marian-nmt/marian
1
1
Fork 0
mirror of https://github.com/marian-nmt/marian.git synced 2024-11-04 14:04:24 +03:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' of github.com:emjotde/Marian

Browse Source
This commit is contained in:
Hieu Hoang 2016-09-14 19:58:22 +02:00
commit adbc97f448
4 changed files with 182 additions and 179 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

103
src/tensor.h
View File

@ -1,6 +1,5 @@
#pragma once
#include <cudnn.h>
#include <cublas_v2.h>
#include <thrust/device_vector.h>
#include <thrust/functional.h>
@ -13,27 +12,27 @@
namespace marian {
struct Handles {
cudnnHandle_t cudnnHandle;
cublasHandle_t cublasHandle;
cudnnOpTensorDescriptor_t add;
Handles() {
cudnnCreate(&cudnnHandle);
cublasCreate(&cublasHandle);
cudnnCreateOpTensorDescriptor(&add);
cudnnSetOpTensorDescriptor(add, CUDNN_OP_TENSOR_ADD, CUDNN_DATA_FLOAT, CUDNN_NOT_PROPAGATE_NAN);
}
~Handles() {
cudnnDestroy(cudnnHandle);
cublasDestroy(cublasHandle);
cudnnDestroyOpTensorDescriptor(add);
}
};
const Handles handles;
//struct Handles {
// //cudnnHandle_t cudnnHandle;
// //cublasHandle_t cublasHandle;
//
// //cudnnOpTensorDescriptor_t add;
//
// Handles() {
// cudnnCreate(&cudnnHandle);
// cublasCreate(&cublasHandle);
// cudnnCreateOpTensorDescriptor(&add);
// cudnnSetOpTensorDescriptor(add, CUDNN_OP_TENSOR_ADD, CUDNN_DATA_FLOAT, CUDNN_NOT_PROPAGATE_NAN);
// }
//
// ~Handles() {
// cudnnDestroy(cudnnHandle);
// cublasDestroy(cublasHandle);
// cudnnDestroyOpTensorDescriptor(add);
// }
//};
//
//const Handles handles;
// typedef std::vector<int> Shape;
@ -60,17 +59,17 @@ class TensorImpl {
private:
Shape shape_;
thrust::device_vector<Float> data_;
cudnnTensorDescriptor_t desc_;
//cudnnTensorDescriptor_t desc_;
size_t tno_;
static size_t tensorCounter;
cudnnDataType_t dataType() {
switch(sizeof(Float)) {
case 2: return CUDNN_DATA_HALF;
case 8: return CUDNN_DATA_DOUBLE;
default: return CUDNN_DATA_FLOAT;
}
}
//cudnnDataType_t dataType() {
// switch(sizeof(Float)) {
// case 2: return CUDNN_DATA_HALF;
// case 8: return CUDNN_DATA_DOUBLE;
// default: return CUDNN_DATA_FLOAT;
// }
//}
public:
typedef Float value_type;
@ -90,28 +89,28 @@ class TensorImpl {
int size = GetTotalSize(shape_);
data_.resize(size, value);
cudnnCreateTensorDescriptor(&desc_);
switch (shape_.size()) {
case 1:
cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
shape_[0], 1, 1, 1); break;
case 2:
cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
shape_[0], shape_[1], 1, 1); break;
case 3:
cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
shape_[0], shape_[1], shape_[2], 1); break;
case 4:
cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
shape_[0], shape_[1], shape_[2], shape_[3]); break;
}
//cudnnCreateTensorDescriptor(&desc_);
//switch (shape_.size()) {
// case 1:
// cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
// shape_[0], 1, 1, 1); break;
// case 2:
// cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
// shape_[0], shape_[1], 1, 1); break;
// case 3:
// cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
// shape_[0], shape_[1], shape_[2], 1); break;
// case 4:
// cudnnSetTensor4dDescriptor(desc_, CUDNN_TENSOR_NCHW, dataType(),
// shape_[0], shape_[1], shape_[2], shape_[3]); break;
//}
}
TensorImpl(const TensorImpl&) = delete;
TensorImpl(TensorImpl&&) = delete;
~TensorImpl() {
cudnnDestroyTensorDescriptor(desc_);
//cudnnDestroyTensorDescriptor(desc_);
}
value_type operator[](size_t i) const {
@ -146,9 +145,9 @@ class TensorImpl {
return thrust::raw_pointer_cast(data_.data());
}
cudnnTensorDescriptor_t desc() const {
return desc_;
}
//cudnnTensorDescriptor_t desc() const {
// return desc_;
//}
size_t id() const {
return tno_;
@ -246,9 +245,9 @@ class Tensor {
return pimpl_->shape();
}
cudnnTensorDescriptor_t desc() const {
return pimpl_->desc();
}
//cudnnTensorDescriptor_t desc() const {
// return pimpl_->desc();
//}
void set(value_type value) {
pimpl_->set(value);

