+ Just (SomeNat (_ :: Proxy outputRows)), Just (SomeNat (_ :: Proxy outputColumns)),
413+ Just (SomeNat (_ :: Proxy cropLeft)), Just (SomeNat (_ :: Proxy cropRight)),
414+ Just (SomeNat (_ :: Proxy cropTop)), Just (SomeNat (_ :: Proxy cropBottom))) ->
415+ case ( natDict pinr %* natDict chan
416+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( ( 'D3 outputRows outputColumns filters) ~ h ))
417+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (outputRows + cropTop + cropBottom) ~ inputRows ) )
418+ , (unsafeCoerce (Dict :: Dict ()) :: Dict ( (outputColumns + cropLeft + cropRight) ~ inputColumns ))) of
419+ (Dict, Dict, Dict, Dict) ->
420+ pure (SomeNetwork (Crop :~> rest :: Network ( Crop cropLeft cropTop cropRight cropBottom ': layers ) ( ('D3 inputRows inputColumns filters) ': h ': hs )))
421+ _ -> Gen.discard -- Can't occur
422+ ]
423+ ]
424+
425+-- | Test a partial derivative numerically for a random network and input
426+--
427+-- This is the most important test.
428+prop_auto_diff :: Property
429+prop_auto_diff = withDiscards 1000 . withTests 10000 . property $ do
430+ SomeNetwork (network :: Network layers shapes) <- forAll genNetwork
431+ (input :: S (Head shapes)) <- forAllWith nice genOfShape
432+ (target :: S (Last shapes)) <- forAllWith nice oneUp
433+ (tested :: S (Head shapes)) <- forAllWith nice oneUp
434+
435+ let (!tapes, !output) = runNetwork network input
436+ let (_, !backgrad) = runGradient network tapes target
437+ let inputDiff = input + tested * 0.00001
438+ let expected = maxVal ( backgrad * tested )
439+ let (_, !outputDiff) = runNetwork network inputDiff
440+ let result = maxVal ( outputDiff * target - output * target ) / 0.00001
441+
442+ result ~~~ expected
443+
444+-- Make a shape where all are 0 except for 1 value, which is 1.
445+oneUp :: forall shape. ( SingI shape ) => Gen (S shape)
446+oneUp =
447+ case ( sing :: Sing shape ) of
448+ D1Sing SNat ->
449+ let x = 0 :: S ( shape )
450+ in case x of
451+ ( S1D x' ) -> do
452+ let ex = extract x'
453+ let len = VS.length ex
454+ ix <- choose 0 (len - 1)
455+ let nx = runST $ do ex' <- VS.thaw ex
456+ VS.write ex' ix 1.0
457+ VS.freeze ex'
458+ maybe Gen.discard pure . fromStorable $ nx
459+
460+ D2Sing SNat SNat ->
461+ let x = 0 :: S ( shape )
462+ in case x of
463+ ( S2D x' ) -> do
464+ let ex = flatten ( extract x' )
465+ let len = VS.length ex
466+ ix <- choose 0 (len - 1)
467+ let nx = runST $ do ex' <- VS.thaw ex
468+ VS.write ex' ix 1
469+ VS.freeze ex'
470+ maybe Gen.discard pure . fromStorable $ nx
471+
472+ D3Sing SNat SNat SNat ->
473+ let x = 0 :: S ( shape )
474+ in case x of
475+ ( S3D x' ) -> do
476+ let ex = flatten ( extract x' )
477+ let len = VS.length ex
478+ ix <- choose 0 (len - 1)
479+ let nx = runST $ do ex' <- VS.thaw ex
480+ VS.write ex' ix 1
481+ VS.freeze ex'
482+ maybe Gen.discard pure . fromStorable $ nx
483+
484+maxVal :: S shape -> Double
485+maxVal ( S1D x ) = norm_Inf x
486+maxVal ( S2D x ) = norm_Inf x
487+maxVal ( S3D x ) = norm_Inf x
488+
489+(~~~) :: (Monad m, HasCallStack) => Double -> Double -> PropertyT m ()
