Merge pull request #5130 from chengduoZH/fix_sequence_conv_op

fix_sequence_conv_op

paddle/operators/math/context_project.h

@@ -16,36 +16,36 @@ limitations under the License. */
 
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/lod_tensor.h"
-#include "paddle/framework/tensor.h"
 #include "paddle/operators/math/im2col.h"
 
 namespace paddle {
 namespace operators {
 namespace math {
 
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
 /*
- * \brief Context projection concatenate features in adjacent time steps in
+ * \brief Context projection concatenates features in adjacent time-steps in
  * a sequence. The i-th row of the output is the concatenation of
  * context_length rows of the input. The context_length rows are the
  * consecutive rows from the i+shift_start row.
-
+ * ContextProjectGradFunctor is the inverse process of ContextProjectFunctor.
+ *
  * \param in            Input data.
- * \param Shape         The shape of Input data,
- *                      [minibatch, number_of_input_features].
- * \param type          A float LoDTensor.
+ * \param Shape         The shape of Input data:
+ *                        [mini-batch, input_hidden_size].
  *
  * \param padding_data  Padding data.
- * \param Shape         The shape of Padding data,
- *                      [up_pad + down_pad, number_of_input_features].
- * \param type          A float Tensor.
+ * \param Shape         The shape of Padding data:
+ *                        [up_pad + down_pad, input_hidden_size].
  *
  * \param col           Col data.
- * \param Shape         The shape of Col data,
- *                      [minibatch, context_length * number_of_input_features].
- * \param type           A float Tensor.
+ * \param Shape         The shape of Col data:
+ *                        [mini-batch, context_length * input_hidden_size].
  *
  * For a mini-batch of 2 variable lengths sentences, containing 3, and 1
  * time-steps:
@@ -65,7 +65,7 @@ using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
  * - 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:
+ *   and the context_length is 3, the output (Out) is:
  *
  *   Out =[[0,  0,  a1, a2, b1, b2;
  *          a1, a2, b1, b2, c1, c2;
@@ -75,7 +75,7 @@ using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
  * - 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:
+ *   and the context_length is 3, the output (Out) is:
  *
  *   Out = [[w1, w2, a1, a2, b1, b2;
  *           a1, a2, b1, b2, c1, c2;
@@ -87,48 +87,146 @@ using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 template <typename Place, typename T>
 class ContextProjectFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  framework::LoDTensor& in, framework::Tensor& padding_data,
-                  framework::Tensor& col, bool padding_trainable,
-                  int context_start, int context_length, int context_stride,
-                  int up_pad, int down_pad, bool gradient, bool input_grad,
-                  bool pad_grad) {
+  void operator()(const platform::DeviceContext& context, const LoDTensor& in,
+                  const Tensor& padding_data, Tensor& col,
+                  bool padding_trainable, int context_start, int context_length,
+                  int context_stride, int up_pad, int down_pad) {
     auto lod_level_0 = in.lod()[0];
 
-    paddle::operators::math::Im2ColFunctor<
-        paddle::operators::math::ColFormat::kOCF, Place, float>
-        im2col_ocf;
-    paddle::operators::math::Col2ImFunctor<
-        paddle::operators::math::ColFormat::kOCF, Place, float>
-        col2im_ocf;
+    math::Im2ColFunctor<math::ColFormat::kOCF, Place, float> im2col_ocf;
 
     int input_row_begin, input_row_end;
     int sequence_height, sequence_width;
     sequence_width = in.dims()[1];
-    input_grad = gradient && input_grad;
-    pad_grad = gradient && pad_grad;
 
-    if (!gradient || input_grad) {
     for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
       input_row_begin = (context_start > 0)
                             ? static_cast<int>(lod_level_0[i]) + context_start
                             : static_cast<int>(lod_level_0[i]);
       input_row_end = static_cast<int>(lod_level_0[i + 1]);
 
