mirror of
https://github.com/marian-nmt/marian.git
synced 2024-09-11 06:15:56 +03:00
Merge branch 'master' of https://github.com/emjotde/Marian
This commit is contained in:
commit
6534403830
@ -5,7 +5,6 @@ cuda_add_library(marian_lib
|
||||
cnpy/cnpy.cpp
|
||||
exception.cpp
|
||||
expression_graph.cu
|
||||
sgd.cu
|
||||
tensor.cu
|
||||
tensor_operators.cu
|
||||
expression_operators.cu
|
||||
|
@ -18,7 +18,7 @@ struct Chainable {
|
||||
|
||||
virtual void allocate(size_t) = 0;
|
||||
virtual std::string graphviz() = 0;
|
||||
|
||||
virtual const std::string &name() const = 0;
|
||||
|
||||
virtual const Shape& shape() = 0;
|
||||
virtual DataType &val() = 0;
|
||||
@ -33,4 +33,4 @@ typedef std::shared_ptr<ChainableStack> ChainableStackPtr;
|
||||
typedef std::shared_ptr<Chainable<Tensor>> ChainPtr;
|
||||
|
||||
|
||||
}
|
||||
}
|
||||
|
@ -39,12 +39,12 @@ std::string Expr::Debug() const
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////
|
||||
ExpressionGraph::ExpressionGraph(int cudaDevice)
|
||||
: stack_(new ChainableStack)
|
||||
{
|
||||
std::srand (time(NULL));
|
||||
cudaSetDevice(0);
|
||||
|
||||
}
|
||||
//ExpressionGraph::ExpressionGraph(int cudaDevice)
|
||||
//: stack_(new ChainableStack)
|
||||
//{
|
||||
// std::srand (time(NULL));
|
||||
// cudaSetDevice(0);
|
||||
//
|
||||
//}
|
||||
|
||||
}
|
||||
|
@ -38,9 +38,14 @@ class Expr {
|
||||
|
||||
class ExpressionGraph {
|
||||
public:
|
||||
ExpressionGraph(int cudaDevice);
|
||||
ExpressionGraph() : stack_(new ChainableStack) {}
|
||||
|
||||
void forward(size_t batchSize) {
|
||||
void backprop(int batchSize) {
|
||||
forward(batchSize);
|
||||
backward();
|
||||
}
|
||||
|
||||
void forward(int batchSize) {
|
||||
for(auto&& v : *stack_) {
|
||||
v->allocate(batchSize);
|
||||
}
|
||||
@ -48,18 +53,6 @@ class ExpressionGraph {
|
||||
v->forward();
|
||||
}
|
||||
|
||||
std::string graphviz() {
|
||||
std::stringstream ss;
|
||||
ss << "digraph ExpressionGraph {" << std::endl;
|
||||
ss << "rankdir=BT" << std::endl;
|
||||
|
||||
typedef typename ChainableStack::reverse_iterator It;
|
||||
for(It it = stack_->rbegin(); it != stack_->rend(); ++it)
|
||||
ss << (*it)->graphviz();
|
||||
ss << "}" << std::endl;
|
||||
return ss.str();
|
||||
}
|
||||
|
||||
void backward() {
|
||||
for(auto&& v : *stack_)
|
||||
v->set_zero_adjoint();
|
||||
@ -70,9 +63,25 @@ class ExpressionGraph {
|
||||
(*it)->backward();
|
||||
}
|
||||
|
||||
std::string graphviz() {
|
||||
std::stringstream ss;
|
||||
ss << "digraph ExpressionGraph {" << std::endl;
|
||||
ss << "rankdir=BT" << std::endl;
|
||||
typedef typename ChainableStack::reverse_iterator It;
|
||||
for(It it = stack_->rbegin(); it != stack_->rend(); ++it) {
|
||||
ss << (*it)->graphviz();
|
||||
}
|
||||
ss << "}" << std::endl;
|
||||
return ss.str();
|
||||
}
|
||||
|
||||
/*********************************************************/
|
||||
|
||||
template <typename ...Args>
|
||||
inline Expr input(Args ...args) {
|
||||
return Expr(this, new InputNode(args...));
|
||||
Expr e(this, new InputNode(args...));
|
||||
inputs_.emplace_back(e);
|
||||
return e;
|
||||
}
|
||||
|
||||
template <typename ...Args>
|
||||
@ -117,14 +126,20 @@ class ExpressionGraph {
|
||||
named_.emplace(name, e);
|
||||
}
|
||||
|
||||
std::vector<Expr>& inputs() {
|
||||
return inputs_;
|
||||
}
|
||||
|
||||
std::vector<Expr>& params() {
|
||||
return params_;
|
||||
}
|
||||
|
||||
private:
|
||||
ChainableStackPtr stack_;
|
||||
|
||||
std::map<std::string, Expr> named_;
|
||||
std::vector<Expr> params_;
|
||||
std::vector<Expr> inputs_;
|
||||
};
|
||||
|
||||
}
|
||||
|
@ -67,6 +67,8 @@ class Node : public Chainable<Tensor>,
|
||||
virtual const Shape& shape() {
|
||||
return shape_;
|
||||
}
|
||||
|
||||
const std::string &name() const { return name_; }
|
||||
|
||||
protected:
|
||||
Shape shape_;
|
||||
|
140
src/sgd.cu
140
src/sgd.cu
@ -1,140 +0,0 @@
|
||||
#include <ctime>
|
||||
#include <algorithm>
|
||||
#include <vector>
|
||||
#include "sgd.h"
|
||||
#include "thrust_functions.h"
|
||||
|
||||
using namespace std;
|
||||
|
||||
namespace marian {
|
||||
SGD::SGD(ExpressionGraph& g, float eta,
|
||||
std::vector<float>& xData, size_t numFeatures,
|
||||
std::vector<float>& yData, size_t numClasses,
|
||||
size_t epochs, size_t batchSize)
|
||||
: graph_(g),
|
||||
eta_(eta),
|
||||
xData_(xData),
|
||||
numFeatures_(numFeatures),
|
||||
yData_(yData),
|
||||
numClasses_(numClasses),
|
||||
epochs_(epochs),
|
||||
maxBatchSize_(batchSize)
|
||||
{}
|
||||
|
||||
void SGD::Run()
|
||||
{
|
||||
size_t numExamples = xData_.size()/ numFeatures_;
|
||||
Tensor xt({(int)maxBatchSize_, (int)numExamples}, 0.0f);
|
||||
Tensor yt({(int)maxBatchSize_, (int)numClasses_}, 0.0f);
|
||||
|
||||
vector<size_t> shuffle = CreateShuffle(numExamples);
|
||||
//vector<size_t> shuffle;
|
||||
|
||||
for (size_t numEpoch = 0; numEpoch < epochs_; ++numEpoch) {
|
||||
std::cerr << "Starting epoch #" << numEpoch << std::endl;
|
||||
size_t startId = 0;
|
||||
|
||||
while (startId < numExamples) {
|
||||
size_t batchSize = std::min(maxBatchSize_, numExamples - startId);
|
||||
size_t endId = startId + batchSize;
|
||||
|
||||
PrepareBatch(startId, endId, batchSize, shuffle, xt, yt);
|
||||
graph_["x"] = xt;
|
||||
graph_["y"] = yt;
|
||||
|
||||
graph_.forward(maxBatchSize_);
|
||||
graph_.backward();
|
||||
|
||||
UpdateModel();
|
||||
|
||||
