m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

clean up unused code.

Browse Source
avx_docs
xutianbing 8 years ago
parent ea4d08dab6
commit ec6b13dbfc
7 changed files with 53 additions and 747 deletions
Show Stats Download Patch File Download Diff File

72
paddle/cuda/include/hl_sequence.h
Unescape View File

@ -48,78 +48,6 @@ extern void hl_max_sequence_forward(real* input,
extern void hl_max_sequence_backward(
real* outputGrad, int* index, real* inputGrad, int numSequences, int dim);
/**
* @brief Context projection forward.
*
* @param[in] input input sequence.
* @param[in] sequence sequence index.
* @param[in] weightData padding data.
* @param[out] output output sequence.
* @param[in] numSequences number of sequences.
* @param[in] inputDim input sequence dimension.
* @param[in] contextLength context length.
* @param[in] contextStart context start.
* @param[in] beginPad number of extra timesteps added at the
* beginning.
* @param[in] isPadding trainable padding.
*
*/
extern void hl_context_projection_forward(real* input,
const int* sequence,
real* weightData,
real* output,
int numSequences,
int inputDim,
int contextLength,
int contextStart,
int beginPad,
bool isPadding);
/**
* @brief Context projection backward data.
*
* @param[in] outputGrad output gradient.
* @param[in] sequence sequence index.
* @param[out] inputGrad input gradient.
* @param[in] numSequences number of sequences.
* @param[in] inputDim input sequence dimension.
* @param[in] contextLength context length.
* @param[in] contextStart context start.
*
*/
extern void hl_context_projection_backward_data(real* outputGrad,
const int* sequence,
real* inputGrad,
int numSequences,
int inputDim,
int contextLength,
int contextStart);
/**
* @brief Context projection backward weight.
*
* @param[in] outputGrad output gradient.
* @param[in] sequence sequence index.
* @param[out] weightGrad weight gradient.
* @param[in] numSequences number of sequences.
* @param[in] weightDim input sequence dimension.
* @param[in] totalPad number of extra timesteps.
* @param[in] contextLength context length.
* @param[in] contextStart context start.
* @param[in] beginPad number of extra timesteps added at the
* beginning.
*
*/
extern void hl_context_projection_backward_weight(real* outputGrad,
const int* sequence,
real* weightGrad,
int numSequences,
int weightDim,
int totalPad,
int contextLength,
int contextStart,
int beginPad);
/**
* @brief Memory copy from sequence to batch.
*

29
paddle/cuda/include/stub/hl_sequence_stub.h
Unescape View File

@ -27,35 +27,6 @@ inline void hl_max_sequence_forward(real* input,
inline void hl_max_sequence_backward(
real* outputGrad, int* index, real* inputGrad, int numSequences, int dim) {}
inline void hl_context_projection_forward(real* input,
const int* sequence,
real* weightData,
real* output,
int numSequences,
int inputDim,
int contextLength,
int contextStart,
int beginPad,
bool isPadding) {}
inline void hl_context_projection_backward_data(real* outputGrad,
const int* sequence,
real* inputGrad,
int numSequences,
int inputDim,
int contextLength,
int contextStart) {}
inline void hl_context_projection_backward_weight(real* outputGrad,
const int* sequence,
real* weightGrad,
int numSequences,
int weightDim,
int totalPad,
int contextLength,
int contextStart,
int beginPad) {}
inline void hl_sequence2batch_copy(real* batch,
real* sequence,
const int* batchIndex,