-        framework::Tensor out_t =
-            col.Slice(static_cast<int>(lod_level_0[i]),
+      Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
                                static_cast<int>(lod_level_0[i + 1]));
 
       sequence_height = static_cast<int>(out_t.dims()[0]);
 
       if (input_row_begin < input_row_end) {
-          framework::Tensor in_t = in.Slice(input_row_begin, input_row_end);
+        Tensor in_t = in.Slice(input_row_begin, input_row_end);
 
         std::vector<int64_t> 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<int64_t> input_shape(
+            {1, input_row_end - input_row_begin,
+             sequence_width});  // input_channels, input_height, input_width
+        in_t.Resize(framework::make_ddim(input_shape));
+
+        im2col_ocf(context, in_t, out_t,
+                   /*stride_height*/ context_stride, /*stride_width*/ 1, up_pad,
+                   down_pad, 0, 0);
+        out_t.Resize({sequence_height, context_length * sequence_width});
+      }
+    }
+    if (padding_trainable) {
+      for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
+        Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
+                                 static_cast<int>(lod_level_0[i + 1]));
+
+        sequence_height = static_cast<int>(out_t.dims()[0]);
+
+        // add up trainable data
+        out_t.Resize({sequence_height * context_length, sequence_width});
+
+        if (up_pad > 0) {  // add up pad
+          int padding_rows = std::min(
+              up_pad, static_cast<int>(lod_level_0[i + 1] - lod_level_0[i]));
+
+          for (int k = 0; k < padding_rows; ++k) {
+            int padding_size =
+                k + context_length < up_pad ? context_length : up_pad - k;
+            Tensor out_t_sub = out_t.Slice(k * context_length,
+                                           k * context_length + padding_size);
+            Tensor w_sub = padding_data.Slice(k, k + padding_size);
+            auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
+            auto w_sub_e = EigenMatrix<T>::From(w_sub);
+            out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
+          }
+        }
+        if (down_pad > 0) {  // add down pad
+          int down_pad_begin_row =
+              std::max(0,
+                       (sequence_height - context_start - context_length) + 1) +
+              1;
+          int padding_begin = std::max(0, context_start - sequence_height);
+          int padding_size =
+              sequence_height - context_start >= context_length
+                  ? 1
+                  : context_length - (sequence_height - context_start);
+          if (context_start >= sequence_height) padding_size = context_length;
+          int padding_idx = padding_begin;
+          for (int t = 0; t + down_pad_begin_row <= sequence_height;
+               ++t, ++padding_size) {
+            if (context_start >= sequence_height) padding_size = context_length;
+            if (padding_size > context_length) {
+              padding_size = context_length;
+              padding_idx++;
+            }
+            if (padding_begin > 0 || sequence_height == context_start)
+              padding_idx = padding_begin + t;
+
+            Tensor out_t_sub = out_t.Slice(
+                (down_pad_begin_row + t) * context_length - padding_size,
+                (down_pad_begin_row + t) * context_length);
+            Tensor w_sub = padding_data.Slice(
+                up_pad + padding_idx, up_pad + padding_idx + padding_size);
+            auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
+            auto w_sub_e = EigenMatrix<T>::From(w_sub);
+            out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
+          }
+        }
+        out_t.Resize({sequence_height, context_length * sequence_width});
+      }
+    }
+  }
+};
+
+template <typename Place, typename T>
+class ContextProjectGradFunctor {
+ public:
+  void operator()(const platform::DeviceContext& context, LoDTensor& in,
+                  Tensor& padding_data, Tensor& col, bool padding_trainable,
+                  int context_start, int context_length, int context_stride,
+                  int up_pad, int down_pad, bool input_grad, bool pad_grad) {
+    auto lod_level_0 = in.lod()[0];
+
+    math::Col2ImFunctor<math::ColFormat::kOCF, Place, float> col2im_ocf;
+
+    int input_row_begin, input_row_end;
+    int sequence_height, sequence_width;
+    sequence_width = in.dims()[1];
+
+    if (input_grad) {
+      for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
+        input_row_begin = (context_start > 0)
+                              ? static_cast<int>(lod_level_0[i]) + context_start
+                              : static_cast<int>(lod_level_0[i]);
+        input_row_end = static_cast<int>(lod_level_0[i + 1]);
+
+        Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
+                                 static_cast<int>(lod_level_0[i + 1]));
+
+        sequence_height = static_cast<int>(out_t.dims()[0]);
+
+        if (input_row_begin < input_row_end) {
+          Tensor in_t = in.Slice(input_row_begin, input_row_end);
 
