From 4c6bccbe205ee578289449c717bdc7d1feeaa7f5 Mon Sep 17 00:00:00 2001 From: chengduoZH Date: Tue, 24 Oct 2017 15:10:22 +0800 Subject: [PATCH] fix doc and remove useless code --- paddle/operators/math/CMakeLists.txt | 4 +- paddle/operators/math/sequence_project.h | 91 +++++++++++++++--------- paddle/operators/sequence_conv_op.cc | 45 ++++-------- paddle/operators/sequence_conv_op.h | 6 -- 4 files changed, 71 insertions(+), 75 deletions(-) diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt index e381545d27..2560c0a5aa 100644 --- a/paddle/operators/math/CMakeLists.txt +++ b/paddle/operators/math/CMakeLists.txt @@ -7,7 +7,7 @@ if(WITH_GPU) nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS operator) nv_library(pooling SRCS pooling.cc pooling.cu DEPS device_context) nv_library(vol2col SRCS vol2col.cc vol2col.cu DEPS device_context) - nv_library(sequence_project SRCS sequence_project.cc sequence_project.cu DEPS device_context math_function) + nv_library(sequence_project SRCS sequence_project.cc sequence_project.cu DEPS device_context) else() cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context operator) cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function) @@ -15,7 +15,7 @@ else() cc_library(cross_entropy SRCS cross_entropy.cc DEPS operator) cc_library(pooling SRCS pooling.cc DEPS device_context) cc_library(vol2col SRCS vol2col.cc DEPS device_context) - cc_library(sequence_project SRCS sequence_project.cc DEPS device_context math_function) + cc_library(sequence_project SRCS sequence_project.cc DEPS device_context) endif() cc_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) diff --git a/paddle/operators/math/sequence_project.h b/paddle/operators/math/sequence_project.h index 64a27d885d..a2ab86f790 100644 --- a/paddle/operators/math/sequence_project.h +++ b/paddle/operators/math/sequence_project.h @@ -18,7 +18,6 @@ limitations under the License. */ #include "paddle/framework/lod_tensor.h" #include "paddle/framework/tensor.h" #include "paddle/operators/math/im2col.h" -#include "paddle/operators/math/math_function.h" namespace paddle { namespace operators { @@ -32,37 +31,59 @@ template using EigenMatrix = framework::EigenMatrix; /* - * \brief Converts the feature data of four dimensions(CDHW) into a colData of - * seven dimensions in the Vol2ColFunctor calculation, - * And in the Col2VolFunctor calculation, it is reversed. + * \brief SequenceProject projects features of context_length time-steps of each + * instance. * - * \param volData Vol data. - * \param volShape The shape of volData, - * [input_channels, input_depth, input_height, input_width]. - * \param colData Column data. - * \param colShape The shape of colData. + * \param in Input data. + * \param inShape The shape of Input data, + * [minibatch, number_of_input_features]. + * \param inShape A float LoDTensor. * - * The shape of colData is: - * [input_channels, filter_depth, filter_height, filter_width, output_depth, - * output_height, output_width] - * So, it is easy to reshape into a convolution matrix for convolution - * calculation based on matrix multiplication. - * The shape of convolution matrix is [height, width], where the height is equal - * input_channels * filter_depth * filter_height * filter_width, and the width - * is equal output_depth * output_height * output_width. + * \param padding_data Padding data. + * \param inShape The shape of Padding data, + * [up_pad + down_pad, number_of_input_features]. + * \param inShape A float LoDTensor. * - * Reshape: - * shape of colData shape of convolution matrix - * [input_channels, - * filter_depth, - * filter_height, - * filter_width, ======> [height, width] - * output_depth, - * output_height, - * output_width] + * \param col Col data. + * \param inShape The shape of Col data, + * [minibatch, 1]. + * \param inShape A float LoDTensor. + * + * For a mini-batch of 2 variable lengths sentences, containing 3, and 1 + * time-steps: + * + * Assumed input (X) is a [4, M, N] float LoDTensor, and X->lod()[0] = [0, 3, + * 4]. + * Besides, for the sake of simplicity, we assume M=1 and N=2. + * + * X = [[a1, a2; + * b1, b2; + * c1, c2] + * [d1, d2]] + * + * This