follow comments

paddle/operators/CMakeLists.txt

@@ -128,7 +128,7 @@ op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
 op_library(sum_op DEPS net_op)
 op_library(pool_op DEPS pooling)
 op_library(pool_with_index_op DEPS pooling)
-op_library(sequence_conv_op DEPS sequence_project)
+op_library(sequence_conv_op DEPS context_project)
 op_library(lstm_op DEPS sequence2batch lstm_compute)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})

paddle/operators/math/CMakeLists.txt

@@ -9,7 +9,7 @@ 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)
+    nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context)
     nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context)
     nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions)
 else()
@@ -19,7 +19,7 @@ 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)
+    cc_library(context_project SRCS context_project.cc DEPS device_context)
     cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context)
     cc_library(lstm_compute SRCS lstm_compute.cc DEPS device_context activation_functions)
 endif()

paddle/operators/math/sequence_project.cc → paddle/operators/math/context_project.cc

@@ -12,14 +12,14 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/operators/math/sequence_project.h"
+#include "paddle/operators/math/context_project.h"
 
 namespace paddle {
 namespace operators {
 namespace math {
 
-template class SequenceProjectFunctor<platform::CPUPlace, float>;
-template class SequenceProjectFunctor<platform::CPUPlace, double>;
+template class ContextProjectFunctor<platform::CPUPlace, float>;
+template class ContextProjectFunctor<platform::CPUPlace, double>;
 
 }  // namespace math
 }  // namespace operators

paddle/operators/math/sequence_project.cu → paddle/operators/math/context_project.cu

@@ -14,14 +14,14 @@ limitations under the License. */
 
 #define EIGEN_USE_GPU
 
-#include "paddle/operators/math/sequence_project.h"
+#include "paddle/operators/math/context_project.h"
 
 namespace paddle {
 namespace operators {
 namespace math {
 
-template class SequenceProjectFunctor<platform::GPUPlace, float>;
-template class SequenceProjectFunctor<platform::GPUPlace, double>;
+template class ContextProjectFunctor<platform::GPUPlace, float>;
+template class ContextProjectFunctor<platform::GPUPlace, double>;
 
 }  // namespace math
 }  // namespace operators

paddle/operators/math/sequence_project.h → paddle/operators/math/context_project.h

@@ -23,31 +23,29 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-//    template <typename T, int MajorType = Eigen::RowMajor,
-//            typename IndexType = Eigen::DenseIndex>
-//    using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 /*
- * \brief SequenceProject projects features of context_length time-steps of each
- * instance.
- *
+ * \brief Context projection concatenate 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.
+
  * \param in            Input data.
- * \param inShape       The shape of Input data,
+ * \param Shape         The shape of Input data,
  *                      [minibatch, number_of_input_features].
- * \param inShape       A float LoDTensor.
+ * \param type          A float LoDTensor.
  *
  * \param padding_data  Padding data.
- * \param inShape       The shape of Padding data,
+ * \param Shape         The shape of Padding data,
  *                      [up_pad + down_pad, number_of_input_features].
- * \param inShape       A float LoDTensor.
+ * \param type          A float Tensor.
  *
  * \param col           Col data.
- * \param inShape       The shape of Col data,
- *                      [minibatch, 1].
- * \param inShape       A float LoDTensor.
+ * \param Shape         The shape of Col data,
+ *                      [minibatch, context_length * number_of_input_features].
+ * \param type           A float Tensor.
  *
  * For a mini-batch of 2 variable lengths sentences, containing 3, and 1
  * time-steps:
@@ -87,7 +85,7 @@ using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
  */
 
 template <typename Place, typename T>
-class SequenceProjectFunctor {
+class ContextProjectFunctor {
  public:
   void operator()(const platform::DeviceContext& context,
                   framework::LoDTensor& in, framework::Tensor& padding_data,
@@ -147,8 +145,7 @@ class SequenceProjectFunctor {
                        /*stride_height*/ context_stride, /*stride_width*/ 1,
                        up_pad, down_pad, 0, 0);
           }
-          out_t.Resize(framework::make_ddim(
-              {sequence_height, context_length * sequence_width}));
+          out_t.Resize({sequence_height, context_length * sequence_width});
         }
       }
     }
@@ -162,8 +159,7 @@ class SequenceProjectFunctor {
           sequence_height = static_cast<int>(out_t.dims()[0]);
 
           // add up trainable data
-          out_t.Resize(framework::make_ddim(
-              {sequence_height * context_length, sequence_width}));
+          out_t.Resize({sequence_height * context_length, sequence_width});
 
           if (up_pad > 0) {  // add up pad
             int padding_rows = std::min(
@@ -223,8 +219,7 @@ class SequenceProjectFunctor {
               }
             }
           }
-          out_t.Resize(framework::make_ddim(
-              {sequence_height, context_length * sequence_width}));
+          out_t.Resize({sequence_height, context_length * sequence_width});
         }
       }
     }

paddle/operators/sequence_conv_op.cc

@@ -38,10 +38,9 @@ class SequenceConvOp : public framework::OperatorWithKernel {
     auto filter_dims = ctx->GetInputDim("Filter");
     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 && filter_dims[1] == in_dims[1],
-        "Filter's shape should be (context_length x "
-        "number_of_input_features).");
+    PADDLE_ENFORCE(filter_dims[0] == context_length * in_dims[1],
+                   "Filter's height should be context_length * "
+                   "number_of_input_features .");
 
     if (padding_trainable) {
       PADDLE_ENFORCE(
@@ -66,8 +65,9 @@ class SequenceConvOp : public framework::OperatorWithKernel {
           "and 'context_length'.");
     }
 
-    in_dims[1] = 1;
+    in_dims[1] = filter_dims[1];
     ctx->SetOutputDim("Out", in_dims);
+    ctx->ShareLoD("X", "Out");
   }
 };
 
@@ -101,35 +101,51 @@ class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker {
   SequenceConvOpMaker(framework::OpProto* proto,
                       framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X",
-             "(A float LoDTensor) the input of SequenceConvOp, a vector of "
-             "2-D matrix of size (minibatch, number_of_input_features).");
+    AddInput(
+        "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.");
     AddInput("PaddingData",
-             "(Tensor) the input of SequenceConvOp, a vector of "
-             "2-D matrix of size (up_pad + down_pad, "
-             "number_of_input_features). ")
+             "(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 "
+             "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",
-             "(Tensor) the input of SequenceConvOp, a vector of "
-             "2-D matrix of size (context_length x number_of_input_features).");
-    AddOutput("Out",
-              "(A float LoDTensor) the output of SequenceConvOp, a vector "
-              "of 2-D matrix of size (minibatch, 1).");
+             "(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.");
+    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.");
 
     AddAttr<bool>("padding_trainable",
                   "(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.")
+                 "(int, default 3) the context_length 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.")
+                 "(int, default 0) the context_start of SequenceConvOp "
+                 "represents the beginning of the convolution of the number of "
+                 "rows of sequence, which can be negative.")
         .SetDefault(0);
     AddAttr<int>("context_stride",
-                 "(int, default 1) the context_stride of SequenceConvOp. "
-                 "Currently, sequence_project_op only support "
+                 "(int, default 1) the context_stride of SequenceConvOp "
+                 "represents the step length of convolution. "
+                 "Currently, SequenceConvOp only supports"
                  "context_stride=1.")
         .SetDefault(1)
         .GreaterThan(0);
@@ -139,14 +155,10 @@ class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker {
     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)
+    parameters is checked in the infer-shape. 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.
     )DOC");
   }
 };