+          std::vector<int64_t> 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<int64_t> input_shape(
@@ -136,53 +234,39 @@ class ContextProjectFunctor {
                sequence_width});  // input_channels, input_height, input_width
           in_t.Resize(framework::make_ddim(input_shape));
 
-          if (gradient) {
           col2im_ocf(context, in_t, out_t,
                      /*stride_height*/ context_stride, /*stride_width*/ 1,
                      up_pad, down_pad, 0, 0);
-          } else {
-            im2col_ocf(context, in_t, out_t,
-                       /*stride_height*/ context_stride, /*stride_width*/ 1,
-                       up_pad, down_pad, 0, 0);
-          }
           out_t.Resize({sequence_height, context_length * sequence_width});
         }
       }
     }
-    if (!gradient || pad_grad) {
+    if (pad_grad) {
       if (padding_trainable) {
         for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
-          framework::Tensor out_t =
-              col.Slice(static_cast<int>(lod_level_0[i]),
+          Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
                                    static_cast<int>(lod_level_0[i + 1]));
 
           sequence_height = static_cast<int>(out_t.dims()[0]);
-
-          // add up trainable data
           out_t.Resize({sequence_height * context_length, sequence_width});
 
-          if (up_pad > 0) {  // add up pad
+          if (up_pad > 0) {
             int padding_rows = std::min(
                 up_pad, static_cast<int>(lod_level_0[i + 1] - lod_level_0[i]));
 
             for (int k = 0; k < padding_rows; ++k) {
               int padding_size =
                   k + context_length < up_pad ? context_length : up_pad - k;
-              framework::Tensor out_t_sub = out_t.Slice(
-                  k * context_length, k * context_length + padding_size);
-              framework::Tensor w_sub = padding_data.Slice(k, k + padding_size);
-              // in this block, using EigenVector<T>::Flatten is ok too.
+              Tensor out_t_sub = out_t.Slice(k * context_length,
+                                             k * context_length + padding_size);
+              Tensor w_sub = padding_data.Slice(k, k + padding_size);
               auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
               auto w_sub_e = EigenMatrix<T>::From(w_sub);
-              if (gradient) {
               w_sub_e.device(*context.GetEigenDevice<Place>()) =
                   w_sub_e + out_t_sub_e;
-              } else {
-                out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
-              }
             }
           }
-          if (down_pad > 0) {  // add down pad
+          if (down_pad > 0) {
             int down_pad_begin_row =
                 std::max(
                     0, (sequence_height - context_start - context_length) + 1) +
@@ -204,19 +288,16 @@ class ContextProjectFunctor {
               }
               if (padding_begin > 0 || sequence_height == context_start)
                 padding_idx = padding_begin + t;
-              framework::Tensor out_t_sub = out_t.Slice(
+
+              Tensor out_t_sub = out_t.Slice(
                   (down_pad_begin_row + t) * context_length - padding_size,
                   (down_pad_begin_row + t) * context_length);
-              framework::Tensor w_sub = padding_data.Slice(
+              Tensor w_sub = padding_data.Slice(
                   up_pad + padding_idx, up_pad + padding_idx + padding_size);
               auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
               auto w_sub_e = EigenMatrix<T>::From(w_sub);
-              if (gradient) {
               w_sub_e.device(*context.GetEigenDevice<Place>()) =
                   w_sub_e + out_t_sub_e;
-              } else {
-                out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
-              }
             }
           }
           out_t.Resize({sequence_height, context_length * sequence_width});

paddle/operators/sequence_conv_op.cc

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

paddle/operators/sequence_conv_op.h

@@ -35,12 +35,11 @@ class SequenceConvKernel : public framework::OpKernel<T> {
     out->mutable_data<T>(context.GetPlace());
     context.ShareLoD("X", "Out");
 
-    int context_start = context.Attr<int>("context_start");
-    int context_length = context.Attr<int>("context_length");
-    int context_stride = context.Attr<int>("context_stride");
-    bool padding_trainable = context.Attr<bool>("padding_trainable");
+    int context_start = context.Attr<int>("contextStart");
+    int context_length = context.Attr<int>("contextLength");
+    int context_stride = context.Attr<int>("contextStride");
+    bool padding_trainable = context.Attr<bool>("paddingTrainable");
 
