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

Merge pull request #5130 from chengduoZH/fix_sequence_conv_op

Browse Source 
fix_sequence_conv_op
fix-typo
chengduo 7 years ago committed by GitHub
parent 03136f6375 b08ae0b1dc
commit 0f9858a1c9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 228 additions and 151 deletions
Show Stats Download Patch File Download Diff File

199
paddle/operators/math/context_project.h
View File

File diff suppressed because it is too large Load Diff

64
paddle/operators/sequence_conv_op.cc
Unescape View File

@ -30,19 +30,20 @@ class SequenceConvOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE(ctx->HasOutput("Out"), PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of SequenceConvOp should not be null."); "Output(Out) of SequenceConvOp should not be null.");
int context_length = ctx->Attrs().Get<int>("context_length"); int context_length = ctx->Attrs().Get<int>("contextLength");
bool padding_trainable = ctx->Attrs().Get<bool>("padding_trainable"); int context_start = ctx->Attrs().Get<int>("contextStart");
int context_start = ctx->Attrs().Get<int>("context_start");
auto in_dims = ctx->GetInputDim("X"); auto in_dims = ctx->GetInputDim("X");
auto filter_dims = ctx->GetInputDim("Filter"); auto filter_dims = ctx->GetInputDim("Filter");
PADDLE_ENFORCE(ctx->Attrs().Get<int>("contextStride") == 1,
"Currently, SequenceConvOp only supports contextStride=1.");
PADDLE_ENFORCE(in_dims.size() == 2 && filter_dims.size() == 2, PADDLE_ENFORCE(in_dims.size() == 2 && filter_dims.size() == 2,
"Input(X, Filter) should be 2-D tensor."); "Input(X, Filter) should be 2-D tensor.");
PADDLE_ENFORCE(filter_dims[0] == context_length * in_dims[1], PADDLE_ENFORCE(filter_dims[0] == context_length * in_dims[1],
"Filter's height should be context_length * " "Filter's height should be context_length * "
"number_of_input_features ."); "input_hidden_size .");
if (padding_trainable) { if (ctx->Attrs().Get<bool>("paddingTrainable")) {
PADDLE_ENFORCE( PADDLE_ENFORCE(
ctx->HasInput("PaddingData"), ctx->HasInput("PaddingData"),
"Input(PaddingData) of SequenceConvOp should not be null."); "Input(PaddingData) of SequenceConvOp should not be null.");
@ -54,7 +55,7 @@ class SequenceConvOp : public framework::OperatorWithKernel {
if (context_start == 0 && context_length == 1) { if (context_start == 0 && context_length == 1) {
PADDLE_THROW( PADDLE_THROW(
"If context_start is 0 and context_length is 1, padding_trainable " "If context_start is 0 and context_length is 1, paddingTrainable "
"should be false."); "should be false.");
} }
PADDLE_ENFORCE(padding_dim.size() == 2, PADDLE_ENFORCE(padding_dim.size() == 2,
@ -81,13 +82,14 @@ class SequenceConvGradOp : public framework::OperatorWithKernel {
"Gradient of output(Out) should not be null."); "Gradient of output(Out) should not be null.");
PADDLE_ENFORCE(ctx->HasInput("X"), "The input(X) should not be null."); PADDLE_ENFORCE(ctx->HasInput("X"), "The input(X) should not be null.");
if (ctx->Attrs().Get<bool>("padding_trainable") && if (ctx->Attrs().Get<bool>("paddingTrainable") &&
ctx->HasOutput(framework::GradVarName("PaddingData"))) { ctx->HasOutput(framework::GradVarName("PaddingData"))) {
ctx->SetOutputDim(framework::GradVarName("PaddingData"), ctx->SetOutputDim(framework::GradVarName("PaddingData"),
ctx->GetInputDim("PaddingData")); ctx->GetInputDim("PaddingData"));
} }
if (ctx->HasOutput(framework::GradVarName("X"))) { if (ctx->HasOutput(framework::GradVarName("X"))) {
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
ctx->ShareLoD(framework::GradVarName("X"), "X");
} }
if (ctx->HasOutput(framework::GradVarName("Filter"))) { if (ctx->HasOutput(framework::GradVarName("Filter"))) {
ctx->SetOutputDim(framework::GradVarName("Filter"), ctx->SetOutputDim(framework::GradVarName("Filter"),
@ -105,54 +107,58 @@ class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker {
"X", "X",
"(LoDTensor) the input(X) is a LodTensor, which support " "(LoDTensor) the input(X) is a LodTensor, which support "
"variable-time length input sequence. The underlying tensor in " "variable-time length input sequence. The underlying tensor in "
"this LoDTensor is a matrix with shape (T, D), where, T is the " "this LoDTensor is a matrix with shape (T, N), where, T is the "