490+(~~~) x y =
491+ if abs (x - y) < 2e-5 then
492+ success
493+ else
494+ withFrozenCallStack $
495+ failWith Nothing $ unlines [
496+ "━━━ Not Simliar ━━━"
497+ , show x
498+ , show y
499+ ]
500+infix 4 ~~~
501+
502+
503+tests :: IO Bool
504+tests = checkParallel $$(discover)
test/Test/Grenade/Recurrent/Layers/LSTM.hs+83−54
@@ -10,7 +10,10 @@
10{-# OPTIONS_GHC -fno-warn-missing-signatures #-}
10{-# OPTIONS_GHC -fno-warn-missing-signatures #-}
11module Test.Grenade.Recurrent.Layers.LSTM where
11module Test.Grenade.Recurrent.Layers.LSTM where
12
12
13−import Disorder.Jack
13+import Hedgehog
14+import Hedgehog.Internal.Source
15+import Hedgehog.Internal.Show
16+import Hedgehog.Internal.Property ( failWith, Diff (..) )
14
17
15import Data.Foldable ( toList )
18import Data.Foldable ( toList )
16import Data.Singletons.TypeLits
19import Data.Singletons.TypeLits
@@ -23,9 +26,9 @@ import qualified Numeric.LinearAlgebra.Static as S
23
26
24
27
25import qualified Test.Grenade.Recurrent.Layers.LSTM.Reference as Reference
28import qualified Test.Grenade.Recurrent.Layers.LSTM.Reference as Reference
26−import Test.Jack.Hmatrix
29+import Test.Hedgehog.Hmatrix
27
30
28−genLSTM :: forall i o. (KnownNat i, KnownNat o) => Jack (LSTM i o)
31+genLSTM :: forall i o. (KnownNat i, KnownNat o) => Gen (LSTM i o)
29genLSTM = do
32genLSTM = do
30 let w = uniformSample
33 let w = uniformSample
31 u = uniformSample
34 u = uniformSample
@@ -39,63 +42,89 @@ genLSTM = do
39 <*> pure (LSTMWeights w0 u0 v0 w0 u0 v0 w0 u0 v0 w0 v0)
42 <*> pure (LSTMWeights w0 u0 v0 w0 u0 v0 w0 u0 v0 w0 v0)
40
43
41prop_lstm_reference_forwards =
44prop_lstm_reference_forwards =
42− gamble randomVector $ \(input :: S.R 3) ->
45+ property $ do
43− gamble randomVector $ \(cell :: S.R 2) ->
46+ input :: S.R 3 <- forAll randomVector
47+ cell :: S.R 2 <- forAll randomVector
44− gamble genLSTM $ \(net@(LSTM lstmWeights _) :: LSTM 3 2) ->
48+ net@(LSTM lstmWeights _) :: LSTM 3 2 <- forAll genLSTM
49+
45− let actual = runRecurrentForwards net (S1D cell) (S1D input)
50+ let actual = runRecurrentForwards net (S1D cell) (S1D input)
46− in case actual of
51+ case actual of
47− ((S1D cellOut) :: S ('D1 2), (S1D output) :: S ('D1 2)) ->
52+ (_, (S1D cellOut) :: S ('D1 2), (S1D output) :: S ('D1 2)) ->
48− let cellOut' = Reference.Vector . H.toList . S.extract $ cellOut
53+ let cellOut' = Reference.Vector . H.toList . S.extract $ cellOut
49− output' = Reference.Vector . H.toList . S.extract $ output
54+ output' = Reference.Vector . H.toList . S.extract $ output
50− refNet = Reference.lstmToReference lstmWeights
55+ refNet = Reference.lstmToReference lstmWeights
51− refCell = Reference.Vector . H.toList . S.extract $ cell
56+ refCell = Reference.Vector . H.toList . S.extract $ cell
52− refInput = Reference.Vector . H.toList . S.extract $ input
57+ refInput = Reference.Vector . H.toList . S.extract $ input
53− (refCO, refO) = Reference.runLSTM refNet refCell refInput
58+ (refCO, refO) = Reference.runLSTM refNet refCell refInput
54− in toList refCO ~~~ toList cellOut' .&&. toList refO ~~~ toList output'
