提交 17152510 编写于 作者: F fengjiayi

update CPU sequence_padding functor

上级 8b9938ac
...@@ -18,37 +18,45 @@ namespace paddle { ...@@ -18,37 +18,45 @@ namespace paddle {
namespace operators { namespace operators {
namespace math { namespace math {
enum CopyType { kSeqToPad, kPadToSeq };
template <typename T> template <typename T>
void CopyDataCPU(framework::LoDTensor* seq_tensor, void CopyValidData(framework::Tensor* dst_tensor,
framework::Tensor* pad_tensor, const framework::Tensor* src_tensor,
const framework::Vector<size_t>& seq_offset, const framework::Vector<size_t>& seq_offsets,
const int64_t& max_seq_len, const int64_t& seq_width, int pad_seq_len, int step_width, bool norm_by_len,
bool seq_to_pad, bool norm_by_len, CopyType type, PadLayout layout) {
OutputLayout output_layout) { int seq_num = seq_offsets.size() - 1;
T* seq_data = seq_tensor->data<T>(); const T* src_data = src_tensor->data<T>();
T* pad_data = pad_tensor->data<T>(); T* dst_data = dst_tensor->data<T>();
int64_t seq_num = seq_offset.size() - 1; int seq_cpy_gap = step_width;
int pad_cpy_gap =
for (int64_t i = 0; i < seq_num; ++i) { layout == kBatchLengthWidth ? step_width : seq_num * step_width;
int64_t seq_start = seq_offset[i]; for (int seq_idx = 0; seq_idx < seq_num; ++seq_idx) {
int64_t seq_len = seq_offset[i + 1] - seq_start; int valid_seq_len = seq_offsets[seq_idx + 1] - seq_offsets[seq_idx];
T scale = norm_by_len ? (1.0f / static_cast<T>(seq_len)) : 1.0f; PADDLE_ENFORCE_GE(
for (int64_t j = 0; j < seq_len; ++j) { pad_seq_len, valid_seq_len,
for (int64_t k = 0; k < seq_width; ++k) { "The padded sequence length can not be less than its original length.");
size_t pad_data_idx = 0; int seq_data_offset = seq_offsets[seq_idx] * step_width;
size_t seq_data_idx = (seq_start + j) * seq_width + k; int pad_data_offset = layout == kBatchLengthWidth
if (output_layout == kBatchLengthWidth) { ? seq_idx * pad_seq_len * step_width
pad_data_idx = (i * max_seq_len + j) * seq_width + k; : seq_idx * step_width;
} else { float scale = 1.0f / static_cast<float>(valid_seq_len);
pad_data_idx = (j * seq_num + i) * seq_width + k;
} for (int step_idx = 0; step_idx < valid_seq_len; ++step_idx) {
if (seq_to_pad) { const T* src =
pad_data[pad_data_idx] = seq_data[seq_data_idx] * scale; src_data + (type == kSeqToPad ? seq_data_offset : pad_data_offset);
} else { T* dst =
seq_data[seq_data_idx] = pad_data[pad_data_idx] * scale; dst_data + (type == kSeqToPad ? pad_data_offset : seq_data_offset);
memcpy(dst, src, step_width * sizeof(T));
if (norm_by_len) {
for (int i = 0; i < step_width; ++i) {
*(dst + i) *= scale;
} }
} }
seq_data_offset += seq_cpy_gap;
pad_data_offset += pad_cpy_gap;
} }
} }
} }
...@@ -58,31 +66,37 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> { ...@@ -58,31 +66,37 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
public: public:
void operator()(const platform::CPUDeviceContext& context, void operator()(const platform::CPUDeviceContext& context,
const framework::LoDTensor& seq_tensor, const framework::LoDTensor& seq_tensor,
framework::Tensor* pad_tensor, framework::LoDTensor* pad_tensor,
T pad_value = static_cast<T>(0), bool norm_by_times = false, std::vector<T> pad_value = {0}, int pad_seq_len = -1,
size_t lod_level = 0, int lod_level = 0, bool norm_by_times = false,
OutputLayout output_layout = kBatchLengthWidth) { const PadLayout layout = kBatchLengthWidth) {
CheckLoD(seq_tensor, lod_level); auto seq_offsets = framework::ToAbsOffset(seq_tensor.lod())[lod_level];
auto& lod = seq_tensor.lod();