252
paddle/cuda/src/hl_cuda_sequence.cu
View File

File diff suppressed because it is too large Load Diff

64
paddle/function/ContextProjectionOpGpu.cu
Unescape View File

@ -73,15 +73,30 @@ __global__ void KeContextProjectionForward(const real* input,
}
}
/**
* @brief Context projection forward.
*
* @param[in] input input sequence.
* @param[in] sequence sequence index.
* @param[in] weight padding data.
* @param[out] output output sequence.
* @param[in] num_sequences number of sequences.
* @param[in] input_dim input sequence dimension.
* @param[in] context_length context length.
* @param[in] context_start context start.
* @param[in] begin_pad number of extra timesteps added at the
* beginning.
*
*/
void hl_context_projection_forward(const real* input,
const int* sequence,
const real* weight,
real* output,
int num_sequences,
int input_dim,
int context_length,
size_t num_sequences,
size_t input_dim,
size_t context_length,
int context_start,
int begin_pad) {
size_t begin_pad) {
CHECK_NOTNULL(input);
CHECK_NOTNULL(sequence);
CHECK_NOTNULL(output);
@ -168,12 +183,24 @@ __global__ void KeContextProjectionBackwardData(real* out_grad,
}
}
/**
* @brief Context projection backward data.
*
* @param[in] out_grad output gradient.
* @param[in] sequence sequence index.
* @param[out] input_grad input gradient.
* @param[in] num_sequences number of sequences.
* @param[in] input_dim input sequence dimension.
* @param[in] context_length context length.
* @param[in] context_start context start.
*
*/
void hl_context_projection_backward_data(real* out_grad,
const int* sequence,
real* input_grad,
int num_sequences,
int input_dim,
int context_length,
size_t num_sequences,
size_t input_dim,
size_t context_length,
int context_start) {
CHECK_NOTNULL(out_grad);
CHECK_NOTNULL(sequence);
@ -278,15 +305,30 @@ __global__ void KeContextProjectionBackwardWeight(real* out_grad,
}
}
/**
* @brief Context projection backward weight.
*
* @param[in] out_grad output gradient.
* @param[in] sequence sequence index.
* @param[out] w_grad weight gradient.
* @param[in] num_sequences number of sequences.
* @param[in] w_dim input sequence dimension.
* @param[in] total_pad number of extra timesteps.
* @param[in] context_length context length.
* @param[in] context_start context start.
* @param[in] begin_pad number of extra timesteps added at the
* beginning.
*
*/
void hl_context_projection_backward_weight(real* out_grad,
const int* sequence,
real* w_grad,
int num_sequences,
int w_dim,
size_t num_sequences,
size_t w_dim,
size_t total_pad,
int context_length,
size_t context_length,
int context_start,
int begin_pad) {
size_t begin_pad) {
CHECK_NOTNULL(out_grad);
CHECK_NOTNULL(sequence);
CHECK_NOTNULL(w_grad);