59+ in do toList refCO ~~~ toList cellOut'
60+ toList refO ~~~ toList output'
61+
55
62
56prop_lstm_reference_backwards =
63prop_lstm_reference_backwards =
57− gamble randomVector $ \(input :: S.R 3) ->
64+ property $ do
58− gamble randomVector $ \(cell :: S.R 2) ->
65+ input :: S.R 3 <- forAll randomVector
59− gamble genLSTM $ \(net@(LSTM lstmWeights _) :: LSTM 3 2) ->
66+ cell :: S.R 2 <- forAll randomVector
67+ net@(LSTM lstmWeights _) :: LSTM 3 2 <- forAll genLSTM
68+ let (tape, _ :: S ('D1 2), _ :: S ('D1 2))
69+ = runRecurrentForwards net (S1D cell) (S1D input)
60− let actualBacks = runRecurrentBackwards net (S1D cell) (S1D input) (S1D (S.konst 1) :: S ('D1 2)) (S1D (S.konst 1) :: S ('D1 2))
70+ actualBacks = runRecurrentBackwards net tape (S1D (S.konst 1) :: S ('D1 2)) (S1D (S.konst 1) :: S ('D1 2))
61− in case actualBacks of
71+ case actualBacks of
62− (actualGradients, _, _) ->
72+ (actualGradients, _, _ :: S ('D1 3)) ->
63− let refNet = Reference.lstmToReference lstmWeights
73+ let refNet = Reference.lstmToReference lstmWeights
64− refCell = Reference.Vector . H.toList . S.extract $ cell
74+ refCell = Reference.Vector . H.toList . S.extract $ cell
65− refInput = Reference.Vector . H.toList . S.extract $ input
75+ refInput = Reference.Vector . H.toList . S.extract $ input
66− refGradients = Reference.runLSTMback refCell refInput refNet
76+ refGradients = Reference.runLSTMback refCell refInput refNet
67− in toList refGradients ~~~ toList (Reference.lstmToReference actualGradients)
77+ in toList refGradients ~~~ toList (Reference.lstmToReference actualGradients)
68
78
69prop_lstm_reference_backwards_input =
79prop_lstm_reference_backwards_input =
70− gamble randomVector $ \(input :: S.R 3) ->
80+ property $ do
71− gamble randomVector $ \(cell :: S.R 2) ->
81+ input :: S.R 3 <- forAll randomVector
72− gamble genLSTM $ \(net@(LSTM lstmWeights _) :: LSTM 3 2) ->
82+ cell :: S.R 2 <- forAll randomVector
83+ net@(LSTM lstmWeights _) :: LSTM 3 2 <- forAll genLSTM
84+ let (tape, _ :: S ('D1 2), _ :: S ('D1 2))
85+ = runRecurrentForwards net (S1D cell) (S1D input)
73− let actualBacks = runRecurrentBackwards net (S1D cell) (S1D input) (S1D (S.konst 1) :: S ('D1 2)) (S1D (S.konst 1) :: S ('D1 2))
86+ actualBacks = runRecurrentBackwards net tape (S1D (S.konst 1) :: S ('D1 2)) (S1D (S.konst 1) :: S ('D1 2))
74− in case actualBacks of
87+ case actualBacks of
75− (_, _, S1D actualGradients) ->
88+ (_, _, S1D actualGradients :: S ('D1 3)) ->
76− let refNet = Reference.lstmToReference lstmWeights
89+ let refNet = Reference.lstmToReference lstmWeights
77− refCell = Reference.Vector . H.toList . S.extract $ cell
90+ refCell = Reference.Vector . H.toList . S.extract $ cell
78− refInput = Reference.Vector . H.toList . S.extract $ input
91+ refInput = Reference.Vector . H.toList . S.extract $ input
79− refGradients = Reference.runLSTMbackOnInput refCell refNet refInput
92+ refGradients = Reference.runLSTMbackOnInput refCell refNet refInput
80− in toList refGradients ~~~ H.toList (S.extract actualGradients)
93+ in toList refGradients ~~~ H.toList (S.extract actualGradients)
81
