use Daoyuan's SequenceArg.

paddle/function/BufferArg.h

@@ -190,7 +190,7 @@ private:
   size_t numSeqs_;
 };
 
-// sequence data
+// sequence data {seqId(vec), buf(matrix)}
 class SequenceArg : public BufferArg {
 public:
   SequenceArg(void* buf,
@@ -210,6 +210,7 @@ public:
 
   void* getIdBuf() const { return startPositions_.data(); }
   size_t numSeqs() const { return startPositions_.numSeqs(); }
+  const SequenceIdArg& getSequenceIds() const { return startPositions_; }
 
 private:
   SequenceIdArg startPositions_;

paddle/function/ContextProjectionOp.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include "paddle/math/Vector.h"
 
 namespace paddle {
-
 /**
  * Context Projection Forward with CPU Matrix Device.
  *
@@ -208,10 +207,10 @@ void ContextProjectionBackward<DEVICE_TYPE_CPU>(const CpuMatrix& out_grad_mat,
  * Context Projection Backward Function.
  * Update the weight gradient and input layer gradient with backprop
  *
- * \param inputs[0]       input sequence.
- * \param inputs[1]       output layer grad.
- * \param outputs[0]      input layer grad.
- * \param outputs[1]      weight grad.
+ * \param inputs[0].seq          input sequence.
+ * \param inputs[0].matrix       output layer grad.
+ * \param outputs[0]             input layer grad.
+ * \param outputs[1]             weight grad.
  */
 template <DeviceType Device>
 class ContextProjectionBackwardFunc : public FunctionBase {
@@ -225,27 +224,28 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ((size_t)2, inputs.size());
+    CHECK_EQ((size_t)1, inputs.size());
     CHECK_EQ((size_t)2, outputs.size());
 
-    CHECK(inputs[0].data() && inputs[1].data());
-    CHECK_EQ(inputs[0].shape().ndims(), (size_t)1);
-    CHECK_EQ(inputs[1].shape().ndims(), (size_t)2);
+    const auto seqArg = dynamic_cast<const SequenceArg&>(inputs[0]);
+    CHECK(seqArg.data() && inputs[0].data());
+    CHECK_EQ(seqArg.shape().ndims(), (size_t)2);
+    CHECK_EQ(seqArg.getSequenceIds().shape().ndims(), (size_t)1);
     CHECK_EQ(outputs[0].shape().ndims(), (size_t)2);
     CHECK_EQ(outputs[1].shape().ndims(), (size_t)2);
 
     /// dim of input grad == dim of weight
     CHECK_EQ(outputs[0].shape()[1], outputs[1].shape()[1]);
     /// input and output grad has the same batch_size
-    CHECK_EQ(outputs[0].shape()[0], inputs[1].shape()[0]);
+    CHECK_EQ(outputs[0].shape()[0], seqArg.shape()[0]);
     /// dim of output val = dim of input grad * context_length
-    CHECK_EQ(inputs[1].shape()[1], outputs[0].shape()[1] * context_length_);
+    CHECK_EQ(seqArg.shape()[1], outputs[0].shape()[1] * context_length_);
 
     CHECK_EQ(outputs[0].getArgType(), ADD_TO);
     CHECK_EQ(outputs[1].getArgType(), ADD_TO);
 
-    const auto seq_vec = inputs[0].vector<int, Device>();
-    const auto out_grad_mat = inputs[1].matrix<Device>();
+    const auto seq_vec = seqArg.getSequenceIds().vector<int, Device>();
+    const auto out_grad_mat = seqArg.matrix<Device>();
     auto in_grad_mat =
         !outputs[0].data()
             ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)

paddle/function/ContextProjectionOpTest.cpp

@@ -122,8 +122,7 @@ void testMatrixProjectionBackward(int context_start,
 
   BufferArgs cpu_inputs;
   BufferArgs cpu_outputs;
-  cpu_inputs.addArg(*cpu_seq);
-  cpu_inputs.addArg(cpu_out_grad);
+  cpu_inputs.addArg(cpu_out_grad, *cpu_seq);
   cpu_outputs.addArg(cpu_in_grad, ADD_TO);
   cpu_outputs.addArg(
       cpu_w_grad ? *cpu_w_grad : CpuMatrix(nullptr, 0, input_dim), ADD_TO);
@@ -132,8 +131,7 @@ void testMatrixProjectionBackward(int context_start,
 
   BufferArgs gpu_inputs;
   BufferArgs gpu_outputs;
-  gpu_inputs.addArg(*gpu_seq);
-  gpu_inputs.addArg(gpu_out_grad);
+  gpu_inputs.addArg(gpu_out_grad, *gpu_seq);
   gpu_outputs.addArg(gpu_in_grad, ADD_TO);
   gpu_outputs.addArg(
       gpu_w_grad ? *gpu_w_grad : GpuMatrix(nullptr, 0, input_dim), ADD_TO);

paddle/function/Function.cpp

@@ -90,6 +90,12 @@ void BufferArgs::addArg(const GpuSparseMatrix& arg, ArgType argType) {
   args_.push_back(std::make_shared<SparseMatrixArg>(arg, argType));
 }
 
+void BufferArgs::addArg(const Matrix& matrix,
+                        const IVector& vector,
+                        ArgType argType) {
+  args_.push_back(std::make_shared<SequenceArg>(matrix, vector, argType));
+}
+
 ClassRegistrar<FunctionBase> FunctionBase::funcRegistrar_;
 
 }  // namespace paddle

paddle/function/Function.h

@@ -77,6 +77,10 @@ public:
   void addArg(const CpuSparseMatrix& arg, ArgType argType = UNSPECIFIED);
   void addArg(const GpuSparseMatrix& arg, ArgType argType = UNSPECIFIED);
 
+  void addArg(const Matrix& matrix,
+              const IVector& vector,
+              ArgType argType = UNSPECIFIED);
+
   // get argument
   const BufferArg& operator[](size_t num) const {
     CHECK_LT(num, args_.size());

paddle/gserver/layers/ContextProjection.cpp

@@ -166,8 +166,7 @@ void ContextProjection::backward(const UpdateCallback& callback) {
 
   BufferArgs inputs;
   BufferArgs outputs;
-  inputs.addArg(*in_->sequenceStartPositions->getVector(useGpu_));
-  inputs.addArg(*out_->grad);
+  inputs.addArg(*out_->grad, *in_->sequenceStartPositions->getVector(useGpu_));
   outputs.addArg(
       CpuMatrix(
           in_->grad ? in_->grad->getData() : nullptr, batch_size, input_dim),