169
paddle/math/Matrix.cpp
Unescape View File

@ -1304,68 +1304,6 @@ void GpuMatrix::maxSequenceBackward(Matrix& outputGrad,
hl_max_sequence_backward(outGrad, maxIndex, inputGrad, numSequences, dim);
}
void GpuMatrix::contextProjectionForward(Matrix& input,
Matrix* weight,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding) {
CHECK(dynamic_cast<GpuMatrix*>(&input));
CHECK(dynamic_cast<const GpuIVector*>(&sequence));
if (weight) CHECK(dynamic_cast<GpuMatrix*>(weight));
CHECK_EQ(getWidth(), input.getWidth() * contextLength);
hl_context_projection_forward(input.getData(),
sequence.getData(),
isPadding ? weight->getData() : NULL,
getData(),
sequence.getSize() - 1,
input.getWidth(),
contextLength,
contextStart,
beginPad,
isPadding);
}
void GpuMatrix::contextProjectionBackwardData(Matrix& inputGrad,
const IVector& sequence,
int contextLength,
int contextStart) {
CHECK(dynamic_cast<GpuMatrix*>(&inputGrad));
CHECK(dynamic_cast<const GpuIVector*>(&sequence));
CHECK_EQ(getWidth(), inputGrad.getWidth() * contextLength);
hl_context_projection_backward_data(getData(),
sequence.getData(),
inputGrad.getData(),
sequence.getSize() - 1,
inputGrad.getWidth(),
contextLength,
contextStart);
}
void GpuMatrix::contextProjectionBackwardWeight(Matrix& weightGrad,
const IVector& sequence,
int contextLength,
int contextStart,
int totalPad,
size_t beginPad) {
CHECK(dynamic_cast<GpuMatrix*>(&weightGrad));
CHECK(dynamic_cast<const GpuIVector*>(&sequence));
CHECK_EQ(getWidth(), weightGrad.getWidth() * contextLength);
hl_context_projection_backward_weight(getData(),
sequence.getData(),
weightGrad.getData(),
sequence.getSize() - 1,
weightGrad.getWidth(),
totalPad,
contextLength,
contextStart,
beginPad);
}
void GpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
CHECK(data.useGpu_ == true && W.useGpu_ == true)
<< "Matrix type are not equal";
@ -2203,113 +2141,6 @@ void CpuMatrix::maxSequenceBackward(Matrix& outputGrad,
}
}
void CpuMatrix::contextProjectionForward(Matrix& input,
Matrix* weight,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding) {
auto input_ptr = dynamic_cast<CpuMatrix*>(&input);
auto seq_ptr = dynamic_cast<const CpuIVector*>(&sequence);
CHECK(input_ptr && seq_ptr);
if (weight) CHECK(dynamic_cast<CpuMatrix*>(weight));
CHECK_EQ(getWidth(), input_ptr->getWidth() * contextLength);
const int* starts = seq_ptr->getData();
size_t numSequences = seq_ptr->getSize() - 1;
for (size_t i = 0; i < numSequences; ++i) {
for (int j = 0; j < contextLength; ++j) {
int begin = starts[i] + contextStart + j;
int end = starts[i + 1] + contextStart + j;
int dstBegin = starts[i];
int dstEnd = starts[i + 1];
if (begin < starts[i]) {
int64_t padSize =
std::min(starts[i] - begin, starts[i + 1] - starts[i]);
MatrixPtr mat = this->subMatrix(starts[i], padSize);
if (isPadding) {
MatrixPtr sub = weight->subMatrix(j, padSize);
mat->addAtOffset(*sub, j * input_ptr->getWidth());
}
dstBegin = starts[i] + padSize;
begin = starts[i];
}
if (end > starts[i + 1]) {
int64_t padSize =
std::min(end - starts[i + 1], starts[i + 1] - starts[i]);
MatrixPtr mat = this->subMatrix(starts[i + 1] - padSize, padSize);
if (isPadding) {
MatrixPtr sub =
weight->subMatrix(beginPad + contextStart + j - padSize, padSize);
mat->addAtOffset(*sub, j * input_ptr->getWidth());
}
dstEnd = starts[i + 1] - padSize;
end = starts[i + 1];
}
if (end <= begin) continue;
MatrixPtr src = input_ptr->subMatrix(begin, end - begin);
MatrixPtr dst = this->subMatrix(dstBegin, dstEnd - dstBegin);
dst->addAtOffset(*src, j * input_ptr->getWidth());
}
}
}
void CpuMatrix::contextProjectionBackward(Matrix* inputGrad,
Matrix* weightGrad,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding) {
if (inputGrad) CHECK(dynamic_cast<CpuMatrix*>(inputGrad));
if (weightGrad) CHECK(dynamic_cast<CpuMatrix*>(weightGrad));
CHECK(dynamic_cast<const CpuIVector*>(&sequence));
int64_t inputDim = inputGrad ? inputGrad->getWidth()
: weightGrad ? weightGrad->getWidth() : 0;
CHECK_EQ(getWidth(), inputDim * contextLength);
const int* starts = sequence.getData();
size_t numSequences = sequence.getSize() - 1;
for (size_t i = 0; i < numSequences; ++i) {
for (int j = 0; j < contextLength; ++j) {
int begin = starts[i] + contextStart + j;
int end = starts[i + 1] + contextStart + j;
int dstBegin = starts[i];
int dstEnd = starts[i + 1];
if (begin < starts[i]) {
int64_t padSize =
std::min(starts[i] - begin, starts[i + 1] - starts[i]);
if (isPadding && weightGrad) {
MatrixPtr mat = this->subMatrix(starts[i], padSize);
MatrixPtr sub = weightGrad->subMatrix(j, padSize);
sub->addAtOffset(*mat, j * inputDim);
}
dstBegin = starts[i] + padSize;
begin = starts[i];
}
if (end > starts[i + 1]) {
int64_t padSize =
std::min(end - starts[i + 1], starts[i + 1] - starts[i]);
if (isPadding && weightGrad) {
MatrixPtr mat = this->subMatrix(starts[i + 1] - padSize, padSize);
MatrixPtr sub = weightGrad->subMatrix(
beginPad + contextStart + j - padSize, padSize);
sub->addAtOffset(*mat, j * inputDim);
}
dstEnd = starts[i + 1] - padSize;
end = starts[i + 1];
}
if (end <= begin) continue;
if (!inputGrad) continue;
MatrixPtr src = inputGrad->subMatrix(begin, end - begin);
MatrixPtr dst = this->subMatrix(dstBegin, dstEnd - dstBegin);
src->addAtOffset(*dst, j * inputDim);
}
}
}
inline void vecAddTo(real* a, const real* b, size_t len) {
for (unsigned int i = 0; i < len; ++i) {
a[i] += b[i];