startId += maxBatchSize_;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
std::vector<size_t> SGD::CreateShuffle(size_t numExamples) const {
|
||||
vector<size_t> ret(numExamples);
|
||||
std::iota(ret.begin(), ret.end(), 0);
|
||||
std::random_shuffle ( ret.begin(), ret.end() );
|
||||
/*
|
||||
cerr << "shuffled" << endl;
|
||||
for (size_t i = 0; i < ret.size(); ++i) {
|
||||
cerr << ret[i] << " ";
|
||||
}
|
||||
*/
|
||||
return ret;
|
||||
}
|
||||
|
||||
void SGD::PrepareBatch(
|
||||
size_t startId,
|
||||
size_t endId,
|
||||
size_t batchSize,
|
||||
const std::vector<size_t> &shuffle,
|
||||
Tensor& xt,
|
||||
Tensor& yt) {
|
||||
/*
|
||||
std::vector<float> x(xData_.begin() + startId * numFeatures_,
|
||||
xData_.begin() + endId * numFeatures_);
|
||||
std::vector<float> y(yData_.begin() + startId * numClasses_,
|
||||
yData_.begin() + endId * numClasses_);
|
||||
*/
|
||||
std::vector<float> x(batchSize * numFeatures_);
|
||||
std::vector<float> y(batchSize * numClasses_);
|
||||
|
||||
//cerr << "batchSize=" << batchSize << endl;
|
||||
/*
|
||||
cerr << "startId=" << startId
|
||||
<< " " << endId
|
||||
<< " " << batchSize
|
||||
<< endl;
|
||||
cerr << "numExamples=" << shuffle.size() << endl;
|
||||
cerr << "numFeatures_=" << numFeatures_ << " " << numClasses_ << endl;
|
||||
cerr << "sizes=" << x.size()
|
||||
<< " " << y.size()
|
||||
<< " " << xData_.size()
|
||||
<< " " << yData_.size()
|
||||
<< endl;
|
||||
*/
|
||||
size_t startXId = 0;
|
||||
size_t startYId = 0;
|
||||
|
||||
for (size_t i = startId; i < endId; ++i) {
|
||||
size_t ind = shuffle[i];
|
||||
size_t startXDataId = ind * numFeatures_;
|
||||
size_t startYDataId = ind * numClasses_;
|
||||
|
||||
size_t endXDataId = startXDataId + numFeatures_;
|
||||
size_t endYDataId = startYDataId + numClasses_;
|
||||
/*
|
||||
cerr << "i=" << i
|
||||
<< " " << ind
|
||||
<< " " << startXDataId << "-" << endXDataId
|
||||
<< " " << startYDataId << "-" << endYDataId
|
||||
<< endl;
|
||||
*/
|
||||
|
||||
std::copy(xData_.begin() + startXDataId,
|
||||
xData_.begin() + endXDataId,
|
||||
x.begin() + startXId);
|
||||
|
||||
std::copy(yData_.begin() + startYDataId,
|
||||
yData_.begin() + endYDataId,
|
||||
y.begin() + startYId);
|
||||
|
||||
startXId += numFeatures_;
|
||||
startYId += numClasses_;
|
||||
}
|
||||
|
||||
xt.set(x);
|
||||
yt.set(y);
|
||||
}
|
||||
|
||||
void SGD::UpdateModel() {
|
||||
for (auto& param : graph_.params()) {
|
||||
using namespace thrust::placeholders;
|
||||
Element(_1 -= eta_ * _2, param.val(), param.grad());
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace
|
||||
|
71
src/sgd.h
71
src/sgd.h
@ -1,43 +1,48 @@
|
||||
#pragma once
|
||||
|
||||
#include <memory>
|
||||
#include <iostream>
|
||||
|
||||
#include "expression_graph.h"
|
||||
#include "thrust_functions.h"
|
||||
#include <map>
|
||||
#include <boost/any.hpp>
|
||||
#include "tensor_operators.h"
|
||||
|
||||
namespace marian {
|
||||
|
||||
class SGD {
|
||||
class Sgd {
|
||||
public:
|
||||
SGD(ExpressionGraph& g, float eta,
|
||||
std::vector<float>& xData, size_t numFeatures,
|
||||
std::vector<float>& yData, size_t numClasses,
|
||||
size_t epochs, size_t batchSize);
|
||||
|
||||
void Run();
|
||||
|
||||
Sgd(float eta=0.1) : eta_(eta) {}
|
||||
|
||||
void operator()(ExpressionGraph& graph, int batchSize) {
|
||||
graph.backprop(batchSize);
|
||||
|
||||
for(auto& param : graph.params())
|
||||
Element(_1 -= eta_ * _2, param.val(), param.grad());
|
||||
}
|
||||
|
||||
private:
|
||||
ExpressionGraph& graph_;
|
||||
const float eta_;
|
||||
std::vector<float>& xData_;
|
||||
const size_t numFeatures_;
|
||||
std::vector<float>& yData_;
|
||||
const size_t numClasses_;
|
||||
const size_t epochs_;
|
||||
const size_t maxBatchSize_;
|
||||
|
||||
std::vector<size_t> CreateShuffle(size_t numExamples) const;
|
||||
void PrepareBatch(
|
||||
size_t startId,
|
||||
size_t endId,
|
||||
size_t batchSize,
|
||||
const std::vector<size_t> &shuffle,
|
||||
Tensor& xt,
|
||||
Tensor& yt);
|
||||
|
||||
void UpdateModel();
|
||||
float eta_;
|
||||
};
|
||||
|
||||
} // namespace marian
|
||||
class Adagrad {
|
||||
public:
|
||||
Adagrad(float eta=0.1) : eta_(eta) {}
|
||||
|
||||
void operator()(ExpressionGraph& graph, int batchSize) {
|
||||
if(history_.size() < graph.params().size())
|
||||
for(auto& param : graph.params())
|
||||
history_.emplace_back(Tensor(param.grad().shape(), 0));
|
||||
|
||||
graph.backprop(batchSize);
|
||||
|
||||
auto it = history_.begin();
|
||||
for(auto& param : graph.params()) {
|
||||
Element(_1 -= eta_ / Sqrt(_2) * _3, param.val(), *it, param.grad());
|
||||
Element(_1 += _2 * _2, *it, param.grad());
|
||||
it++;
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
float eta_;
|
||||
std::vector<Tensor> history_;
|
||||
};
|
||||
|
||||
}
|
@ -1,8 +1,6 @@
|
||||
#include <fstream>
|
||||
#include "tensor.h"
|
||||
|
||||
using namespace std;
|
||||
|
||||
namespace marian {
|
||||
|
||||
void Tensor::set(const std::vector<float>& data)
|
||||
|
254
src/tensor.h
254
src/tensor.h
@ -1,4 +1,21 @@
|
||||
#pragma once
|
||||
/* Copyright (C)
|
||||
* 2016 - MLAMU & friends
|
||||
* This program is free software; you can redistribute it and/or
|
||||
* modify it under the terms of the GNU General Public License
|
||||
* as published by the Free Software Foundation; either version 2
|
||||
* of the License, or (at your option) any later version.