"total time steps in this mini-batch, D is the input feature size."); "total time steps in this mini-batch, N is the input_hidden_size.");
AddInput("PaddingData", AddInput("PaddingData",
"(Tensor, optional) the input(PaddingData) is an optional " "(Tensor, optional) the input(PaddingData) is an optional "
"parameter, and it is learnable. " "parameter, and it is learnable. "
"This is a tensor with shape (N, D), where N is the " "This is a tensor with shape (P, N), where P is the "
"top_pad + bottom_pad, D is the input feature size. In order to " "top_pad + bottom_pad, N is the input_hidden_size. In order to "
"ensure the equal length of sequence before and after " "ensure the equal length of sequence before and after "
"convolution, it is necessary to fill the top and bottom of each " "convolution, it is necessary to fill the top and bottom of each "
"sequence according to context_length, context_stride and " "sequence according to context_length, context_stride and "
"context_start") "context_start")
.AsDispensable(); .AsDispensable();
AddInput("Filter", AddInput(
"Filter",
"(Tensor) the input(Filter) is an learnable parameter." "(Tensor) the input(Filter) is an learnable parameter."
"This is a tensor with shape (N, D), where N is the " "This is a tensor with shape (K, M), where K is the "
"context_length, D is the output feature size."); "context_length * input_hidden_size, M is the output feature size.");
AddOutput( AddOutput(
"Out", "Out",
"(LoDTensor) the output(Out) is a LodTensor, which support " "(LoDTensor) the output(Out) is a LodTensor, which support "
"variable-time length output sequence. The underlying tensor in " "variable-time length output sequence. The underlying tensor in "
"this LoDTensor is a matrix with shape (T, D), where, T is the " "this LoDTensor is a matrix with shape (T, M), where, T is the "
"total time steps in this mini-batch, D is the output feature size."); "total time steps in this mini-batch, M is the output feature size.");
AddAttr<bool>("padding_trainable", AddAttr<bool>("paddingTrainable",
"(bool, default false) the padding data of SequenceConvOp " "(bool, default:false) the padding data of SequenceConvOp "
"is trainable or not.") "is trainable or not.")
.SetDefault(false); .SetDefault(false);
AddAttr<int>("context_length", AddAttr<int>("contextLength",
"(int, default 3) the context_length of SequenceConvOp is the " "(int) the contextLength of SequenceConvOp is the "
"height of the convolution kernel.") "height of the convolution kernel.")
.SetDefault(3)
.GreaterThan(0); .GreaterThan(0);
AddAttr<int>("context_start", AddAttr<int>("contextStart",
"(int, default 0) the context_start of SequenceConvOp " "(int, default:0) the contextStart of SequenceConvOp "
"represents the beginning of the convolution of the number of " "represents the beginning of the convolution of the number of "
"rows of sequence, which can be negative.") "rows of sequence, which can be negative. The negative number "
"means to pad contextStart time-steps of zeros or learnable "
"parameters at the beginning of each instance. The positive "
"number means to skip contextStart time-steps of each "
"instance.")
.SetDefault(0); .SetDefault(0);
AddAttr<int>("context_stride", AddAttr<int>("contextStride",
"(int, default 1) the context_stride of SequenceConvOp " "(int, default:1) the contextStride of SequenceConvOp "
"represents the step length of convolution. " "represents the stride length of convolution kernel. "
"Currently, SequenceConvOp only supports" "Currently, SequenceConvOp only supports"
"context_stride=1.") "contextStride=1.")
.SetDefault(1) .SetDefault(1)
.GreaterThan(0); .GreaterThan(0);
AddComment(R"DOC( AddComment(R"DOC(
SequenceConvOp performs convolution operation on features of SequenceConvOp performs convolution operation on features of
context_length time-steps of each instance. contextLength time-steps of each instance.
The convolution operation calculates the output based on the input, filter The convolution operation calculates the output based on the input, filter
and strides, paddings parameters. The size of each dimension of the and strides, paddings parameters. The size of each dimension of the
parameters is checked in the infer-shape. In order to ensure the equal parameters is checked in the infer-shape. In order to ensure the equal

