marian-nmt/marian
1
1
Fork 0
mirror of https://github.com/marian-nmt/marian.git synced 2024-11-03 20:13:47 +03:00
Code Issues Packages Projects Releases Wiki Activity

function keywords

Browse Source
This commit is contained in:
Marcin Junczys-Dowmunt 2016-05-08 18:42:19 +02:00
parent 29ef0fee9e
commit 3db0f30312
5 changed files with 300 additions and 54 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
CMakeLists.txt
View File

@ -2,8 +2,8 @@ cmake_minimum_required(VERSION 3.5.1)
set(CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake) set(CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake)
project(marian CXX) project(marian CXX)
SET(CMAKE_CXX_FLAGS " -std=c++11 -g -O0 -funroll-loops -Wno-unused-result -Wno-deprecated") SET(CMAKE_CXX_FLAGS " -std=c++11 -g -O3 -funroll-loops -Wno-unused-result -Wno-deprecated")
LIST(APPEND CUDA_NVCC_FLAGS --default-stream per-thread; -std=c++11; -g; -O0; -arch=sm_35; -lineinfo; --use_fast_math;) LIST(APPEND CUDA_NVCC_FLAGS --default-stream per-thread; -std=c++11; -g; -O3; -arch=sm_35; -lineinfo; --use_fast_math;)
add_definitions(-DCUDA_API_PER_THREAD_DEFAULT_STREAM) add_definitions(-DCUDA_API_PER_THREAD_DEFAULT_STREAM)
SET(CUDA_PROPAGATE_HOST_FLAGS OFF) SET(CUDA_PROPAGATE_HOST_FLAGS OFF)

38
scripts/xor.py Normal file
View File

@ -0,0 +1,38 @@
import numpy as np
import keras
from keras.layers import Input, Dense
from keras.models import Model
from keras.optimizers import SGD
import time
inputs = Input(shape=(2,))
x = Dense(5000, activation='sigmoid')(inputs)
x = Dense(5000, activation='sigmoid')(x)
x = Dense(5000, activation='sigmoid')(x)
predictions = Dense(1, activation='sigmoid')(x)
X = np.array([
0, 0,
0, 1,
1, 0,
1, 1]).reshape((4,2))
Y = np.array([0, 1, 1, 0]).reshape((4,1))
#sgd = SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=False)
model = Model(input=inputs, output=predictions)
model.compile(optimizer='adadelta',
loss='binary_crossentropy',
metrics=['accuracy'])
start = time.time()
for i in range(10):
model.fit(X, Y, nb_epoch=200, verbose=0)
print model.predict(X)
print model.evaluate(X, Y, verbose=0)
end = time.time()
print(end - start)