|
||||
*
|
||||
* This program is distributed in the hope that it will be useful,
|
||||
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
* GNU General Public License for more details.
|
||||
*
|
||||
* You should have received a copy of the GNU General Public License
|
||||
* along with this program; if not, write to the Free Software
|
||||
* Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
|
||||
*
|
||||
*/
|
||||
|
||||
#include <cublas_v2.h>
|
||||
#include <thrust/device_vector.h>
|
||||
@ -12,6 +29,13 @@
|
||||
|
||||
namespace marian {
|
||||
|
||||
/**
|
||||
* @brief Debug shape by printing it.
|
||||
*
|
||||
* @param shape Shape of Tensor.
|
||||
*
|
||||
* @return String of shape.
|
||||
*/
|
||||
inline std::string Debug(const Shape &shape)
|
||||
{
|
||||
std::stringstream strm;
|
||||
@ -23,6 +47,13 @@ inline std::string Debug(const Shape &shape)
|
||||
return strm.str();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Calculate the vector size based on Tensor shape.
|
||||
*
|
||||
* @param shape Shape of Tensor.
|
||||
*
|
||||
* @return Size of Tensor vector.
|
||||
*/
|
||||
inline size_t GetTotalSize(const Shape &shape)
|
||||
{
|
||||
size_t ret = std::accumulate(shape.begin(), shape.end(),
|
||||
@ -30,17 +61,28 @@ inline size_t GetTotalSize(const Shape &shape)
|
||||
return ret;
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief This class manages the Tensor on the GPU.
|
||||
*
|
||||
* @tparam Float Data type.
|
||||
*/
|
||||
template<class Float>
|
||||
class TensorImpl {
|
||||
private:
|
||||
Shape shape_;
|
||||
thrust::device_vector<Float> data_;
|
||||
size_t tno_;
|
||||
static size_t tensorCounter;
|
||||
Shape shape_; /*!< Dimenions of Tensor */
|
||||
thrust::device_vector<Float> data_; /*< Vector of data that Tensor is managing on GPU. */
|
||||
size_t tno_; /*< Tensor number */
|
||||
static size_t tensorCounter; /*< Static counter of created Tensors */
|
||||
|
||||
public:
|
||||
typedef Float value_type;
|
||||
typedef Float value_type; /*< Tensor value type */
|
||||
|
||||
/**
|
||||
* @brief Constructor
|
||||
*
|
||||
* @param shape Shape of Tensor.
|
||||
* @param value Value to fill Tensor's vector with.
|
||||
*/
|
||||
TensorImpl(const Shape& shape, value_type value = 0)
|
||||
: shape_(shape), tno_(tensorCounter++)
|
||||
{
|
||||
@ -59,54 +101,122 @@ class TensorImpl {
|
||||
TensorImpl(const TensorImpl&) = delete;
|
||||
TensorImpl(TensorImpl&&) = delete;
|
||||
|
||||
/**
|
||||
* @brief Get the i-th element of Tensor vector.
|
||||
*
|
||||
* @param i Index.
|
||||
*
|
||||
* @return Value of Tensor vector indexed with i.
|
||||
*/
|
||||
value_type operator[](size_t i) const {
|
||||
return data_[i];
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get begin iterator of Tensor's vector.
|
||||
*
|
||||
* @return Vector begin iterator.
|
||||
*/
|
||||
auto begin() -> decltype( data_.begin() ) {
|
||||
return data_.begin();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get begin iterator of Tensor's vector (const).
|
||||
*
|
||||
* @return Vector begin iterator (const)
|
||||
*/
|
||||
auto begin() const -> decltype( data_.begin() ) {
|
||||
return data_.begin();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get end iterator of Tensor's vector.
|
||||
*
|
||||
* @return Vector end iterator
|
||||
*/
|
||||
auto end() -> decltype( data_.end() ) {
|
||||
return data_.end();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get end iterator of Tensor's vector (const).
|
||||
*
|
||||
* @return Vector end iterator (const)
|
||||
*/
|
||||
auto end() const -> decltype( data_.end() ) {
|
||||
return data_.end();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get Tensor's shape (const)
|
||||
*
|
||||
* @return Shape of Tensor
|
||||
*/
|
||||
const Shape& shape() const {
|
||||
return shape_;
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get size of Tensor's vector.
|
||||
*
|
||||
* @return Length of Tensor's vector.
|
||||
*/
|
||||
size_t size() const {
|
||||
return data_.size();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Cast data from Tensor's GPU to value_type.
|
||||
*
|
||||
* @return Pointer of value_type array.
|
||||
*/
|
||||
value_type* data() {
|
||||
return thrust::raw_pointer_cast(data_.data());
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get Tensor id (number).
|
||||
*
|
||||
* @return Tensor id.
|
||||
*/
|
||||
size_t id() const {
|
||||
return tno_;
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Fill Tensor's vector with specified value on the GPU.
|
||||
*
|
||||
* @param value Value to fill vector with.
|
||||
*/
|
||||
void set(value_type value) {
|
||||
thrust::fill(data_.begin(), data_.end(), value);
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Set Tensor's vector to values of specified vector by copying it to GPU.
|
||||
*
|
||||
* @param begin Begin iterator of a vector.
|
||||
* @param end End iterator of a vector.
|
||||
*/
|
||||
void set(const std::vector<float>::const_iterator &begin, const std::vector<float>::const_iterator &end) {
|
||||
thrust::copy(begin, end, data_.begin());
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Copy Tensor's vector from GPU to vector variable on CPU.
|
||||
*
|
||||
* @param out Vector to copy data to.
|
||||
*/
|
||||
void get(std::vector<float>::iterator out) {
|
||||
thrust::copy(data_.begin(), data_.end(), out);
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Debug function.
|
||||
*
|
||||
* @return Vector in string form.
|
||||
*/
|
||||
std::string Debug() const
|
||||
{
|
||||
std::stringstream strm;
|
||||
@ -133,78 +243,170 @@ class TensorImpl {
|
||||
template <typename Type>
|
||||
size_t TensorImpl<Type>::tensorCounter = 0;
|
||||
|
||||
/**
|
||||
* @brief Class that communicates with GPU's Tensor.