auto& seq_offset = framework::ToAbsOffset(lod)[lod_level];
auto seq_tensor_dims = seq_tensor.dims(); auto seq_tensor_dims = seq_tensor.dims();
auto pad_tensor_dims = pad_tensor->dims(); auto pad_tensor_dims = pad_tensor->dims();
int64_t max_seq_len = MaximumSequenceLength(seq_offset); if (pad_seq_len == -1) {
int64_t seq_num = seq_offset.size() - 1; pad_seq_len = MaximumSequenceLength(seq_offsets);
int64_t seq_width = seq_tensor.numel() / seq_tensor_dims[0]; }
int step_width = seq_tensor.numel() / seq_tensor_dims[0];
CheckDims(seq_tensor_dims, seq_offset.back(), pad_tensor_dims, max_seq_len, CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len,
seq_num, seq_width, output_layout); step_width, layout);
PADDLE_ENFORCE(pad_value.size() == 1 ||
static_cast<int>(pad_value.size()) == step_width,
"The size of 'pad_value' can only be 1 or be equal to the "
"'step_width'.");
T* pad_data = pad_tensor->data<T>(); if (pad_value.size() == 1) {
pad_value = std::vector<T>(step_width, pad_value[0]);
}
memset(pad_data, pad_value, max_seq_len * seq_num * seq_width * sizeof(T)); // fill padding value
T* pad_data = pad_tensor->data<T>();
for (int i = 0; i < pad_tensor->numel() / step_width; ++i) {
memcpy(pad_data, pad_value.data(), step_width * sizeof(T));
}
CopyDataCPU<T>(const_cast<framework::LoDTensor*>(&seq_tensor), pad_tensor, CopyValidData<T>(pad_tensor, &seq_tensor, seq_offsets, pad_seq_len,
seq_offset, max_seq_len, seq_width, true /* seq_to_pad */, step_width, norm_by_times, kSeqToPad, layout);
norm_by_times, output_layout);
} }
}; };
...@@ -90,30 +104,23 @@ template <typename T> ...@@ -90,30 +104,23 @@ template <typename T>
class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> { class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
public: public:
void operator()(const platform::CPUDeviceContext& context, void operator()(const platform::CPUDeviceContext& context,
framework::LoDTensor* seq_tensor, const framework::LoDTensor& pad_tensor,
const framework::Tensor& pad_tensor, framework::LoDTensor* seq_tensor, int pad_seq_len = -1,
bool norm_by_times = false, size_t lod_level = 0, int lod_level = 0, bool norm_by_times = false,
OutputLayout output_layout = kBatchLengthWidth) { const PadLayout& layout = kBatchLengthWidth) {
CheckLoD(*seq_tensor, lod_level); auto seq_offsets = framework::ToAbsOffset(seq_tensor->lod())[lod_level];
auto seq_tensor_dims = seq_tensor->dims();
auto& lod = seq_tensor->lod(); auto pad_tensor_dims = pad_tensor.dims();
auto& seq_offset = framework::ToAbsOffset(lod)[lod_level]; if (pad_seq_len == -1) {
pad_seq_len = MaximumSequenceLength(seq_offsets);
auto& seq_tensor_dims = seq_tensor->dims(); }
auto& pad_tensor_dims = pad_tensor.dims(); int step_width = seq_tensor->numel() / seq_tensor_dims[0];
int64_t max_seq_len = MaximumSequenceLength(seq_offset);
int64_t seq_num = seq_offset.size() - 1; CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len,
int64_t seq_width = seq_tensor->numel() / seq_tensor_dims[0]; step_width, layout);
CheckDims(seq_tensor_dims, seq_offset.back(), pad_tensor_dims, max_seq_len, CopyValidData<T>(seq_tensor, &pad_tensor, seq_offsets, pad_seq_len,
seq_num, seq_width, output_layout); step_width, norm_by_times, kPadToSeq, layout);
T* seq_data = seq_tensor->data<T>();
memset(seq_data, static_cast<T>(0), seq_tensor->numel() * sizeof(T));
CopyDataCPU<T>(seq_tensor, const_cast<framework::Tensor*>(&pad_tensor),
seq_offset, max_seq_len, seq_width, false /* seq_to_pad */,
norm_by_times, output_layout);
} }
}; };
......