64
paddle/operators/sequence_conv_op.h
Unescape View File

@ -35,12 +35,11 @@ class SequenceConvKernel : public framework::OpKernel<T> {
out->mutable_data<T>(context.GetPlace()); out->mutable_data<T>(context.GetPlace());
context.ShareLoD("X", "Out"); context.ShareLoD("X", "Out");
int context_start = context.Attr<int>("context_start"); int context_start = context.Attr<int>("contextStart");
int context_length = context.Attr<int>("context_length"); int context_length = context.Attr<int>("contextLength");
int context_stride = context.Attr<int>("context_stride"); int context_stride = context.Attr<int>("contextStride");
bool padding_trainable = context.Attr<bool>("padding_trainable"); bool padding_trainable = context.Attr<bool>("paddingTrainable");
// InferShape by in_lod
PADDLE_ENFORCE_EQ(in->lod().size(), 1UL, PADDLE_ENFORCE_EQ(in->lod().size(), 1UL,
"Only support one level sequence now."); "Only support one level sequence now.");
@ -51,26 +50,21 @@ class SequenceConvKernel : public framework::OpKernel<T> {
int up_pad = std::max(0, -context_start); int up_pad = std::max(0, -context_start);
int down_pad = std::max(0, context_start + context_length - 1); int down_pad = std::max(0, context_start + context_length - 1);
int sequence_width; int sequence_width = static_cast<int>(in->dims()[1]);
sequence_width = static_cast<int>(in->dims()[1]);
// Use col_shape in the im2col calculation.
framework::DDim col_shape = {in->dims()[0], framework::DDim col_shape = {in->dims()[0],
sequence_width * context_length}; context_length * sequence_width};
Tensor col; Tensor col;
col.mutable_data<T>(col_shape, context.GetPlace()); col.mutable_data<T>(col_shape, context.GetPlace());
math::SetConstant<Place, T> set_zero;
// Because if padding_trainable is false, padding data should be zeros. // Because if padding_trainable is false, padding data should be zeros.
math::SetConstant<Place, T> set_zero;
set_zero(context.device_context(), &col, static_cast<T>(0)); set_zero(context.device_context(), &col, static_cast<T>(0));
paddle::operators::math::ContextProjectFunctor<Place, T> math::ContextProjectFunctor<Place, T> seq_project_functor;
seq_project_functor;
LoDTensor* input = const_cast<LoDTensor*>(in);
Tensor* pad_data = const_cast<Tensor*>(padding_data);
seq_project_functor(context.device_context(), *input, *pad_data, col, seq_project_functor(context.device_context(), *in, *padding_data, col,
padding_trainable, context_start, context_length, padding_trainable, context_start, context_length,
context_stride, up_pad, down_pad, false, false, false); context_stride, up_pad, down_pad);
math::matmul<Place, T>(context.device_context(), col, false, filter, false, math::matmul<Place, T>(context.device_context(), col, false, filter, false,
static_cast<T>(1.0), out, static_cast<T>(0.0)); static_cast<T>(1.0), out, static_cast<T>(0.0));
@ -81,18 +75,18 @@ template <typename Place, typename T>
class SequenceConvGradKernel : public framework::OpKernel<T> { class SequenceConvGradKernel : public framework::OpKernel<T> {
public: public:
void Compute(const framework::ExecutionContext& context) const override { void Compute(const framework::ExecutionContext& context) const override {
auto* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X")); auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
auto* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
auto* filter_g = context.Output<Tensor>(framework::GradVarName("Filter")); auto* filter_g = context.Output<Tensor>(framework::GradVarName("Filter"));
auto* padding_data_g = auto* padding_data_g =
context.Output<Tensor>(framework::GradVarName("PaddingData")); context.Output<Tensor>(framework::GradVarName("PaddingData"));
auto* in = context.Input<LoDTensor>("X"); auto* in = context.Input<LoDTensor>("X");
auto* filter = context.Input<Tensor>("Filter"); auto* filter = context.Input<Tensor>("Filter");
int context_start = context.Attr<int>("context_start"); int context_start = context.Attr<int>("contextStart");
int context_length = context.Attr<int>("context_length"); int context_length = context.Attr<int>("contextLength");