|
||||
*/
|
||||
class Tensor {
|
||||
private:
|
||||
std::shared_ptr<TensorImpl<Float>> pimpl_;
|
||||
std::shared_ptr<TensorImpl<Float>> pimpl_; /*< Pointer to Tensor working on GPU */
|
||||
|
||||
public:
|
||||
typedef TensorImpl<Float>::value_type value_type;
|
||||
typedef TensorImpl<Float>::value_type value_type; /*< Get value type of GPU's Tensor data */
|
||||
|
||||
/**
|
||||
* @brief Default constructor
|
||||
*/
|
||||
Tensor() {}
|
||||
|
||||
/**
|
||||
* @brief Constructor that allocates memory.
|
||||
*
|
||||
* @param shape Shape of Tensor.
|
||||
* @param value Value to fill Tensor's vector with.
|
||||
*/
|
||||
Tensor(const Shape& shape, value_type value = 0) {
|
||||
allocate(shape, value);
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Default destructor
|
||||
*/
|
||||
~Tensor() {}
|
||||
|
||||
/**
|
||||
* @brief Allocate memory if Tensor doesn't exist on GPU. Otherwise, do nothing.
|
||||
*
|
||||
* @param shape Shape of Tensor.
|
||||
* @param value Value to fill Tensor's vector with.
|
||||
*/
|
||||
void allocate(const Shape& shape, value_type value = 0) {
|
||||
if(!pimpl_)
|
||||
pimpl_.reset(new TensorImpl<Float>(shape, value));
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get i-th element of GPU Tensor vector (const).
|
||||
*
|
||||
* @param i Index.
|
||||
*
|
||||
* @return Value of specified element of Tensor.
|
||||
*/
|
||||
value_type operator[](size_t i) const {
|
||||
return (*pimpl_)[i];
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get size of GPU Tensor's vector.
|
||||
*
|
||||
* @return Size of Tensor vector.
|
||||
*/
|
||||
size_t size() const {
|
||||
return pimpl_->size();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Return pointer to GPU Tensor's data.
|
||||
*
|
||||
* @return Pointer to GPU Tensor's data.
|
||||
*/
|
||||
value_type* data() {
|
||||
return pimpl_->data();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Return pointer to GPU Tensor's data (const).
|
||||
*
|
||||
* @return Pointer to GPU Tensor's data.
|
||||
*/
|
||||
const value_type* data() const {
|
||||
return pimpl_->data();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get begin iterator of GPU Tensor's vector.
|
||||
*
|
||||
* @return Vector begin iterator.
|
||||
*/
|
||||
auto begin() -> decltype( pimpl_->begin() ) {
|
||||
return pimpl_->begin();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get begin iterator of GPU Tensor's vector (const).
|
||||
*
|
||||
* @return Vector begin iterator (const)
|
||||
*/
|
||||
auto begin() const -> decltype( pimpl_->begin() ) {
|
||||
return pimpl_->begin();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get end iterator of Tensor's vector.
|
||||
*
|
||||
* @return Vector end iterator
|
||||
*/
|
||||
auto end() -> decltype( pimpl_->end() ) {
|
||||
return pimpl_->end();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get end iterator of Tensor's vector (const).
|
||||
*
|
||||
* @return Vector end iterator (const)
|
||||
*/
|
||||
auto end() const -> decltype( pimpl_->end() ) {
|
||||
return pimpl_->end();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get GPU Tensor's shape.
|
||||
*
|
||||
* @return Tensor's shape.
|
||||
*/
|
||||
const Shape& shape() const {
|
||||
return pimpl_->shape();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Fill GPU Tensor's vector with specified value.
|
||||
*
|
||||
* @param value Value to fill Tensor with.
|
||||
*/
|
||||
void set(value_type value) {
|
||||
pimpl_->set(value);
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Get GPU Tensor id (number).
|
||||
*
|
||||
* @return Tensor id.
|
||||
*/
|
||||
size_t id() const {
|
||||
return pimpl_->id();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Check if Tensor is allocated.
|
||||
*
|
||||
* @return True or False
|
||||
*/
|
||||
operator bool() {
|
||||
return pimpl_ != nullptr;
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Run Debug on GPU Tensor.
|
||||
*
|
||||
* @return String of Tensor's data.
|
||||
*/
|
||||
std::string Debug() const
|
||||
{
|
||||
return pimpl_->Debug();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Print Tensor data on CPU (?) (const).
|
||||
*/
|
||||
void Print() const {
|
||||
for (int i = 0; i < size(); ++i) {
|
||||
std::cerr << (*this)[i] << " ";
|
||||
@ -213,21 +415,59 @@ class Tensor {
|
||||
}
|
||||
|
||||
//void Load(const std::string &path);
|
||||
|
||||
/**
|
||||
* @brief Set GPU Tensor's vector to values of specified vector.
|
||||
*
|
||||
* @param data Vector copied to GPU.
|
||||
*/
|
||||
void set(const std::vector<float>& data);
|
||||
/**
|
||||
* @brief Fill GPU Tensor's vector using values from the specified vector.
|
||||
*
|
||||
* @param begin Begin iterator of vector being copied.
|
||||
* @param end End iterator of vector being copied.
|
||||
*/
|
||||
void set(const std::vector<float>::const_iterator &begin, const std::vector<float>::const_iterator &end);
|
||||
|
||||
/**
|
||||
* @brief Copy Tensor's vector from GPU to vector variable on CPU (const).
|
||||
*
|
||||
* @param out Vector iterator used in copying.
|
||||
*/
|
||||
void get(std::vector<float>::iterator out) const {
|
||||
pimpl_->get(out);
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Copy Tensor's vector from GPU to vector variable on CPU.
|
||||
*
|
||||
* @param out Vector to copy data to.
|
||||
*/
|
||||
void get(std::vector<float> &vout) const {
|
||||
vout.resize(size());
|
||||
pimpl_->get(vout.begin());
|
||||
}
|
||||
};
|
||||
|
||||
/**
|
||||
* @brief Operator to set data on Tensor using vector.
|
||||
*
|
||||
* @param t Tensor.
|
||||
* @param vec Vector used to set data in Tensor.
|
||||
*
|
||||
* @return Tensor with assigned data.
|
||||
*/
|
||||
Tensor& operator<<(Tensor& t, const std::vector<float> &vec);
|
||||
|
||||
/**
|
||||
* @brief Operator to get data from Tensor to vector.
|
||||
*
|
||||
* @param vec Vector to save copied data.
|
||||
* @param t Tensor to copy data from.
|
||||
*
|
||||
* @return Vector with copied data.