72
paddle/math/Matrix.h
Unescape View File

@ -972,42 +972,6 @@ public:
LOG(FATAL) << "Not implemeted";
}
virtual void contextProjectionForward(Matrix& input,
Matrix* weight,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding) {
LOG(FATAL) << "Not implemeted";
}
virtual void contextProjectionBackward(Matrix* inputGrad,
Matrix* weightGrad,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding) {
LOG(FATAL) << "Not implemeted";
}
virtual void contextProjectionBackwardData(Matrix& inputGrad,
const IVector& sequence,
int contextLength,
int contextStart) {
LOG(FATAL) << "Not implemeted";
}
virtual void contextProjectionBackwardWeight(Matrix& weightGrad,
const IVector& sequence,
int contextLength,
int contextStart,
int totalPad,
size_t beginPad) {
LOG(FATAL) << "Not implemeted";
}
/**
* @code
* this.row[i] += table.row[ids[i]]
@ -1442,26 +1406,6 @@ public:
const IVector& sequence,
IVector& index);
void contextProjectionForward(Matrix& input,
Matrix* weight,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding);
void contextProjectionBackwardData(Matrix& inputGrad,
const IVector& sequence,
int contextLength,
int contextStart);
void contextProjectionBackwardWeight(Matrix& weightGrad,
const IVector& sequence,
int contextLength,
int contextStart,
int totalPad,
size_t beginPad);
void bilinearForward(const Matrix& in,
const size_t inImgH,
const size_t inImgW,
@ -1648,22 +1592,6 @@ public:
const IVector& sequence,
IVector& index);
void contextProjectionForward(Matrix& input,
Matrix* weight,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding);
void contextProjectionBackward(Matrix* inputGrad,
Matrix* weightGrad,
const IVector& sequence,
int contextLength,
int contextStart,
size_t beginPad,
bool isPadding);
real* getRow(size_t row) { return BaseMatrix::rowBuf(row); }
virtual real* getRowBuf(size_t row) { return getRow(row); }