94
82prop_lstm_reference_backwards_cell =
95prop_lstm_reference_backwards_cell =
83− gamble randomVector $ \(input :: S.R 3) ->
96+ property $ do
84− gamble randomVector $ \(cell :: S.R 2) ->
97+ input :: S.R 3 <- forAll randomVector
85− gamble genLSTM $ \(net@(LSTM lstmWeights _) :: LSTM 3 2) ->
98+ cell :: S.R 2 <- forAll randomVector
99+ net@(LSTM lstmWeights _) :: LSTM 3 2 <- forAll genLSTM
100+ let (tape, _ :: S ('D1 2), _ :: S ('D1 2))
101+ = runRecurrentForwards net (S1D cell) (S1D input)
86− let actualBacks = runRecurrentBackwards net (S1D cell) (S1D input) (S1D (S.konst 1) :: S ('D1 2)) (S1D (S.konst 1) :: S ('D1 2))
102+ actualBacks = runRecurrentBackwards net tape (S1D (S.konst 1) :: S ('D1 2)) (S1D (S.konst 1) :: S ('D1 2))
87− in case actualBacks of
103+ case actualBacks of
88− (_, S1D actualGradients, _) ->
104+ (_, S1D actualGradients, _ :: S ('D1 3)) ->
89− let refNet = Reference.lstmToReference lstmWeights
105+ let refNet = Reference.lstmToReference lstmWeights
90− refCell = Reference.Vector . H.toList . S.extract $ cell
106+ refCell = Reference.Vector . H.toList . S.extract $ cell
91− refInput = Reference.Vector . H.toList . S.extract $ input
107+ refInput = Reference.Vector . H.toList . S.extract $ input
92− refGradients = Reference.runLSTMbackOnCell refInput refNet refCell
108+ refGradients = Reference.runLSTMbackOnCell refInput refNet refCell
93− in toList refGradients ~~~ H.toList (S.extract actualGradients)
109+ in toList refGradients ~~~ H.toList (S.extract actualGradients)
94−
110+
95−
111+(~~~) :: (Monad m, Eq a, Ord a, Num a, Fractional a, Show a, HasCallStack) => [a] -> [a] -> PropertyT m ()
112+(~~~) x y =
96−(~~~) as bs = all (< 1e-8) (zipWith (-) as bs)
113+ if all (< 1e-8) (zipWith (-) x y) then
114+ success
115+ else
116+ case valueDiff <$> mkValue x <*> mkValue y of
117+ Nothing ->
118+ withFrozenCallStack $
119+ failWith Nothing $ unlines [
120+ "━━━ Not Simliar ━━━"
121+ , showPretty x
122+ , showPretty y
123+ ]
124+ Just differ ->
125+ withFrozenCallStack $
126+ failWith (Just $ Diff "Failed (" "- lhs" "~/~" "+ rhs" ")" differ) ""
97infix 4 ~~~
127infix 4 ~~~
98
128
99−return []
100tests :: IO Bool
129tests :: IO Bool
101−tests = $quickCheckAll
130+tests = checkParallel $$(discover)
test/Test/Grenade/Recurrent/Layers/LSTM/Reference.hs+25−21
@@ -6,7 +6,6 @@
6{-# LANGUAGE DeriveFunctor #-}
6{-# LANGUAGE DeriveFunctor #-}
7{-# LANGUAGE DeriveFoldable #-}
7{-# LANGUAGE DeriveFoldable #-}
8{-# LANGUAGE DeriveTraversable #-}
8{-# LANGUAGE DeriveTraversable #-}
9−{-# LANGUAGE RecordWildCards #-}
10{-# LANGUAGE FlexibleContexts #-}
9{-# LANGUAGE FlexibleContexts #-}
11{-# LANGUAGE RankNTypes #-}
10{-# LANGUAGE RankNTypes #-}
12
11
@@ -15,9 +14,13 @@ module Test.Grenade.Recurrent.Layers.LSTM.Reference where
15
14
16import Data.Reflection
15import Data.Reflection
17import Numeric.AD.Mode.Reverse
16import Numeric.AD.Mode.Reverse
18−import Numeric.AD.Internal.Reverse ( Tape )
17+import Numeric.AD.Internal.Reverse (Tape)
19
18
19+import GHC.TypeLits (KnownNat)
20+
21+import Grenade.Recurrent.Layers.LSTM (LSTMWeights (..))