|
||||
*/
|
||||
std::vector<float>& operator<<(std::vector<float> &vec, const Tensor& t);
|
||||
|
||||
}
|
||||
|
@ -1,5 +1,7 @@
|
||||
#include "tensor_operators.h"
|
||||
|
||||
using namespace std;
|
||||
|
||||
namespace marian {
|
||||
|
||||
__global__ void gSubtractMean(float* out, float* weights,
|
||||
@ -53,6 +55,7 @@ void SubtractMean(Tensor* Out, Tensor &Weights) {
|
||||
cudaStreamSynchronize(0);
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////
|
||||
__global__ void gSoftMax(float* softMaxP, size_t rows, size_t cols) {
|
||||
for(int bid = 0; bid < rows; bid += gridDim.x) {
|
||||
int j = bid + blockIdx.x;
|
||||
@ -97,6 +100,35 @@ void Softmax(Tensor* Out) {
|
||||
gSoftMax<<<blocks, threads, shared>>>(Out->data(), m, k);
|
||||
cudaStreamSynchronize(0);
|
||||
}
|
||||
///////////////////////////////////////////////////////
|
||||
__global__ void gArgMax(float *out, const float *data, size_t rows, size_t cols) {
|
||||
size_t row = blockIdx.x;
|
||||
size_t startInd = row * cols;
|
||||
float maxScore = -99999;
|
||||
size_t maxInd;
|
||||
for (size_t col = 0; col < cols; ++col) {
|
||||
size_t ind = startInd + col;
|
||||
float score = data[ind];
|
||||
if (score > maxScore) {
|
||||
maxScore = score;
|
||||
maxInd = col;
|
||||
}
|
||||
}
|
||||
out[row] = maxInd;
|
||||
}
|
||||
|
||||
void Argmax(Tensor* Out, const Tensor* In) {
|
||||
size_t m = In->shape()[0];
|
||||
size_t k = In->shape()[1];
|
||||
|
||||
int blocks = m; //std::min(MAX_BLOCKS, (int) m);
|
||||
int threads = k; //std::min(MAX_THREADS, (int) k);
|
||||
//int shared = sizeof(float) * threads * 2;
|
||||
gArgMax<<<blocks, threads>>>(Out->data(), In->data(), m, k);
|
||||
cudaStreamSynchronize(0);
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////
|
||||
|
||||
Tensor Prod(cublasHandle_t handle, Tensor C, const Tensor A, const Tensor B,
|
||||
bool transA, bool transB, Float beta) {
|
||||
@ -137,4 +169,4 @@ Tensor Prod(Tensor C, const Tensor A, const Tensor B,
|
||||
return temp;
|
||||
}
|
||||
|
||||
}
|
||||
}
|
||||
|
@ -151,6 +151,10 @@ __global__ void gSoftMax(float* softMaxP, size_t rows, size_t cols);
|
||||
|
||||
void Softmax(Tensor* Out);
|
||||
|
||||
__global__ void gArgMax(float *out, const float *data, size_t rows, size_t cols);
|
||||
|
||||
void Argmax(Tensor* Out, const Tensor* In);
|
||||
|
||||
Tensor Prod(cublasHandle_t handle, Tensor C, const Tensor A, const Tensor B,
|
||||
bool transA, bool transB, Float beta);
|
||||
|
||||
|
53
src/test.cu
53
src/test.cu
@ -3,7 +3,51 @@
|
||||
#include "mnist.h"
|
||||
#include "vocab.h"
|
||||
|
||||
#include "tensor_operators.h"
|
||||
|
||||
using namespace std;
|
||||
|
||||
///////////////////////////////////////////////////////
|
||||
string output(const std::vector<float> &vec)
|
||||
{
|
||||
stringstream strm;
|
||||
for (size_t i = 0; i < vec.size(); ++i) {
|
||||
strm << vec[i] << " ";
|
||||
}
|
||||
return strm.str();
|
||||
}
|
||||
|
||||
void temp()
|
||||
{
|
||||
using namespace std;
|
||||
using namespace marian;
|
||||
|
||||
std::vector<float> hVec({29,19, 49,39, 79,99, 79,39});
|
||||
cerr << "hVec =" << output(hVec) << endl;
|
||||
|
||||
thrust::device_vector<float> dVec(8);
|
||||
thrust::copy(hVec.begin(), hVec.end(), dVec.begin());
|
||||
float *data = thrust::raw_pointer_cast(dVec.data());
|
||||
|
||||
thrust::device_vector<float> dLabel(4);
|
||||
float *labelPtr = thrust::raw_pointer_cast(dLabel.data());
|
||||
|
||||
gArgMax<<<4, 1, sizeof(float)>>>(labelPtr, data, 4, 2);
|
||||
|
||||
std::vector<float> hVec2(8);
|
||||
thrust::copy(dVec.begin(), dVec.end(), hVec2.begin());
|
||||
cerr << "hVec2=" << output(hVec2) << endl;
|
||||
|
||||
std::vector<float> hLabel(4);
|
||||
thrust::copy(dLabel.begin(), dLabel.end(), hLabel.begin());
|
||||
cerr << "hLabel=" << output(hLabel) << endl;
|
||||
|
||||
exit(0);
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////
|
||||
int main(int argc, char** argv) {
|
||||
temp();
|
||||
|
||||
using namespace std;
|
||||
using namespace marian;
|
||||
@ -21,7 +65,7 @@ int main(int argc, char** argv) {
|
||||
std::vector<Expr> Y;
|
||||
std::vector<Expr> H;
|
||||
|
||||
ExpressionGraph g(0);
|
||||
ExpressionGraph g;
|
||||
|
||||
for (int t = 0; t < num_inputs; ++t) {
|
||||
X.emplace_back(g.input(shape={batch_size, input_size}));
|
||||
@ -39,10 +83,9 @@ int main(int argc, char** argv) {
|
||||
|
||||
string sourceLine, targetLine;
|
||||
while (getline(sourceFile, sourceLine)) {
|
||||
getline(targetFile, targetLine);
|
||||
|
||||
std::vector<size_t> sourceIds = sourceVocab.ProcessSentence(sourceLine);
|
||||
std::vector<size_t> targetIds = sourceVocab.ProcessSentence(targetLine);
|
||||
getline(targetFile, targetLine);
|
||||
std::vector<size_t> sourceIds = sourceVocab.ProcessSentence(sourceLine);
|
||||
std::vector<size_t> targetIds = sourceVocab.ProcessSentence(targetLine);
|
||||
}
|
||||
|
||||
std::cerr << "Building RNN..." << std::endl;
|
||||
|
@ -85,6 +85,19 @@ namespace thrust
|
||||