64
src/compile_time_crc32.h Normal file
View File

@ -0,0 +1,64 @@
#pragma once
static constexpr uint32_t crc_table[256] = {
0x00000000, 0x77073096, 0xee0e612c, 0x990951ba, 0x076dc419, 0x706af48f,
0xe963a535, 0x9e6495a3, 0x0edb8832, 0x79dcb8a4, 0xe0d5e91e, 0x97d2d988,
0x09b64c2b, 0x7eb17cbd, 0xe7b82d07, 0x90bf1d91, 0x1db71064, 0x6ab020f2,
0xf3b97148, 0x84be41de, 0x1adad47d, 0x6ddde4eb, 0xf4d4b551, 0x83d385c7,
0x136c9856, 0x646ba8c0, 0xfd62f97a, 0x8a65c9ec, 0x14015c4f, 0x63066cd9,
0xfa0f3d63, 0x8d080df5, 0x3b6e20c8, 0x4c69105e, 0xd56041e4, 0xa2677172,
0x3c03e4d1, 0x4b04d447, 0xd20d85fd, 0xa50ab56b, 0x35b5a8fa, 0x42b2986c,
0xdbbbc9d6, 0xacbcf940, 0x32d86ce3, 0x45df5c75, 0xdcd60dcf, 0xabd13d59,
0x26d930ac, 0x51de003a, 0xc8d75180, 0xbfd06116, 0x21b4f4b5, 0x56b3c423,
0xcfba9599, 0xb8bda50f, 0x2802b89e, 0x5f058808, 0xc60cd9b2, 0xb10be924,
0x2f6f7c87, 0x58684c11, 0xc1611dab, 0xb6662d3d, 0x76dc4190, 0x01db7106,
0x98d220bc, 0xefd5102a, 0x71b18589, 0x06b6b51f, 0x9fbfe4a5, 0xe8b8d433,
0x7807c9a2, 0x0f00f934, 0x9609a88e, 0xe10e9818, 0x7f6a0dbb, 0x086d3d2d,
0x91646c97, 0xe6635c01, 0x6b6b51f4, 0x1c6c6162, 0x856530d8, 0xf262004e,
0x6c0695ed, 0x1b01a57b, 0x8208f4c1, 0xf50fc457, 0x65b0d9c6, 0x12b7e950,
0x8bbeb8ea, 0xfcb9887c, 0x62dd1ddf, 0x15da2d49, 0x8cd37cf3, 0xfbd44c65,
0x4db26158, 0x3ab551ce, 0xa3bc0074, 0xd4bb30e2, 0x4adfa541, 0x3dd895d7,
0xa4d1c46d, 0xd3d6f4fb, 0x4369e96a, 0x346ed9fc, 0xad678846, 0xda60b8d0,
0x44042d73, 0x33031de5, 0xaa0a4c5f, 0xdd0d7cc9, 0x5005713c, 0x270241aa,
0xbe0b1010, 0xc90c2086, 0x5768b525, 0x206f85b3, 0xb966d409, 0xce61e49f,
0x5edef90e, 0x29d9c998, 0xb0d09822, 0xc7d7a8b4, 0x59b33d17, 0x2eb40d81,
0xb7bd5c3b, 0xc0ba6cad, 0xedb88320, 0x9abfb3b6, 0x03b6e20c, 0x74b1d29a,
0xead54739, 0x9dd277af, 0x04db2615, 0x73dc1683, 0xe3630b12, 0x94643b84,
0x0d6d6a3e, 0x7a6a5aa8, 0xe40ecf0b, 0x9309ff9d, 0x0a00ae27, 0x7d079eb1,
0xf00f9344, 0x8708a3d2, 0x1e01f268, 0x6906c2fe, 0xf762575d, 0x806567cb,
0x196c3671, 0x6e6b06e7, 0xfed41b76, 0x89d32be0, 0x10da7a5a, 0x67dd4acc,
0xf9b9df6f, 0x8ebeeff9, 0x17b7be43, 0x60b08ed5, 0xd6d6a3e8, 0xa1d1937e,
0x38d8c2c4, 0x4fdff252, 0xd1bb67f1, 0xa6bc5767, 0x3fb506dd, 0x48b2364b,
0xd80d2bda, 0xaf0a1b4c, 0x36034af6, 0x41047a60, 0xdf60efc3, 0xa867df55,
0x316e8eef, 0x4669be79, 0xcb61b38c, 0xbc66831a, 0x256fd2a0, 0x5268e236,
0xcc0c7795, 0xbb0b4703, 0x220216b9, 0x5505262f, 0xc5ba3bbe, 0xb2bd0b28,
0x2bb45a92, 0x5cb36a04, 0xc2d7ffa7, 0xb5d0cf31, 0x2cd99e8b, 0x5bdeae1d,
0x9b64c2b0, 0xec63f226, 0x756aa39c, 0x026d930a, 0x9c0906a9, 0xeb0e363f,
0x72076785, 0x05005713, 0x95bf4a82, 0xe2b87a14, 0x7bb12bae, 0x0cb61b38,
0x92d28e9b, 0xe5d5be0d, 0x7cdcefb7, 0x0bdbdf21, 0x86d3d2d4, 0xf1d4e242,
0x68ddb3f8, 0x1fda836e, 0x81be16cd, 0xf6b9265b, 0x6fb077e1, 0x18b74777,
0x88085ae6, 0xff0f6a70, 0x66063bca, 0x11010b5c, 0x8f659eff, 0xf862ae69,
0x616bffd3, 0x166ccf45, 0xa00ae278, 0xd70dd2ee, 0x4e048354, 0x3903b3c2,
0xa7672661, 0xd06016f7, 0x4969474d, 0x3e6e77db, 0xaed16a4a, 0xd9d65adc,
0x40df0b66, 0x37d83bf0, 0xa9bcae53, 0xdebb9ec5, 0x47b2cf7f, 0x30b5ffe9,
0xbdbdf21c, 0xcabac28a, 0x53b39330, 0x24b4a3a6, 0xbad03605, 0xcdd70693,
0x54de5729, 0x23d967bf, 0xb3667a2e, 0xc4614ab8, 0x5d681b02, 0x2a6f2b94,
0xb40bbe37, 0xc30c8ea1, 0x5a05df1b, 0x2d02ef8d
};
template<size_t idx>
constexpr uint32_t crc32(const char * str)
{
return (crc32<idx-1>(str) >> 8) ^ crc_table[(crc32<idx-1>(str) ^ str[idx]) & 0x000000FF];
}
// This is the stop-recursion function
template<>
constexpr uint32_t crc32<size_t(-1)>(const char * str)
{
return 0xFFFFFFFF;
}
// This don't take into account the nul char
#define COMPILE_TIME_CRC32_STR(x) (crc32<sizeof(x) - 2>(x) ^ 0xFFFFFFFF)