20import qualified Grenade.Recurrent.Layers.LSTM as LSTM
22import qualified Grenade.Recurrent.Layers.LSTM as LSTM
23+
21import qualified Numeric.LinearAlgebra.Static as S
24import qualified Numeric.LinearAlgebra.Static as S
22import qualified Numeric.LinearAlgebra as H
25import qualified Numeric.LinearAlgebra as H
23
26
@@ -54,31 +57,32 @@ data RefLSTM a = RefLSTM
54 , refLstmBc :: Vector a -- Bias Cell (b_c)
57 , refLstmBc :: Vector a -- Bias Cell (b_c)
55 } deriving (Functor, Foldable, Traversable, Eq, Show)
58 } deriving (Functor, Foldable, Traversable, Eq, Show)
56
59
57−lstmToReference :: LSTM.LSTMWeights a b -> RefLSTM Double
60+lstmToReference :: (KnownNat a, KnownNat b) => LSTM.LSTMWeights a b -> RefLSTM Double
58−lstmToReference LSTM.LSTMWeights {..} =
61+lstmToReference lw =
62+ RefLSTM
59− let refLstmWf = Matrix . H.toLists . S.extract $ lstmWf -- Weight Forget (W_f)
63+ { refLstmWf = Matrix . H.toLists . S.extract $ lstmWf lw -- Weight Forget (W_f)
60− refLstmUf = Matrix . H.toLists . S.extract $ lstmUf -- Cell State Forget (U_f)
64+ , refLstmUf = Matrix . H.toLists . S.extract $ lstmUf lw -- Cell State Forget (U_f)
61− refLstmBf = Vector . H.toList . S.extract $ lstmBf -- Bias Forget (b_f)
65+ , refLstmBf = Vector . H.toList . S.extract $ lstmBf lw -- Bias Forget (b_f)
62− refLstmWi = Matrix . H.toLists . S.extract $ lstmWi -- Weight Input (W_i)
66+ , refLstmWi = Matrix . H.toLists . S.extract $ lstmWi lw -- Weight Input (W_i)
63− refLstmUi = Matrix . H.toLists . S.extract $ lstmUi -- Cell State Input (U_i)
67+ , refLstmUi = Matrix . H.toLists . S.extract $ lstmUi lw -- Cell State Input (U_i)
64− refLstmBi = Vector . H.toList . S.extract $ lstmBi -- Bias Input (b_i)
68+ , refLstmBi = Vector . H.toList . S.extract $ lstmBi lw -- Bias Input (b_i)
65− refLstmWo = Matrix . H.toLists . S.extract $ lstmWo -- Weight Output (W_o)
69+ , refLstmWo = Matrix . H.toLists . S.extract $ lstmWo lw -- Weight Output (W_o)
66− refLstmUo = Matrix . H.toLists . S.extract $ lstmUo -- Cell State Output (U_o)
70+ , refLstmUo = Matrix . H.toLists . S.extract $ lstmUo lw -- Cell State Output (U_o)
67− refLstmBo = Vector . H.toList . S.extract $ lstmBo -- Bias Output (b_o)
71+ , refLstmBo = Vector . H.toList . S.extract $ lstmBo lw -- Bias Output (b_o)
68− refLstmWc = Matrix . H.toLists . S.extract $ lstmWc -- Weight Cell (W_c)
72+ , refLstmWc = Matrix . H.toLists . S.extract $ lstmWc lw -- Weight Cell (W_c)
69− refLstmBc = Vector . H.toList . S.extract $ lstmBc -- Bias Cell (b_c)
73+ , refLstmBc = Vector . H.toList . S.extract $ lstmBc lw -- Bias Cell (b_c)
70− in RefLSTM {..}
74+ }
71
75
72runLSTM :: Floating a => RefLSTM a -> Vector a -> Vector a -> (Vector a, Vector a)