return compose(unary_operator<unary_tanh>(), _1);
|
||||
}
|
||||
|
||||
template<typename T>
|
||||
struct unary_sqrt : public thrust::unary_function<T,T> {
|
||||
__host__ __device__
|
||||
T operator()(const T &x) const { return sqrtf(x); }
|
||||
};
|
||||
|
||||
template<typename Eval>
|
||||
__host__ __device__
|
||||
actor<composite<unary_operator<unary_sqrt>, actor<Eval>>>
|
||||
Sqrt(const actor<Eval> &_1) {
|
||||
return compose(unary_operator<unary_sqrt>(), _1);
|
||||
}
|
||||
|
||||
template<typename T1, typename T2>
|
||||
__host__ __device__
|
||||
actor<composite<binary_operator<thrust::maximum>, actor<T1>, actor<T2>>>
|
||||
|
@ -11,12 +11,12 @@ int main(int argc, char** argv) {
|
||||
int numofdata;
|
||||
|
||||
vector<float> trainImages = datasets::mnist::ReadImages("../examples/mnist/t10k-images-idx3-ubyte", numofdata, IMAGE_SIZE);
|
||||
vector<float>trainLabels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", numofdata, LABEL_SIZE);
|
||||
vector<float> trainLabels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", numofdata, LABEL_SIZE);
|
||||
|
||||
using namespace marian;
|
||||
using namespace keywords;
|
||||
|
||||
ExpressionGraph g(0);
|
||||
ExpressionGraph g;
|
||||
|
||||
Expr x = named(g.input(shape={whatevs, IMAGE_SIZE}), "x");
|
||||
Expr y = named(g.input(shape={whatevs, LABEL_SIZE}), "y");
|
||||
@ -24,16 +24,13 @@ int main(int argc, char** argv) {
|
||||
Expr w = named(g.param(shape={IMAGE_SIZE, LABEL_SIZE}), "w");
|
||||
Expr b = named(g.param(shape={1, LABEL_SIZE}), "b");
|
||||
|
||||
std::vector<Expr*> params;
|
||||
params.push_back(&w);
|
||||
params.push_back(&b);
|
||||
|
||||
auto scores = dot(x, w) + b;
|
||||
auto lr = softmax_fast(scores);
|
||||
auto cost = named(-mean(sum(y * log(lr), axis=1), axis=0), "cost");
|
||||
cerr << "lr=" << lr.Debug() << endl;
|
||||
|
||||
SGD opt(g, 0.9, trainImages, IMAGE_SIZE, trainLabels, LABEL_SIZE, 3, 24);
|
||||
opt.Run();
|
||||
Adagrad opt;
|
||||
opt(g, 300);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
@ -1,97 +1,129 @@
|
||||
|
||||
#include "marian.h"
|
||||
#include "mnist.h"
|
||||
#include "vocab.h"
|
||||
#include <assert.h>
|
||||
|
||||
using namespace marian;
|
||||
using namespace keywords;
|
||||
|
||||
const int input_size = 10;
|
||||
const int output_size = 15;
|
||||
const int embedding_size = 8;
|
||||
const int hidden_size = 5;
|
||||
const int batch_size = 25;
|
||||
const int num_inputs = 8;
|
||||
const int num_outputs = 6;
|
||||
|
||||
#if 0
|
||||
ExpressionGraph build_graph() {
|
||||
std::cerr << "Loading model params...";
|
||||
std::cerr << "Building computation graph..." << std::endl;
|
||||
|
||||
ExpressionGraph g;
|
||||
std::vector<Expr> X, Y, H, S;
|
||||
|
||||
// We're including the stop symbol here.
|
||||
for (int t = 0; t <= num_inputs; ++t) {
|
||||
std::stringstream ss;
|
||||
ss << "X" << t;
|
||||
X.emplace_back(named(g.input(shape={batch_size, input_size}), ss.str()));
|
||||
}
|
||||
|
||||
// We're including the stop symbol here.
|
||||
for (int t = 0; t <= num_outputs; ++t) {
|
||||
std::stringstream ss;
|
||||
ss << "Y" << t;
|
||||
Y.emplace_back(named(g.input(shape={batch_size, output_size}), ss.str()));
|
||||
}
|
||||
|
||||
// Source embeddings.
|
||||
Expr E = named(g.param(shape={input_size, embedding_size},
|
||||
init=uniform()), "E");
|
||||
|
||||
// Source RNN parameters.
|
||||
Expr Wxh = named(g.param(shape={embedding_size, hidden_size},
|
||||
init=uniform()), "Wxh");
|
||||
Expr Whh = named(g.param(shape={hidden_size, hidden_size},
|
||||
init=uniform()), "Whh");
|
||||
Expr bh = named(g.param(shape={1, hidden_size},
|
||||
init=uniform()), "bh");
|
||||
Expr h0 = named(g.param(shape={1, hidden_size},
|
||||
init=uniform()), "h0");
|
||||
|
||||
std::cerr << "Building encoder RNN..." << std::endl;
|
||||
H.emplace_back(tanh(dot(dot(X[0], E), Wxh) + dot(h0, Whh) + bh));
|
||||
for (int t = 1; t <= num_inputs; ++t) {
|
||||
H.emplace_back(tanh(dot(dot(X[t], E), Wxh) + dot(H[t-1], Whh) + bh));
|
||||
}
|
||||
|
||||
// Target RNN parameters.
|
||||
Expr Wxh_d = named(g.param(shape={output_size, hidden_size},
|
||||
init=uniform()), "Wxh_d");
|
||||
Expr Whh_d = named(g.param(shape={hidden_size, hidden_size},
|
||||
init=uniform()), "Whh_d");
|
||||
Expr bh_d = named(g.param(shape={1, hidden_size},
|
||||
init=uniform()), "bh_d");
|
||||
|
||||
std::cerr << "Building decoder RNN..." << std::endl;
|
||||
auto h0_d = H[num_inputs];
|
||||
S.emplace_back(tanh(dot(Y[0], Wxh_d) + dot(h0_d, Whh_d) + bh_d));
|
||||
for (int t = 1; t < num_outputs; ++t) {
|
||||
S.emplace_back(tanh(dot(Y[t], Wxh_d) + dot(S[t-1], Whh_d) + bh_d));
|
||||
}
|
||||
|
||||
// Output linear layer before softmax.
|
||||
Expr Why = named(g.param(shape={hidden_size, output_size},
|
||||
init=uniform()), "Why");
|
||||
Expr by = named(g.param(shape={1, output_size},
|
||||
init=uniform()), "by");
|
||||
|
||||
std::cerr << "Building output layer..." << std::endl;
|
||||
|
||||
// Softmax layer and cost function.