103
src/keywords.h Normal file
View File

@ -0,0 +1,103 @@
#pragma once
#include <iostream>
#include <vector>
#include <typeinfo>
#include <typeindex>
#include <unordered_map>
#include <boost/any.hpp>
#include "compile_time_crc32.h"
namespace marian {
namespace keywords {
template <int key, typename Value>
class Keyword {
public:
typedef Value value_type;
struct pair {
Keyword<key, Value> first;
Value second;
};
Keyword(const std::string& name)
: name_(name) {}
pair operator=(Value value) {
return pair{*this, value};
}
const std::string& operator()() const {
return name_;
}
private:
std::string name_;
};
struct Keywords {
Keywords() {}
template <typename ...Args>
Keywords(Args ...args) {
add(args...);
}
template <typename Head>
void add(Head head) {
map_[std::type_index(typeid(head.first))] = head.second;
}
template <typename Head, typename ...Tail>
void add(Head head, Tail ...tail) {
map_[std::type_index(typeid(head.first))] = head.second;
add(tail...);
}
template <typename Value, typename Key>
Value Get(Key key, Value default_value) {
auto it = map_.find(std::type_index(typeid(key)));
if(it != map_.end())
return boost::any_cast<Value>(map_[std::type_index(typeid(key))]);
else
return default_value;
}
private:
std::unordered_map<std::type_index, boost::any> map_;
};
#define KEY(name, value_type) \
typedef Keyword<COMPILE_TIME_CRC32_STR(#name),value_type> name ## _k; \
name ## _k name(#name);
KEY(shape, std::vector<int>)
KEY(prefix, std::string)
KEY(axis, size_t);
}
class demo : public keywords::Keywords {
public:
template <typename ...Args>
demo(size_t size, Args ...args)
: Keywords(args...),
size_(size),
prefix_(Get<std::string>(keywords::prefix, std::string("_"))),
shape_(Get<std::vector<int>>(keywords::shape, std::vector<int>()))
{}
private:
size_t size_;
std::string prefix_;
std::vector<int> shape_;
};
void demo_main() {
using namespace keywords;
demo(300, shape={1,3}, prefix="layer1_", axis=0);
}
}