76runLSTM :: Floating a => RefLSTM a -> Vector a -> Vector a -> (Vector a, Vector a)
73−runLSTM RefLSTM {..} cell input =
77+runLSTM rl cell input =
74 let -- Forget state vector
78 let -- Forget state vector
75− f_t = sigmoid $ refLstmBf #+ refLstmWf #> input #+ refLstmUf #> cell
79+ f_t = sigmoid $ refLstmBf rl #+ refLstmWf rl #> input #+ refLstmUf rl #> cell
76 -- Input state vector
80 -- Input state vector
77− i_t = sigmoid $ refLstmBi #+ refLstmWi #> input #+ refLstmUi #> cell
81+ i_t = sigmoid $ refLstmBi rl #+ refLstmWi rl #> input #+ refLstmUi rl #> cell
78 -- Output state vector
82 -- Output state vector
79− o_t = sigmoid $ refLstmBo #+ refLstmWo #> input #+ refLstmUo #> cell
83+ o_t = sigmoid $ refLstmBo rl #+ refLstmWo rl #> input #+ refLstmUo rl #> cell
80 -- Cell input state vector
84 -- Cell input state vector
81− c_x = fmap tanh $ refLstmBc #+ refLstmWc #> input
85+ c_x = fmap tanh $ refLstmBc rl #+ refLstmWc rl #> input
82 -- Cell state
86 -- Cell state
83 c_t = f_t #* cell #+ i_t #* c_x
87 c_t = f_t #* cell #+ i_t #* c_x
84 -- Output (it's sometimes recommended to use tanh c_t)
88 -- Output (it's sometimes recommended to use tanh c_t)
test/Test/Hedgehog/Compat.hsnew file+21
@@ -0,0 +1,21 @@
1+{-# LANGUAGE RankNTypes #-}
2+module Test.Hedgehog.Compat (
3+ (...)
4+ , choose
5+ , blindForAll
6+ )where
7+
8+import Hedgehog (Gen)
9+import qualified Hedgehog.Gen as Gen
10+import qualified Hedgehog.Range as Range
11+import Hedgehog.Internal.Property
12+
13+(...) :: (c -> d) -> (a -> b -> c) -> a -> b -> d
14+(...) = (.) . (.)
15+{-# INLINE (...) #-}
16+
17+choose :: Integral a => a -> a -> Gen a
18+choose = Gen.integral ... Range.constant
19+
20+blindForAll :: Monad m => Gen a -> PropertyT m a
21+blindForAll = forAllWith (const "blind")
test/Test/Hedgehog/Hmatrix.hsnew file+42
@@ -0,0 +1,42 @@
1+{-# LANGUAGE DataKinds #-}
2+{-# LANGUAGE KindSignatures #-}
3+{-# LANGUAGE GADTs #-}
4+{-# LANGUAGE RankNTypes #-}
5+{-# LANGUAGE ScopedTypeVariables #-}
6+
7+module Test.Hedgehog.Hmatrix where
8+
9+import Grenade
10+import Data.Singletons
11+import Data.Singletons.TypeLits
12+
13+import Hedgehog (Gen)
14+import qualified Hedgehog.Gen as Gen
15+import qualified Hedgehog.Range as Range
16+
17+import qualified Numeric.LinearAlgebra.Static as HStatic
18+
19+randomVector :: forall n. ( KnownNat n ) => Gen (HStatic.R n)
20+randomVector = (\s -> HStatic.randomVector s HStatic.Uniform * 2 - 1) <$> Gen.int Range.linearBounded
21+
22+uniformSample :: forall m n. ( KnownNat m, KnownNat n ) => Gen (HStatic.L m n)
23+uniformSample = (\s -> HStatic.uniformSample s (-1) 1 ) <$> Gen.int Range.linearBounded
24+
25+-- | Generate random data of the desired shape