-    // InferShape by in_lod
     PADDLE_ENFORCE_EQ(in->lod().size(), 1UL,
                       "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 down_pad = std::max(0, context_start + context_length - 1);
-    int sequence_width;
-    sequence_width = static_cast<int>(in->dims()[1]);
+    int sequence_width = static_cast<int>(in->dims()[1]);
 
-    // Use col_shape in the im2col calculation.
     framework::DDim col_shape = {in->dims()[0],
-                                 sequence_width * context_length};
+                                 context_length * sequence_width};
     Tensor col;
     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.
+    math::SetConstant<Place, T> set_zero;
     set_zero(context.device_context(), &col, static_cast<T>(0));
 
-    paddle::operators::math::ContextProjectFunctor<Place, T>
-        seq_project_functor;
-    LoDTensor* input = const_cast<LoDTensor*>(in);
-    Tensor* pad_data = const_cast<Tensor*>(padding_data);
+    math::ContextProjectFunctor<Place, T> seq_project_functor;
 
-    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,
-                        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,
                            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> {
  public:
   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* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
     auto* filter_g = context.Output<Tensor>(framework::GradVarName("Filter"));
     auto* padding_data_g =
         context.Output<Tensor>(framework::GradVarName("PaddingData"));
     auto* in = context.Input<LoDTensor>("X");
     auto* filter = context.Input<Tensor>("Filter");
 
-    int context_start = context.Attr<int>("context_start");
-    int context_length = context.Attr<int>("context_length");
-    int context_stride = context.Attr<int>("context_stride");
-    bool padding_trainable = context.Attr<bool>("padding_trainable");
+    int context_start = context.Attr<int>("contextStart");
+    int context_length = context.Attr<int>("contextLength");
+    int context_stride = context.Attr<int>("contextStride");
+    bool padding_trainable = context.Attr<bool>("paddingTrainable");
 
     PADDLE_ENFORCE_EQ(in->lod().size(), 1UL,
                       "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,
                              true, T(1.0), &col, T(1.0));
     }
-    paddle::operators::math::ContextProjectFunctor<Place, T>
-        seq_project_functor;
+    math::ContextProjectFunctor<Place, T> seq_project_functor;
+    math::ContextProjectGradFunctor<Place, T> seq_project_grad_functor;
 
     if (in_g) {
       in_g->mutable_data<T>(context.GetPlace());
       in_g->set_lod(in->lod());
       set_zero(context.device_context(), in_g, static_cast<T>(0));
 
-      seq_project_functor(context.device_context(), *in_g, *padding_data_g, col,
-                          padding_trainable, context_start, context_length,
-                          context_stride, up_pad, down_pad, true, true, false);
+      seq_project_grad_functor(context.device_context(), *in_g, *padding_data_g,
+                               col, padding_trainable, context_start,
+                               context_length, context_stride, up_pad, down_pad,
+                               true, false);
     }
 
     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));
 
       LoDTensor* input = const_cast<LoDTensor*>(in);
-      seq_project_functor(context.device_context(), *input, *padding_data_g,
-                          col, padding_trainable, context_start, context_length,
-                          context_stride, up_pad, down_pad, true, false, true);
+      seq_project_grad_functor(context.device_context(), *input,
+                               *padding_data_g, col, padding_trainable,
+                               context_start, context_length, context_stride,
+                               up_pad, down_pad, false, true);
     }
 
     if (filter_g) {
@@ -150,15 +146,9 @@ class SequenceConvGradKernel : public framework::OpKernel<T> {
         padding_data = context.Input<Tensor>("PaddingData");
       }
 
-      sequence_width = static_cast<int>(in->dims()[1]);
-
-      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,
-                          context_stride, up_pad, down_pad, false, false,
-                          false);
+                          context_stride, up_pad, down_pad);
 
       math::matmul<Place, T>(context.device_context(), col, true, out_grad,
                              false, T(1.0), &filter_grad, T(1.0));

python/paddle/v2/framework/tests/test_seq_conv.py

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