|
||||
std::vector<Expr> Yp;
|
||||
Yp.emplace_back(named(softmax_fast(dot(h0_d, Why) + by), "pred"));
|
||||
Expr cross_entropy = sum(Y[0] * log(Yp[0]), axis=1);
|
||||
for (int t = 1; t <= num_outputs; ++t) {
|
||||
Yp.emplace_back(named(softmax_fast(dot(S[t-1], Why) + by), "pred"));
|
||||
cross_entropy = cross_entropy + sum(Y[t] * log(Yp[t]), axis=1);
|
||||
}
|
||||
auto cost = named(-mean(cross_entropy, axis=0), "cost");
|
||||
|
||||
std::cerr << "Done." << std::endl;
|
||||
|
||||
return g;
|
||||
}
|
||||
|
||||
int main(int argc, char** argv) {
|
||||
#if 1
|
||||
std::cerr << "Loading the data... ";
|
||||
Vocab sourceVocab, targetVocab;
|
||||
|
||||
// read parallel corpus from file
|
||||
std::fstream sourceFile("../examples/mt/dev/newstest2013.de");
|
||||
std::fstream targetFile("../examples/mt/dev/newstest2013.en");
|
||||
|
||||
std::vector<std::vector<size_t> > source_sentences, target_sentences;
|
||||
std::string sourceLine, targetLine;
|
||||
while (getline(sourceFile, sourceLine)) {
|
||||
getline(targetFile, targetLine);
|
||||
std::vector<size_t> sourceIds = sourceVocab.ProcessSentence(sourceLine);
|
||||
std::vector<size_t> targetIds = sourceVocab.ProcessSentence(targetLine);
|
||||
std::vector<size_t> targetIds = targetVocab.ProcessSentence(targetLine);
|
||||
source_sentences.push_back(sourceIds);
|
||||
target_sentences.push_back(targetIds);
|
||||
}
|
||||
std::cerr << "Done." << std::endl;
|
||||
std::cerr << source_sentences.size()
|
||||
<< " sentence pairs read." << std::endl;
|
||||
std::cerr << "Source vocabulary size: " << sourceVocab.Size() << std::endl;
|
||||
std::cerr << "Target vocabulary size: " << targetVocab.Size() << std::endl;
|
||||
#endif
|
||||
|
||||
// Build the encoder-decoder computation graph.
|
||||
ExpressionGraph g = build_graph();
|
||||
|
||||
int main(int argc, char** argv) {
|
||||
|
||||
using namespace marian;
|
||||
using namespace keywords;
|
||||
|
||||
int input_size = 10;
|
||||
int output_size = 15;
|
||||
int batch_size = 25;
|
||||
int hidden_size = 5;
|
||||
int num_inputs = 8;
|
||||
int num_outputs = 6;
|
||||
|
||||
ExpressionGraph g(0);
|
||||
std::vector<Expr*> X(num_inputs+1); // For the stop symbol.
|
||||
std::vector<Expr*> Y(num_outputs);
|
||||
std::vector<Expr*> H(num_inputs+1); // For the stop symbol.
|
||||
std::vector<Expr*> S(num_outputs);
|
||||
|
||||
// For the stop symbol.
|
||||
for (int t = 0; t <= num_inputs; ++t) {
|
||||
X[t] = new Expr(g.input(shape={batch_size, input_size}));
|
||||
}
|
||||
|
||||
// For the stop symbol.
|
||||
for (int t = 0; t <= num_outputs; ++t) {
|
||||
Y[t] = new Expr(g.input(shape={batch_size, output_size}));
|
||||
}
|
||||
|
||||
Expr Wxh = g.param(shape={input_size, hidden_size}, init=uniform(), name="Wxh");
|
||||
Expr Whh = g.param(shape={hidden_size, hidden_size}, init=uniform(), name="Whh");
|
||||
Expr bh = g.param(shape={1, hidden_size}, init=uniform(), name="bh");
|
||||
Expr h0 = g.param(shape={1, hidden_size}, init=uniform(), name="h0");
|
||||
|
||||
std::cerr << "Building encoder RNN..." << std::endl;
|
||||
H[0] = new Expr(tanh(dot(*X[0], Wxh) + dot(h0, Whh) + bh));
|
||||
for (int t = 1; t <= num_inputs; ++t) {
|
||||
H[t] = new Expr(tanh(dot(*X[t], Wxh) + dot(*H[t-1], Whh) + bh));
|
||||
}
|
||||
|
||||
Expr Wxh_d = g.param(shape={output_size, hidden_size}, init=uniform(), name="Wxh_d");
|
||||
Expr Whh_d = g.param(shape={hidden_size, hidden_size}, init=uniform(), name="Whh_d");
|
||||
Expr bh_d = g.param(shape={1, hidden_size}, init=uniform(), name="bh_d");
|
||||
|
||||
std::cerr << "Building decoder RNN..." << std::endl;
|
||||
auto h0_d = *H[num_inputs];
|
||||
S[0] = new Expr(tanh(dot(*Y[0], Wxh_d) + dot(h0_d, Whh_d) + bh_d));
|
||||
for (int t = 1; t < num_outputs; ++t) {
|
||||
S[t] = new Expr(tanh(dot(*Y[t], Wxh_d) + dot(*S[t-1], Whh_d) + bh_d));
|
||||
}
|
||||
|
||||
Expr Why = g.param(shape={hidden_size, output_size}, init=uniform(), name="Why");
|
||||
Expr by = g.param(shape={1, output_size}, init=uniform(), name="by");
|
||||
|
||||
std::cerr << "Building output layer..." << std::endl;
|
||||
std::vector<Expr*> Yp(num_outputs+1); // For the stop symbol.
|
||||
|
||||
Expr* cross_entropy = NULL;
|
||||
for (int t = 0; t <= num_outputs; ++t) {
|
||||
if (t == 0) {
|
||||
Yp[t] = new Expr(named(softmax_fast(dot(h0_d, Why) + by), "pred"));
|
||||
cross_entropy = new Expr(sum(*Y[t] * log(*Yp[t]), axis=1));
|
||||
} else {
|
||||
Yp[t] = new Expr(named(softmax_fast(dot(*S[t-1], Why) + by), "pred"));
|
||||
*cross_entropy = *cross_entropy + sum(*Y[t] * log(*Yp[t]), axis=1);
|
||||
}
|
||||
}
|
||||
auto graph = -mean(*cross_entropy, axis=0, name="cost");
|
||||
|
||||
// For the stop symbol.
|
||||
// Generate input data (include the stop symbol).
|
||||
for (int t = 0; t <= num_inputs; ++t) {
|
||||
Tensor Xt({batch_size, input_size});
|
||||
|
||||
float max = 1.;
|
||||
std::vector<float> values(batch_size * input_size);
|
||||
std::vector<float> classes(batch_size * output_size, 0.0);
|
||||
@ -101,13 +133,14 @@ int main(int argc, char** argv) {
|
||||
values[k] = max * (2.0*static_cast<float>(rand()) / RAND_MAX - 1.0);
|
||||
}
|
||||
}
|
||||
|
||||
thrust::copy(values.begin(), values.end(), Xt.begin());
|
||||
|
||||
*X[t] = Xt;
|
||||
std::stringstream ss;
|
||||
ss << "X" << t;
|
||||
g[ss.str()] = Xt;
|
||||
}
|
||||
|
||||
for (int t = 0; t < num_outputs; ++t) {
|
||||
// Generate output data (include the stop symbol).