26+genOfShape :: forall x. ( SingI x ) => Gen (S x)
27+genOfShape =
28+ case (sing :: Sing x) of
29+ D1Sing l ->
30+ withKnownNat l $
31+ S1D <$> randomVector
32+ D2Sing r c ->
33+ withKnownNat r $ withKnownNat c $
34+ S2D <$> uniformSample
35+ D3Sing r c d ->
36+ withKnownNat r $ withKnownNat c $ withKnownNat d $
37+ S3D <$> uniformSample
38+
39+nice :: S shape -> String
40+nice (S1D x) = show . HStatic.extract $ x
41+nice (S2D x) = show . HStatic.extract $ x
42+nice (S3D x) = show . HStatic.extract $ x
test/Test/Hedgehog/TypeLits.hsnew file+65
@@ -0,0 +1,65 @@
1+{-# LANGUAGE CPP #-}
2+{-# LANGUAGE RankNTypes #-}
3+{-# LANGUAGE DataKinds #-}
4+{-# LANGUAGE ScopedTypeVariables #-}
5+{-# LANGUAGE PolyKinds #-}
6+{-# LANGUAGE TypeOperators #-}
7+{-# LANGUAGE GADTs #-}
8+module Test.Hedgehog.TypeLits where
9+
10+import Data.Constraint (Dict (..))
11+import Data.Singletons (Proxy (..), Sing (..), SomeSing (..), sing)
12+#if MIN_VERSION_singletons(2,6,0)
13+import Data.Singletons.TypeLits (SNat (..))
14+#endif
15+
16+import Hedgehog (Gen)
17+import qualified Hedgehog.Gen as Gen
18+
19+import Grenade
20+
21+import GHC.TypeLits (SomeNat (..), natVal, someNatVal)
22+import GHC.TypeLits.Witnesses ((%*), natDict)
23+import Test.Hedgehog.Compat (choose)
24+
25+genNat :: Gen SomeNat
26+genNat = do
27+ ~(Just n) <- someNatVal <$> choose 1 10
28+ return n
29+
30+genShape :: Gen (SomeSing Shape)
31+genShape
32+ = Gen.choice [
33+ genD1
34+ , genD2
35+ , genD3
36+ ]
37+
38+genD1 :: Gen (SomeSing Shape)
39+genD1 = do
40+ n <- genNat
41+ return $ case n of
42+ SomeNat (_ :: Proxy x) -> SomeSing (sing :: Sing ('D1 x))
43+
44+genD2 :: Gen (SomeSing Shape)
45+genD2 = do
46+ n <- genNat
47+ m <- genNat
48+ return $ case (n, m) of
49+ (SomeNat (_ :: Proxy x), SomeNat (_ :: Proxy y)) -> SomeSing (sing :: Sing ('D2 x y))
50+
51+genD3 :: Gen (SomeSing Shape)
52+genD3 = do
53+ n <- genNat
54+ m <- genNat
55+ o <- genNat
56+ return $ case (n, m, o) of
57+ (SomeNat (px :: Proxy x), SomeNat (_ :: Proxy y), SomeNat (pz :: Proxy z)) ->
58+ case natDict px %* natDict pz of
59+ Dict -> SomeSing (sing :: Sing ('D3 x y z))
60+
61+rss :: SomeSing Shape -> String
62+rss (SomeSing (r :: Sing s)) = case r of
63+ (D1Sing a@SNat) -> "D1 " ++ show (natVal a)
64+ (D2Sing a@SNat b@SNat) -> "D2 " ++ show (natVal a) ++ " " ++ show (natVal b)
65+ (D3Sing a@SNat b@SNat c@SNat) -> "D3 " ++ show (natVal a) ++ " " ++ show (natVal b) ++ " " ++ show (natVal c)
test/Test/Jack/Hmatrix.hsdeleted−17
@@ -1,17 +0,0 @@
1−{-# LANGUAGE RankNTypes #-}
2−{-# LANGUAGE DataKinds #-}
3−{-# LANGUAGE ScopedTypeVariables #-}
4−
5−module Test.Jack.Hmatrix where
6−
7−import Disorder.Jack
8−
9−import GHC.TypeLits