int context_stride = context.Attr<int>("context_stride"); int context_stride = context.Attr<int>("contextStride");
bool padding_trainable = context.Attr<bool>("padding_trainable"); bool padding_trainable = context.Attr<bool>("paddingTrainable");
PADDLE_ENFORCE_EQ(in->lod().size(), 1UL, PADDLE_ENFORCE_EQ(in->lod().size(), 1UL,
"Only support one level sequence now."); "Only support one level sequence now.");
@ -115,17 +109,18 @@ class SequenceConvGradKernel : public framework::OpKernel<T> {
math::matmul<Place, T>(context.device_context(), *out_g, false, *filter, math::matmul<Place, T>(context.device_context(), *out_g, false, *filter,
true, T(1.0), &col, T(1.0)); true, T(1.0), &col, T(1.0));
} }
paddle::operators::math::ContextProjectFunctor<Place, T> math::ContextProjectFunctor<Place, T> seq_project_functor;
seq_project_functor; math::ContextProjectGradFunctor<Place, T> seq_project_grad_functor;
if (in_g) { if (in_g) {
in_g->mutable_data<T>(context.GetPlace()); in_g->mutable_data<T>(context.GetPlace());
in_g->set_lod(in->lod()); in_g->set_lod(in->lod());
set_zero(context.device_context(), in_g, static_cast<T>(0)); set_zero(context.device_context(), in_g, static_cast<T>(0));
seq_project_functor(context.device_context(), *in_g, *padding_data_g, col, seq_project_grad_functor(context.device_context(), *in_g, *padding_data_g,
padding_trainable, context_start, context_length, col, padding_trainable, context_start,
context_stride, up_pad, down_pad, true, true, false); context_length, context_stride, up_pad, down_pad,
true, false);
} }
if (padding_trainable && padding_data_g) { if (padding_trainable && padding_data_g) {
@ -133,9 +128,10 @@ class SequenceConvGradKernel : public framework::OpKernel<T> {
set_zero(context.device_context(), padding_data_g, static_cast<T>(0)); set_zero(context.device_context(), padding_data_g, static_cast<T>(0));
LoDTensor* input = const_cast<LoDTensor*>(in); LoDTensor* input = const_cast<LoDTensor*>(in);
seq_project_functor(context.device_context(), *input, *padding_data_g, seq_project_grad_functor(context.device_context(), *input,
col, padding_trainable, context_start, context_length, *padding_data_g, col, padding_trainable,
context_stride, up_pad, down_pad, true, false, true); context_start, context_length, context_stride,
up_pad, down_pad, false, true);
} }
if (filter_g) { if (filter_g) {
@ -150,15 +146,9 @@ class SequenceConvGradKernel : public framework::OpKernel<T> {
padding_data = context.Input<Tensor>("PaddingData"); padding_data = context.Input<Tensor>("PaddingData");
} }
sequence_width = static_cast<int>(in->dims()[1]); seq_project_functor(context.device_context(), *in, *padding_data, col,
LoDTensor* input = const_cast<LoDTensor*>(in);
Tensor* pad_data = const_cast<Tensor*>(padding_data);
seq_project_functor(context.device_context(), *input, *pad_data, col,
padding_trainable, context_start, context_length, padding_trainable, context_start, context_length,
context_stride, up_pad, down_pad, false, false, context_stride, up_pad, down_pad);
false);
math::matmul<Place, T>(context.device_context(), col, true, out_grad, math::matmul<Place, T>(context.device_context(), col, true, out_grad,
false, T(1.0), &filter_grad, T(1.0)); false, T(1.0), &filter_grad, T(1.0));

8
python/paddle/v2/framework/tests/test_seq_conv.py
Unescape View File

@ -45,10 +45,10 @@ class TestSeqProject(OpTest):
self.inputs_val_no_f = ['PaddingData', 'X'] self.inputs_val_no_f = ['PaddingData', 'X']
self.attrs = { self.attrs = {
'context_start': self.context_start, 'contextStart': self.context_start,
'context_length': self.context_length, 'contextLength': self.context_length,
'padding_trainable': self.padding_trainable, 'paddingTrainable': self.padding_trainable,
'context_stride': self.context_stride 'contextStride': self.context_stride
} }
out = np.zeros( out = np.zeros(
(self.input_size[0], self.output_represention)).astype('float32') (self.input_size[0], self.output_represention)).astype('float32')

Write Preview
Loading…
Cancel
Save