|
||||
for (int t = 0; t <= num_outputs; ++t) {
|
||||
Tensor Yt({batch_size, output_size});
|
||||
|
||||
std::vector<float> classes(batch_size * output_size, 0.0);
|
||||
@ -117,26 +150,31 @@ int main(int argc, char** argv) {
|
||||
classes[l + gold] = 1.0;
|
||||
l += output_size;
|
||||
}
|
||||
|
||||
thrust::copy(classes.begin(), classes.end(), Yt.begin());
|
||||
|
||||
*Y[t] = Yt;
|
||||
std::stringstream ss;
|
||||
ss << "Y" << t;
|
||||
g[ss.str()] = Yt;
|
||||
}
|
||||
|
||||
std::cerr << "Printing the computation graph..." << std::endl;
|
||||
std::cout << g.graphviz() << std::endl;
|
||||
|
||||
std::cerr << "Running the forward step..." << std::endl;
|
||||
g.forward(batch_size);
|
||||
std::cerr << "Running the backward step..." << std::endl;
|
||||
g.backward();
|
||||
std::cerr << "Done." << std::endl;
|
||||
|
||||
std::cerr << graph.val().Debug() << std::endl;
|
||||
std::cerr << g["cost"].val().Debug() << std::endl;
|
||||
|
||||
std::cerr << X[0]->val().Debug() << std::endl;
|
||||
std::cerr << Y[0]->val().Debug() << std::endl;
|
||||
|
||||
std::cerr << Whh.grad().Debug() << std::endl;
|
||||
std::cerr << bh.grad().Debug() << std::endl;
|
||||
std::cerr << Why.grad().Debug() << std::endl;
|
||||
std::cerr << by.grad().Debug() << std::endl;
|
||||
std::cerr << Wxh.grad().Debug() << std::endl;
|
||||
std::cerr << h0.grad().Debug() << std::endl;
|
||||
std::cerr << g["X0"].val().Debug() << std::endl;
|
||||
std::cerr << g["Y0"].val().Debug() << std::endl;
|
||||
std::cerr << g["Whh"].grad().Debug() << std::endl;
|
||||
std::cerr << g["bh"].grad().Debug() << std::endl;
|
||||
std::cerr << g["Why"].grad().Debug() << std::endl;
|
||||
std::cerr << g["by"].grad().Debug() << std::endl;
|
||||
std::cerr << g["Wxh"].grad().Debug() << std::endl;
|
||||
std::cerr << g["h0"].grad().Debug() << std::endl;
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
@ -10,7 +10,7 @@ const size_t IMAGE_SIZE = 784;
|
||||
const size_t LABEL_SIZE = 10;
|
||||
int BATCH_SIZE = 10000;
|
||||
|
||||
ExpressionGraph build_graph(int cudaDevice) {
|
||||
ExpressionGraph build_graph() {
|
||||
std::cerr << "Loading model params...";
|
||||
NpzConverter converter("../scripts/test_model_single/model.npz");
|
||||
|
||||
@ -22,7 +22,7 @@ ExpressionGraph build_graph(int cudaDevice) {
|
||||
|
||||
std::cerr << "Building model...";
|
||||
|
||||
ExpressionGraph g(cudaDevice);
|
||||
ExpressionGraph g;
|
||||
auto x = named(g.input(shape={whatevs, IMAGE_SIZE}), "x");
|
||||
auto y = named(g.input(shape={whatevs, LABEL_SIZE}), "y");
|
||||
|
||||
@ -52,7 +52,7 @@ int main(int argc, char** argv) {
|
||||
std::vector<float> testLabels = datasets::mnist::ReadLabels("../examples/mnist/t10k-labels-idx1-ubyte", BATCH_SIZE, LABEL_SIZE);
|
||||
std::cerr << "Done." << std::endl;
|
||||
|
||||
ExpressionGraph g = build_graph(0);
|
||||
ExpressionGraph g = build_graph();
|
||||
|
||||
Tensor xt({BATCH_SIZE, IMAGE_SIZE});
|
||||
Tensor yt({BATCH_SIZE, LABEL_SIZE});
|
||||
|
@ -56,8 +56,7 @@ int main(int argc, char** argv) {
|
||||
|
||||
std::cerr << "\tDone." << std::endl;
|
||||
|
||||
|
||||
ExpressionGraph g(0);
|
||||
ExpressionGraph g;
|
||||
|
||||
auto x = g.input(shape={whatevs, IMAGE_SIZE}, name="X");
|
||||
auto y = g.input(shape={whatevs, LABEL_SIZE}, name="Y");
|
||||
|
@ -24,22 +24,6 @@ inline std::vector<std::string> Tokenize(const std::string& str,
|
||||
|
||||
return tokens;
|
||||
}
|
||||
////////////////////////////////////////////////////////
|
||||
size_t Vocab::GetUNK() const
|
||||
{
|
||||
return std::numeric_limits<size_t>::max();
|
||||
}
|
||||
|
||||
size_t Vocab::GetPad() const
|
||||
{
|
||||
return std::numeric_limits<size_t>::max() - 1;
|
||||
}
|
||||
|
||||
size_t Vocab::GetEOS() const
|
||||
{
|
||||
return std::numeric_limits<size_t>::max() - 2;
|
||||
}
|
||||
|
||||
|
||||
size_t Vocab::GetOrCreate(const std::string &word)
|
||||
{
|
||||
@ -55,6 +39,12 @@ size_t Vocab::GetOrCreate(const std::string &word)
|
||||
return id;
|
||||
}
|
||||
|
||||
size_t Vocab::Get(const std::string &word) const
|
||||
{
|
||||
Coll::const_iterator iter = coll_.find(word);
|
||||
return iter->second;
|
||||
}
|
||||
|
||||
std::vector<size_t> Vocab::ProcessSentence(const std::string &sentence)
|
||||
{
|
||||
vector<string> toks = Tokenize(sentence);
|
||||
|
18
src/vocab.h
18
src/vocab.h
@ -7,12 +7,22 @@
|
||||
class Vocab
|
||||
{
|
||||
public:
|
||||
size_t GetOrCreate(const std::string &word);
|
||||
Vocab() {
|
||||
GetOrCreate("__UNK__");
|
||||
GetOrCreate("__PAD__");
|
||||
GetOrCreate("__EOS__");
|
||||
}
|
||||
virtual ~Vocab() {}
|
||||
|
||||
public:
|
||||
size_t Size() const { return coll_.size(); }
|
||||
size_t Get(const std::string &word) const;
|
||||
size_t GetOrCreate(const std::string &word);
|
||||
std::vector<size_t> ProcessSentence(const std::string &sentence);
|
||||
|
||||
size_t GetUNK() const;
|
||||
size_t GetPad() const;
|
||||
size_t GetEOS() const;
|
||||
size_t GetUNK() const { return Get("__UNK__"); }
|
||||
size_t GetPAD() const { return Get("__PAD__"); }
|
||||
size_t GetEOS() const { return Get("__EOS__"); }
|
||||
protected:
|
||||
typedef std::unordered_map<std::string, size_t> Coll;
|
||||
Coll coll_;
|
||||
|
Loading…
Reference in New Issue
Block a user