10−
11−import qualified Numeric.LinearAlgebra.Static as HStatic
12−
13−randomVector :: forall n. KnownNat n => Jack (HStatic.R n)
14−randomVector = (\s -> HStatic.randomVector s HStatic.Uniform * 2 - 1) <$> sizedNat
15−
16−uniformSample :: forall m n. (KnownNat m, KnownNat n) => Jack (HStatic.L m n)
17−uniformSample = (\s -> HStatic.uniformSample s (-1) 1 ) <$> sizedNat
test/Test/Jack/TypeLits.hsdeleted−14
@@ -1,14 +0,0 @@
1−{-# LANGUAGE RankNTypes #-}
2−{-# LANGUAGE DataKinds #-}
3−{-# LANGUAGE ScopedTypeVariables #-}
4−
5−module Test.Jack.TypeLits where
6−
7−import Disorder.Jack
8−
9−import GHC.TypeLits
10−
11−genNat :: Jack SomeNat
12−genNat = do
13− Just n <- someNatVal <$> choose (1, 10)
14− return n
test/test.hs+26−2
@@ -1,23 +1,47 @@
1−import Disorder.Core.Main
1+import Control.Monad
2+
3+import qualified Test.Grenade.Network
2
4
3import qualified Test.Grenade.Layers.Pooling
5import qualified Test.Grenade.Layers.Pooling
4import qualified Test.Grenade.Layers.Convolution
6import qualified Test.Grenade.Layers.Convolution
5import qualified Test.Grenade.Layers.FullyConnected
7import qualified Test.Grenade.Layers.FullyConnected
8+import qualified Test.Grenade.Layers.Nonlinear
9+import qualified Test.Grenade.Layers.PadCrop
6
10
7import qualified Test.Grenade.Layers.Internal.Convolution
11import qualified Test.Grenade.Layers.Internal.Convolution
8import qualified Test.Grenade.Layers.Internal.Pooling
12import qualified Test.Grenade.Layers.Internal.Pooling
9
13
10import qualified Test.Grenade.Recurrent.Layers.LSTM
14import qualified Test.Grenade.Recurrent.Layers.LSTM
11
15
16+import System.Exit
17+import System.IO
18+
12main :: IO ()
19main :: IO ()
13main =
20main =
14 disorderMain [
21 disorderMain [
22+ Test.Grenade.Network.tests
23+
15− Test.Grenade.Layers.Pooling.tests
24+ , Test.Grenade.Layers.Pooling.tests
16 , Test.Grenade.Layers.Convolution.tests
25 , Test.Grenade.Layers.Convolution.tests
17 , Test.Grenade.Layers.FullyConnected.tests
26 , Test.Grenade.Layers.FullyConnected.tests
27+ , Test.Grenade.Layers.Nonlinear.tests
28+ , Test.Grenade.Layers.PadCrop.tests
18
29
19 , Test.Grenade.Layers.Internal.Convolution.tests
30 , Test.Grenade.Layers.Internal.Convolution.tests
20 , Test.Grenade.Layers.Internal.Pooling.tests
31 , Test.Grenade.Layers.Internal.Pooling.tests
21
32
22 , Test.Grenade.Recurrent.Layers.LSTM.tests
33 , Test.Grenade.Recurrent.Layers.LSTM.tests
34+
23 ]
35 ]
36+
37+disorderMain :: [IO Bool] -> IO ()
38+disorderMain tests = do
39+ lineBuffer
40+ rs <- sequence tests
41+ unless (and rs) exitFailure
42+
43+
44+lineBuffer :: IO ()
45+lineBuffer = do
46+ hSetBuffering stdout LineBuffering
47+ hSetBuffering stderr LineBuffering