is to say that input (X) has 4 words and the dimension of each word + * representation is 2. + * + * - Case1: + * If context_start is -1 and padding_trainable is false, we use zero to pad + * instead of learned weight to pad, + * and the context_lenth is 3, the output (Out) is: + * + * Out =[[0, 0, a1, a2, b1, b2; + * a1, a2, b1, b2, c1, c2; + * b1, b2, c1, c2, 0, 0 ] + * [0, 0, d1, d2, 0, 0 ]] + * + * - Case2: + * If context_start is -1 and padding_trainable is true, we use learned weight + * to pad, + * and the context_lenth is 3, the output (Out) is: + * + * Out = [[w1, w2, a1, a2, b1, b2; + * a1, a2, b1, b2, c1, c2; + * b1, b2, c1, c2, w3, w4] + * [w1, w2, d1, d2, w3, w4]] * - * \note The caller needs to ensure that volShape.inputChannels is equal to - * colShape.inputChannels. */ template @@ -96,14 +117,16 @@ class SequenceProjectFunctor { sequence_height = static_cast(out_t.dims()[0]); - std::vector output_shape( - {sequence_height, 1, 1, context_length, - sequence_width}); // output_height, output_width, - // input_channels, filter_height, filter_width - out_t.Resize(framework::make_ddim(output_shape)); - if (input_row_begin < input_row_end) { framework::Tensor in_t = in->Slice(input_row_begin, input_row_end); + + std::vector output_shape( + {sequence_height, 1, 1, context_length, + sequence_width}); // output_height, output_width, + // input_channels, filter_height, filter_width + + out_t.Resize(framework::make_ddim(output_shape)); + std::vector input_shape( {1, input_row_end - input_row_begin, sequence_width}); // input_channels, input_height, input_width diff --git a/paddle/operators/sequence_conv_op.cc b/paddle/operators/sequence_conv_op.cc index 1fc23302dc..d286d334a2 100644 --- a/paddle/operators/sequence_conv_op.cc +++ b/paddle/operators/sequence_conv_op.cc @@ -135,39 +135,18 @@ class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker { .GreaterThan(0); AddComment(R"DOC( - SequenceConvOp projects features of context_length time-steps of each instance. - - For a mini-batch of 2 variable lengths sentences, containing 3, and 1 time-steps: - - Assumed input (X) is a [4, M, N] float LoDTensor, and X->lod()[0] = [0, 3, 4]. - Besides, for the sake of simplicity, we assume M=1 and N=2. - - X = [[a1, a2; - b1, b2; - c1, c2] - [d1, d2]] - - This is to say that input (X) has 4 words and the dimension of each word - representation is 2. - - - Case1: - If context_start is -1 and padding_trainable is false, we use zero to pad instead of learned weight to pad, - and the context_lenth is 3, the output (Out) is: - - Out =[[0, 0, a1, a2, b1, b2; - a1, a2, b1, b2, c1, c2; - b1, b2, c1, c2, 0, 0 ] - [0, 0, d1, d2, 0, 0 ]] - - - Case2: - If context_start is -1 and padding_trainable is true, we use learned weight to pad, - and the context_lenth is 3, the output (Out) is: - - Out = [[w1, w2, a1, a2, b1, b2; - a1, a2, b1, b2, c1, c2; - b1, b2, c1, c2, w3, w4] - [w1, w2, d1, d2, w3, w4]] - + SequenceConvOp performs convolution operation on features of + context_length time-steps of each instance. + The convolution operation calculates the output based on the input, filter + and strides, paddings parameters. The size of each dimension of the + parameters is checked in the infer-shape. + +Example: + Input: + X shape: (minibatch, number_of_input_features) + Filter shape: (context_length, number_of_input_features) + Output: + Out shape: (minibatch, 1) )DOC"); } }; diff --git a/paddle/operators/sequence_conv_op.h b/paddle/operators/sequence_conv_op.h index a8bda2f046..b6ae12f6bb 100644 --- a/paddle/operators/sequence_conv_op.h +++ b/paddle/operators/sequence_conv_op.h @@ -182,12 +182,6 @@ class SequenceConvGradKernel : public framework::OpKernel { functor(context.device_context(), padding_data_g, 0); for (int i = 0; i < static_cast(lod_g_level_0.size()) - 1; ++i) { - input_row_begin = - (context_start > 0) - ? static_cast(lod_g_level_0[i]) + context_start - : static_cast(lod_g_level_0[i]); - input_row_end = static_cast(lod_g_level_0[i + 1]); - Tensor col_t = col.Slice(static_cast(lod_g_level_0[i]), static_cast(lod_g_level_0[i + 1])); -- GitLab