6
src/tensor_operators.cu
View File

@ -130,7 +130,11 @@ Tensor Prod(cublasHandle_t handle, Tensor C, const Tensor A, const Tensor B,
Tensor Prod(Tensor C, const Tensor A, const Tensor B,
bool transA, bool transB, Float beta) {
return Prod(handles.cublasHandle, C, A, B, transA, transB, beta);
cublasHandle_t cublasHandle;
cublasCreate(&cublasHandle);
Tensor temp = Prod(cublasHandle, C, A, B, transA, transB, beta);
cublasDestroy(cublasHandle);
return temp;
}
}

208
src/test.cu
View File

@ -2,9 +2,9 @@
#include "marian.h"
#include "mnist.h"
using namespace std;
int main(int argc, char** argv) {
cudaSetDevice(0);
/*int numImg = 0;*/
/*auto images = datasets::mnist::ReadImages("../examples/mnist/t10k-images-idx3-ubyte", numImg);*/
/*auto labels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", numImg);*/
@ -12,118 +12,104 @@ int main(int argc, char** argv) {
using namespace marian;
using namespace keywords;
const size_t BATCH_SIZE = 500;
const size_t IMAGE_SIZE = 784;
const size_t LABEL_SIZE = 10;
Expr x = input(shape={whatevs, IMAGE_SIZE}, name="X");
Expr y = input(shape={whatevs, LABEL_SIZE}, name="Y");
Expr x = input(shape={1, 2});
Expr y = input(shape={1, 2});
Expr w = param(shape={IMAGE_SIZE, LABEL_SIZE}, name="W0");
// Expr w = param(shape={IMAGE_SIZE, LABEL_SIZE}, name="W0", init=randreal);
Expr b = param(shape={1, LABEL_SIZE}, name="b0");
Expr z = dot(x, w) + b;
Expr lr = softmax(z, axis=1, name="pred");
Expr graph = -mean(sum(y * log(lr), axis=1), axis=0, name="cost");
//cerr << "x=" << Debug(lr.val().shape()) << endl;
Expr w = param(shape={2, 2}, name="W0");
//Expr b = param(shape={1, 2}, name="b0");
int numofdata;
//vector<float> images = datasets::mnist::ReadImages("../examples/mnist/t10k-images-idx3-ubyte", numofdata, IMAGE_SIZE);
//vector<float> labels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", numofdata, LABEL_SIZE);
vector<float> images = datasets::mnist::ReadImages("../examples/mnist/train-images-idx3-ubyte", numofdata, IMAGE_SIZE);
vector<float> labels = datasets::mnist::ReadLabels("../examples/mnist/train-labels-idx1-ubyte", numofdata, LABEL_SIZE);
cerr << "images=" << images.size() << " labels=" << labels.size() << endl;
cerr << "numofdata=" << numofdata << endl;
size_t startInd = 0;
size_t startIndData = 0;
while (startInd < numofdata) {
size_t batchSize = (startInd + BATCH_SIZE < numofdata) ? BATCH_SIZE : numofdata - startInd;
cerr << "startInd=" << startInd
<< " startIndData=" << startIndData
<< " batchSize=" << batchSize << endl;
Tensor tx({numofdata, IMAGE_SIZE}, 1);
Tensor ty({numofdata, LABEL_SIZE}, 1);
tx.Load(images.begin() + startIndData, images.begin() + startIndData + batchSize * IMAGE_SIZE);
ty.Load(labels.begin() + startInd, labels.begin() + startInd + batchSize);
//cerr << "tx=" << Debug(tx.shape()) << endl;
//cerr << "ty=" << Debug(ty.shape()) << endl;
x = tx;
y = ty;
cerr << "x=" << Debug(x.val().shape()) << endl;
cerr << "y=" << Debug(y.val().shape()) << endl;
graph.forward(batchSize);
cerr << "w=" << Debug(w.val().shape()) << endl;
cerr << "b=" << Debug(b.val().shape()) << endl;
std::cerr << "z: " << Debug(z.val().shape()) << endl;
std::cerr << "lr: " << Debug(lr.val().shape()) << endl;
std::cerr << "Log-likelihood: " << Debug(graph.val().shape()) << endl ;
//std::cerr << "scores=" << scores.val().Debug() << endl;
//std::cerr << "lr=" << lr.val().Debug() << endl;
graph.backward();
//std::cerr << graph["pred"].val()[0] << std::endl;
startInd += batchSize;
startIndData += batchSize * IMAGE_SIZE;
}
std::cerr << "Building model...";
auto predict = softmax(dot(x, w),
axis=1, name="pred");
auto graph = -mean(sum(y * log(predict), axis=1),
axis=0, name="cost");
// XOR
/*
Expr x = input(shape={whatevs, 2}, name="X");
Expr y = input(shape={whatevs, 2}, name="Y");
Expr w = param(shape={2, 1}, name="W0");
Expr b = param(shape={1, 1}, name="b0");