...@@ -15,6 +15,7 @@ limitations under the License. */ ...@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once #pragma once
#include <algorithm> #include <algorithm>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/device_context.h"
...@@ -22,7 +23,7 @@ namespace paddle { ...@@ -22,7 +23,7 @@ namespace paddle {
namespace operators { namespace operators {
namespace math { namespace math {
enum OutputLayout { kBatchLengthWidth = 0, kLengthBatchWidth }; enum PadLayout { kBatchLengthWidth = 0, kLengthBatchWidth };
inline static size_t MaximumSequenceLength( inline static size_t MaximumSequenceLength(
const framework::Vector<size_t>& seq_offset) { const framework::Vector<size_t>& seq_offset) {
...@@ -34,35 +35,22 @@ inline static size_t MaximumSequenceLength( ...@@ -34,35 +35,22 @@ inline static size_t MaximumSequenceLength(
return max_seq_len; return max_seq_len;
} }
inline static void CheckLoD(const framework::LoDTensor& seq_tensor,
const size_t& lod_level) {
PADDLE_ENFORCE(lod_level < seq_tensor.lod().size(),
"Invalid lod level which should be at least 0 and less "
"than maximum lod level of sequence tensor.");
}
inline static void CheckDims(const framework::DDim& seq_tensor_dims, inline static void CheckDims(const framework::DDim& seq_tensor_dims,
const size_t& last_offset,
const framework::DDim& pad_tensor_dims, const framework::DDim& pad_tensor_dims,
const int64_t& max_seq_len, const int64_t& seq_num, const framework::Vector<size_t>& seq_offset,
const int64_t& seq_width, int64_t padded_seq_len, int64_t step_width,
const OutputLayout& output_layout) { const PadLayout& layout) {
PADDLE_ENFORCE_EQ(static_cast<size_t>(seq_tensor_dims[0]), last_offset, PADDLE_ENFORCE_EQ(static_cast<size_t>(seq_tensor_dims[0]), seq_offset.back(),
"Value of 1st dimension of the sequence tensor should be " "Value of 1st dimension of the sequence tensor should be "
"equal to sum of lengths of all sequences."); "equal to sum of lengths of all sequences.");
PADDLE_ENFORCE_EQ(pad_tensor_dims.size(), 3UL, PADDLE_ENFORCE(seq_tensor_dims.size() == 1 || seq_tensor_dims.size() == 2,
"Padded tensor should be a 3-D tensor."); "seq_tensor's rank should be 1 or 2.");
if (output_layout == kBatchLengthWidth) { PADDLE_ENFORCE(seq_tensor_dims.size() + 1 == pad_tensor_dims.size() ||
PADDLE_ENFORCE_EQ(pad_tensor_dims, seq_tensor_dims.size() == pad_tensor_dims.size(),
framework::make_ddim({seq_num, max_seq_len, seq_width})); "pad_tensor's rank should be 1 greater than seq_tensor's "
} else if (output_layout == kLengthBatchWidth) { "rank, or be equal with it.");
PADDLE_ENFORCE_EQ(pad_tensor_dims,
framework::make_ddim({max_seq_len, seq_num, seq_width}));
} else {
PADDLE_THROW("Unsupported output layout.");
}
} }
/* /*
...@@ -94,22 +82,22 @@ inline static void CheckDims(const framework::DDim& seq_tensor_dims, ...@@ -94,22 +82,22 @@ inline static void CheckDims(const framework::DDim& seq_tensor_dims,
template <typename DeviceContext, typename T> template <typename DeviceContext, typename T>
class PaddingLoDTensorFunctor { class PaddingLoDTensorFunctor {
public: public:
void operator()(const DeviceContext& context, void operator()(const platform::CPUDeviceContext& context,
const framework::LoDTensor& seq_tensor, const framework::LoDTensor& seq_tensor,
framework::Tensor* pad_tensor, framework::LoDTensor* pad_tensor,
T pad_value = static_cast<T>(0), bool norm_by_times = false, std::vector<T> pad_value = {0}, int pad_seq_len = -1,
size_t lod_level = 0, int lod_level = 0, bool norm_by_times = false,
OutputLayout output_layout = kBatchLengthWidth); const PadLayout layout = kBatchLengthWidth);
}; };
template <typename DeviceContext, typename T> template <typename DeviceContext, typename T>
class UnpaddingLoDTensorFunctor { class UnpaddingLoDTensorFunctor {
public: public:
void operator()(const DeviceContext& context, void operator()(const platform::CPUDeviceContext& context,
framework::LoDTensor* seq_tensor, const framework::LoDTensor& pad_tensor,
const framework::Tensor& pad_tensor, framework::LoDTensor* seq_tensor, int pad_seq_len = -1,
bool norm_by_times = false, size_t lod_level = 0, int lod_level = 0, bool norm_by_times = false,
OutputLayout output_layout = kBatchLengthWidth); const PadLayout& layout = kBatchLengthWidth);
}; };
} // namespace math } // namespace math
......