141
src/test.cu
View File

@ -4,68 +4,109 @@
#include <algorithm> #include <algorithm>
#include <random> #include <random>
#include <boost/timer/timer.hpp> #include <boost/timer/timer.hpp>
#include <typeinfo>
#include <typeindex>
#include <unordered_map>
#include <boost/any.hpp>
#include "marian.h" #include "marian.h"
#include "operators.h" #include "operators.h"
#include "keywords.h"
using namespace marian; using namespace marian;
int main(int argc, char** argv) { int main(int argc, char** argv) {
boost::timer::auto_cpu_timer t;
Var x = input("X", Tensor({4, 2})); using namespace keywords;
Var y = input("Y", Tensor({4, 2}));
std::vector<float> vx = { auto layer = demo(300, prefix="test_");
0, 0,
0, 1,
1, 0,
1, 1
};
std::vector<float> vy = { //auto x = input("X", shape={1, 768});
1, 0, //auto y = input("Y", shape={1, 10});
1, 0, //
0, 1, //auto l = x;
1, 0 //for(auto n : { 300, 200, 100, 50, 20 })
}; // l = dense(n, l, activation=tanh);
//
//auto w = param("W", init=orthogonal, shape={20, 10});
//auto b = param("b", init=orthogonal, shape={1, 10});
//l = sigmoid(dot(w, l) + b);
//
//auto lp = dense(10, l, activation=softmax(axis=1));
//auto cost = -mean(sum(y * log(lp), axis=1));
thrust::copy(vx.begin(), vx.end(), x.val().begin());
thrust::copy(vy.begin(), vy.end(), y.val().begin());
Var w0 = forsave("W0", uniform(Tensor({2, 2}))); //auto x1 = input(k::name="x0", k::shape={1,100});
Var b0 = forsave("b0", uniform(Tensor({1, 2}))); //auto x2 = input(k::name="x1", k::shape={1,100});
//auto y = output(k::name="y", k::shape={1,10});
Var w1 = forsave("W1", uniform(Tensor({2, 2}))); //
Var b1 = forsave("b1", uniform(Tensor({1, 2}))); //auto l1 = dense(100,
// k::name="layer1",
std::vector<Var> params = { w0, w1, b0, b1 }; // k::input={x1, x2},
// k::activation=sigmoid,
Var ry = sigma(dot(x, w0) + b0); // k::init_w=orthogonal,
ry = softmax(dot(ry, w1) + b1, Axis::axis1); // k::init_b=uniform(-0.1,0.1)
Var cost = -mean(sum(y * log(ry), Axis::axis1), Axis::axis0); // k::merge=concat);
//auto l2 = dense(100, k::input=l1, k::name="charlie"
float alpha = 0.1; // k::activation=tanh);
for(size_t i = 0; i < 30000; ++i) { //auto lout = dense(10, k::input=l2,
cost.forward(); // k::activation=softmax);
//
if(i % 100 == 0) { //auto cost = -mean(sum(y * log(lout), k::axis=1));
for(size_t j = 0; j < 4; ++j) { //
std::cerr << ry.val()[j*2] << std::endl; //auto w = cost["charlie_w"];
} //auto b = cost["layer1_b"];
std::cerr << i << " ct: " << cost.val()[0] << std::endl; //
// alpha = alpha * 0.9; //auto opt = optimizer(cost,
} // k::method=adadelta);
//
cost.backward(); //Tensor X(k::shape={60, 768}, k::init=mnist(""));
for(auto p : params) { //Tensor Y(k::shape={60, 10}, k::init=mnist(""));
//std::cerr << p.grad()[0] << std::endl; //
auto update = //float c = opt.fit_batch({X1, X2}, Y, k::logger=logger);
_1 -= alpha * _2; //
//Tensor xTrain
Element(update, p.val(), p.grad()); // (shape, {60000, 784})
} // (init, mnist("train.ubyte"));
} //
//Tensor yTrain
// (shape, {60000, 10})
// (init, mnist("train.ubyte", true));
//
//Tensor xBatch = slice(xTrain, {0, 50, 5});
//
//Var x = input("X");
//Var y = input("Y");
//
//ry = dense(input=x, size=200, activation=tanh,
// init_w=orthogonal, init_b=uniform(-0.1. 0.1));
//
//ry = dense(ry)(size, 100)(activation, tanh);
//ry = dense(ry)(size, 10)(activation, softmax);
//
//Var cost = -mean(y * log(ry) + (1 - y) * log(1 - ry));
//
//boost::timer::auto_cpu_timer t;
//float eta = 0.01;
//for(size_t i = 0; i < 2000; ++i) {
// cost.forward();
//
// if(i % 200 == 0) {
// for(size_t j = 0; j < 4; ++j) {
// std::cerr << ry.val()[j] << std::endl;
// }
// std::cerr << i << " ct: " << cost.val()[0] << std::endl;
// }
//
// cost.backward();
// for(auto p : params) {
// auto update =
// _1 -= eta * _2;
// Element(update, p.val(), p.grad());
// }
//}
return 0; return 0;
} }