Expr n5 = dot(x, w);
Expr n6 = n5 + b;
Expr lr = softmax(n6, axis=1, name="pred");
cerr << "lr=" << lr.Debug() << endl;
Expr graph = -mean(sum(y * log(lr), axis=1), axis=0, name="cost");
Tensor tx({4, 2}, 1);
Tensor ty({4, 1}, 1);
cerr << "tx=" << tx.Debug() << endl;
cerr << "ty=" << ty.Debug() << endl;
tx.Load("../examples/xor/train.txt");
ty.Load("../examples/xor/label.txt");
*/
#if 0
hook0(graph);
graph.autodiff();
std::cerr << graph["cost"].val()[0] << std::endl;
//hook1(graph);
for(auto p : graph.params()) {
auto update = _1 = _1 - alpha * _2;
Element(update, p.val(), p.grad());
}
hook2(graph);
Tensor x1t({1, 2});
std::vector<float> xv = { 0.6, 0.1 };
thrust::copy(xv.begin(), xv.end(), x1t.begin());
Tensor x2t({1, 2});
std::vector<float> yv = { 0, 1 };
thrust::copy(yv.begin(), yv.end(), x2t.begin());
x = x1t;
y = x2t;
graph.forward(1);
graph.backward();
std::cerr << graph.val().Debug() << std::endl;
std::cerr << w.grad().Debug() << std::endl;
//std::cerr << b.grad().Debug() << std::endl;
// using namespace marian;
// using namespace keywords;
//
// const size_t BATCH_SIZE = 500;
// const size_t IMAGE_SIZE = 784;
// const size_t LABEL_SIZE = 10;
//
// Expr x = input(shape={whatevs, IMAGE_SIZE}, name="X");
// Expr y = input(shape={whatevs, LABEL_SIZE}, name="Y");
//
// Expr w = param(shape={IMAGE_SIZE, LABEL_SIZE}, name="W0");
// Expr b = param(shape={1, LABEL_SIZE}, name="b0");
//
// Expr z = dot(x, w) + b;
// Expr lr = softmax(z, axis=1, name="pred");
// Expr graph = -mean(sum(y * log(lr), axis=1), axis=0, name="cost");
// //cerr << "x=" << Debug(lr.val().shape()) << endl;
//
// int numofdata;
// //vector<float> images = datasets::mnist::ReadImages("../examples/mnist/t10k-images-idx3-ubyte", numofdata, IMAGE_SIZE);
// //vector<float> labels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", numofdata, LABEL_SIZE);
// vector<float> images = datasets::mnist::ReadImages("../examples/mnist/train-images-idx3-ubyte", numofdata, IMAGE_SIZE);
// vector<float> labels = datasets::mnist::ReadLabels("../examples/mnist/train-labels-idx1-ubyte", numofdata, LABEL_SIZE);
// cerr << "images=" << images.size() << " labels=" << labels.size() << endl;
// cerr << "numofdata=" << numofdata << endl;
//
// size_t startInd = 0;
// size_t startIndData = 0;
// while (startInd < numofdata) {
// size_t batchSize = (startInd + BATCH_SIZE < numofdata) ? BATCH_SIZE : numofdata - startInd;
// cerr << "startInd=" << startInd
// << " startIndData=" << startIndData
// << " batchSize=" << batchSize << endl;
//
// Tensor tx({numofdata, IMAGE_SIZE}, 1);
// Tensor ty({numofdata, LABEL_SIZE}, 1);
//
// tx.Load(images.begin() + startIndData, images.begin() + startIndData + batchSize * IMAGE_SIZE);
// ty.Load(labels.begin() + startInd, labels.begin() + startInd + batchSize);
//
// //cerr << "tx=" << Debug(tx.shape()) << endl;
// //cerr << "ty=" << Debug(ty.shape()) << endl;
//
// x = tx;
// y = ty;
//
// cerr << "x=" << Debug(x.val().shape()) << endl;
// cerr << "y=" << Debug(y.val().shape()) << endl;
//
//
// graph.forward(batchSize);
//
// cerr << "w=" << Debug(w.val().shape()) << endl;
// cerr << "b=" << Debug(b.val().shape()) << endl;
// std::cerr << "z: " << Debug(z.val().shape()) << endl;
// std::cerr << "lr: " << Debug(lr.val().shape()) << endl;
// std::cerr << "Log-likelihood: " << Debug(graph.val().shape()) << endl ;
//
// //std::cerr << "scores=" << scores.val().Debug() << endl;
// //std::cerr << "lr=" << lr.val().Debug() << endl;
//
// //graph.backward();
//
// //std::cerr << graph["pred"].val()[0] << std::endl;
//
// startInd += batchSize;
// startIndData += batchSize * IMAGE_SIZE;
// }
auto opt = adadelta(cost_function=cost,
eta=0.9, gamma=0.1,
set_batch=set,
before_update=before,
after_update=after,
set_valid=valid,
validation_freq=100,
verbose=1, epochs=3, early_stopping=10);
opt.run();
#endif
return 0;
}