...@@ -23,7 +23,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod, ...@@ -23,7 +23,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
paddle::framework::LoDTensor cpu_seq_back; paddle::framework::LoDTensor cpu_seq_back;
paddle::framework::LoDTensor seq; paddle::framework::LoDTensor seq;
paddle::framework::LoDTensor seq_back; paddle::framework::LoDTensor seq_back;
paddle::framework::Tensor padding; paddle::framework::LoDTensor padding;
const size_t level = lod.size() - 1; const size_t level = lod.size() - 1;
auto seq_dims = auto seq_dims =
...@@ -56,13 +56,13 @@ void TestSequencePadding(const paddle::framework::LoD& lod, ...@@ -56,13 +56,13 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
padding.mutable_data<T>(padding_dims, *place); padding.mutable_data<T>(padding_dims, *place);
paddle::operators::math::PaddingLoDTensorFunctor<DeviceContext, T>()( paddle::operators::math::PaddingLoDTensorFunctor<DeviceContext, T>()(
*context, seq, &padding, 0, false, 0, *context, seq, &padding, {0}, -1, 0, false,
paddle::operators::math::kLengthBatchWidth); paddle::operators::math::kLengthBatchWidth);
seq_back.set_lod(lod); seq_back.set_lod(lod);
seq_back.mutable_data<T>(seq_dims, *place); seq_back.mutable_data<T>(seq_dims, *place);
paddle::operators::math::UnpaddingLoDTensorFunctor<DeviceContext, T>()( paddle::operators::math::UnpaddingLoDTensorFunctor<DeviceContext, T>()(
*context, &seq_back, padding, false, 0, *context, padding, &seq_back, -1, 0, false,
paddle::operators::math::kLengthBatchWidth); paddle::operators::math::kLengthBatchWidth);
if (paddle::platform::is_cpu_place(*place)) { if (paddle::platform::is_cpu_place(*place)) {
......
...@@ -153,7 +153,7 @@ class WarpCTCKernel : public framework::OpKernel<T> { ...@@ -153,7 +153,7 @@ class WarpCTCKernel : public framework::OpKernel<T> {
framework::make_ddim({static_cast<int64_t>(num_sequences), 1}); framework::make_ddim({static_cast<int64_t>(num_sequences), 1});
// warpctc needs sequences data stored in transposed padding format // warpctc needs sequences data stored in transposed padding format
Tensor warpctc_logits; LoDTensor warpctc_logits;
const size_t max_sequence_length = const size_t max_sequence_length =
math::MaximumSequenceLength(logits_lod[level]); math::MaximumSequenceLength(logits_lod[level]);
auto warpctc_logits_dims = auto warpctc_logits_dims =
...@@ -163,7 +163,7 @@ class WarpCTCKernel : public framework::OpKernel<T> { ...@@ -163,7 +163,7 @@ class WarpCTCKernel : public framework::OpKernel<T> {
warpctc_logits.mutable_data<T>(warpctc_logits_dims, ctx.GetPlace()); warpctc_logits.mutable_data<T>(warpctc_logits_dims, ctx.GetPlace());
math::PaddingLoDTensorFunctor<DeviceContext, T>()( math::PaddingLoDTensorFunctor<DeviceContext, T>()(
ctx.template device_context<DeviceContext>(), *logits, &warpctc_logits, ctx.template device_context<DeviceContext>(), *logits, &warpctc_logits,
static_cast<T>(0), false /* norm_by_times */, 0, {static_cast<T>(0)}, -1, 0, false /* norm_by_times */,
math::kLengthBatchWidth); math::kLengthBatchWidth);
const T* warpctc_logits_data = warpctc_logits.data<T>(); const T* warpctc_logits_data = warpctc_logits.data<T>();
...@@ -210,15 +210,15 @@ template <typename DeviceContext, typename T> ...@@ -210,15 +210,15 @@ template <typename DeviceContext, typename T>
class WarpCTCGradKernel : public framework::OpKernel<T> { class WarpCTCGradKernel : public framework::OpKernel<T> {
public: public:
void Compute(const framework::ExecutionContext& ctx) const override { void Compute(const framework::ExecutionContext& ctx) const override {
auto* warpctc_grad = ctx.Input<Tensor>("WarpCTCGrad"); auto* warpctc_grad = ctx.Input<LoDTensor>("WarpCTCGrad");
auto* logits_grad = ctx.Output<LoDTensor>(framework::GradVarName("Logits")); auto* logits_grad = ctx.Output<LoDTensor>(framework::GradVarName("Logits"));
const Tensor* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss")); const Tensor* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));
logits_grad->mutable_data<T>(ctx.GetPlace()); logits_grad->mutable_data<T>(ctx.GetPlace());
bool norm_by_times = ctx.Attr<bool>("norm_by_times"); bool norm_by_times = ctx.Attr<bool>("norm_by_times");
math::UnpaddingLoDTensorFunctor<DeviceContext, T>()( math::UnpaddingLoDTensorFunctor<DeviceContext, T>()(
ctx.template device_context<DeviceContext>(), logits_grad, ctx.template device_context<DeviceContext>(), *warpctc_grad,
*warpctc_grad, norm_by_times, 0, math::kLengthBatchWidth); logits_grad, -1, 0, norm_by_times, math::kLengthBatchWidth);
const T* loss_grad_data = loss_grad->data<T>(); const T* loss_grad_data = loss_grad->data<T>();
math::ScaleLoDTensorFunctor<DeviceContext, T>()( math::ScaleLoDTensorFunctor<DeviceContext, T>()(
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册