142
paddle/math/tests/test_matrixCompare.cpp
Unescape View File

@ -29,148 +29,6 @@ using namespace std; // NOLINT
using autotest::TensorCheckEqual;
using autotest::TensorCheckErr;
void testMatrixProjectionForward(int contextStart,
int contextLength,
bool padding,
int batchSize,
int inputDim) {
MatrixPtr cpuInput = std::make_shared<CpuMatrix>(batchSize, inputDim);
MatrixPtr gpuInput = std::make_shared<GpuMatrix>(batchSize, inputDim);
cpuInput->randomizeUniform();
gpuInput->copyFrom(*cpuInput);
int pad = std::max(0, -contextStart) +
std::max(0, contextStart + contextLength - 1);
if (pad == 0) padding = false;
MatrixPtr cpuWeight = nullptr;
MatrixPtr gpuWeight = nullptr;
if (padding) {
cpuWeight = std::make_shared<CpuMatrix>(pad, inputDim);
gpuWeight = std::make_shared<GpuMatrix>(pad, inputDim);
cpuWeight->randomizeUniform();
gpuWeight->copyFrom(*cpuWeight);
}
IVectorPtr cpuSequence;
generateSequenceStartPositions(batchSize, cpuSequence);
IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true);
gpuSequence->copyFrom(*cpuSequence);
MatrixPtr cpuOutput =
std::make_shared<CpuMatrix>(batchSize, inputDim * contextLength);
MatrixPtr gpuOutput =
std::make_shared<GpuMatrix>(batchSize, inputDim * contextLength);
cpuOutput->randomizeUniform();
gpuOutput->copyFrom(*cpuOutput);
// calculate
int beginPad = std::max(0, -contextStart);
cpuOutput->contextProjectionForward(*cpuInput,
cpuWeight.get(),
*cpuSequence,
contextLength,
contextStart,
beginPad,
padding);
gpuOutput->contextProjectionForward(*gpuInput,
gpuWeight.get(),
*gpuSequence,
contextLength,
contextStart,
beginPad,
padding);
TensorCheckEqual(*cpuOutput, *gpuOutput);
}
void testMatrixProjectionBackward(int contextStart,
int contextLength,
bool padding,
int batchSize,
int inputDim) {
MatrixPtr cpuOutputGrad =
std::make_shared<CpuMatrix>(batchSize, inputDim * contextLength);
MatrixPtr gpuOutputGrad =
std::make_shared<GpuMatrix>(batchSize, inputDim * contextLength);
cpuOutputGrad->randomizeUniform();
gpuOutputGrad->copyFrom(*cpuOutputGrad);
IVectorPtr cpuSequence;
generateSequenceStartPositions(batchSize, cpuSequence);
IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true);
gpuSequence->copyFrom(*cpuSequence);
MatrixPtr cpuInputGrad = std::make_shared<CpuMatrix>(batchSize, inputDim);
MatrixPtr gpuInputGrad = std::make_shared<GpuMatrix>(batchSize, inputDim);
cpuInputGrad->randomizeUniform();
gpuInputGrad->copyFrom(*cpuInputGrad);
int pad = std::max(0, -contextStart) +
std::max(0, contextStart + contextLength - 1);
if (pad == 0) padding = false;
MatrixPtr cpuWeightGrad = nullptr;
MatrixPtr gpuWeightGrad = nullptr;
if (padding) {
cpuWeightGrad = std::make_shared<CpuMatrix>(pad, inputDim);
gpuWeightGrad = std::make_shared<GpuMatrix>(pad, inputDim);
cpuWeightGrad->randomizeUniform();
gpuWeightGrad->copyFrom(*cpuWeightGrad);
}
// calculate
int beginPad = std::max(0, -contextStart);
cpuOutputGrad->contextProjectionBackward(cpuInputGrad.get(),
cpuWeightGrad.get(),
*cpuSequence,
contextLength,
contextStart,
beginPad,
padding);
gpuOutputGrad->contextProjectionBackwardData(
*gpuInputGrad, *gpuSequence, contextLength, contextStart);
if (padding) {
gpuOutputGrad->contextProjectionBackwardWeight(*gpuWeightGrad,
*gpuSequence,
contextLength,
contextStart,
pad,
beginPad);
}
TensorCheckErr(*cpuInputGrad, *gpuInputGrad);
if (padding) {
TensorCheckErr(*cpuWeightGrad, *gpuWeightGrad);
}
}
TEST(Matrix, projection) {
for (auto contextStart : {-5, -3, -1, 0, 3}) {
for (auto contextLength : {1, 2, 5, 7}) {
for (auto trainablePadding : {false, true}) {
for (auto batchSize : {1, 2, 5, 20, 100}) {
for (auto inputDim : {15, 32, 63, 128, 200}) {
VLOG(3) << " contextStart=" << contextStart
<< " contextLength=" << contextLength
<< " trainablePadding=" << trainablePadding
<< " batchSize=" << batchSize << " inputDim=" << inputDim;
testMatrixProjectionForward(contextStart,
contextLength,
trainablePadding,
batchSize,
inputDim);
testMatrixProjectionBackward(contextStart,
contextLength,
trainablePadding,
batchSize,
inputDim);
}
}
}
}
}
}
void testMatrixMaxSequence(int batchSize, int inputDim) {
// forward
MatrixPtr cpuInput = std::make_shared<CpuMatrix>(batchSize, inputDim);

Write Preview
Loading…
Cancel
Save