44
src/validate_mnist.cu
View File

@ -7,13 +7,16 @@ using namespace marian;
using namespace keywords;
int main(int argc, char** argv) {
cudaSetDevice(0);
const size_t IMAGE_SIZE = 784;
const size_t LABEL_SIZE = 10;
int numofdata;
std::cerr << "Loading test set...";
std::vector<float> testImages = datasets::mnist::ReadImages("../examples/mnist/t10k-images-idx3-ubyte", numofdata, IMAGE_SIZE);
std::vector<float>testLabels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", numofdata, LABEL_SIZE);
std::vector<float> testLabels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", numofdata, LABEL_SIZE);
std::cerr << "\tDone." << std::endl;
std::cerr << "Loading model params...";
@ -27,11 +30,11 @@ int main(int argc, char** argv) {
Shape bShape;
converter.Load("bias", bData, bShape);
auto initW = [&wData](Tensor t) {
auto initW = [wData](Tensor t) {
thrust::copy(wData.begin(), wData.end(), t.begin());
};
auto initB = [&bData](Tensor t) {
auto initB = [bData](Tensor t) {
thrust::copy(bData.begin(), bData.end(), t.begin());
};
@ -39,24 +42,35 @@ int main(int argc, char** argv) {
Expr x = input(shape={whatevs, IMAGE_SIZE}, name="X");
Expr y = input(shape={whatevs, LABEL_SIZE}, name="Y");
Expr w = param(shape={IMAGE_SIZE, LABEL_SIZE}, name="W0", init=initW);
Expr b = param(shape={1, LABEL_SIZE}, name="b0", init=initB);
std::cerr << "Building model...";
auto scores = dot(x, w) + b;
auto predict = softmax(scores, axis=1, name="pred");
auto predict = softmax(dot(x, w) + b,
axis=1, name="pred");
auto graph = -mean(sum(y * log(predict), axis=1),
axis=0, name="cost");
std::cerr << "\tDone." << std::endl;
Tensor xt({numofdata, IMAGE_SIZE});
xt.Load(testImages);
predict.forward(numofdata);
Tensor yt({numofdata, LABEL_SIZE});
yt.Load(testLabels);
x = xt;
y = yt;
graph.forward(numofdata);
auto results = predict.val();
graph.backward();
std::cerr << b.grad().Debug() << std::endl;
size_t acc = 0;
for (size_t i = 0; i < testLabels.size(); i += LABEL_SIZE) {
size_t correct = 0;
size_t predicted = 0;
@ -65,11 +79,11 @@ int main(int argc, char** argv) {
if (results[i + j] > results[i + predicted]) predicted = j;
}
acc += (correct == predicted);
std::cerr << "corect: " << correct << " | " << predicted << "(";
for (size_t j = 0; j < LABEL_SIZE; ++j) {
std::cerr << results[i+j] << " ";
}
std::cerr << std::endl;
//std::cerr << "corect: " << correct << " | " << predicted << "(";
//for (size_t j = 0; j < LABEL_SIZE; ++j) {
// std::cerr << results[i+j] << " ";
//}
//std::cerr << std::endl;
}
std::cerr << "ACC: " << float(acc)/numofdata << std::endl;