Support input_grad = null or filter_grad = null.

paddle/operators/conv2d_op.cc

@@ -28,6 +28,13 @@ class Conv2DOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Input"),
+                            "Input(Input) of Conv2DOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Filter"),
+                            "Input(Filter) of Conv2DOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Output"),
+                            "Output(Output) of Conv2DOp should not be null.");
+
     auto in = ctx.Input<Tensor>("Input");
     auto filter = ctx.Input<Tensor>("Filter");
     auto out = ctx.Output<framework::LoDTensor>("Output");
@@ -108,8 +115,8 @@ class Conv2DOpGrad : public framework::OperatorWithKernel {
         ctx.Output<framework::LoDTensor>(framework::GradVarName("Input"));
     auto d_filter =
         ctx.Output<framework::LoDTensor>(framework::GradVarName("Filter"));
-    d_in->Resize(in->dims());
-    d_filter->Resize(filter->dims());
+    if (d_in) d_in->Resize(in->dims());
+    if (d_filter) d_filter->Resize(filter->dims());
   }
 };
 

paddle/operators/gemm_conv2d_op.h

@@ -111,14 +111,16 @@ class GemmConvGrad2DKernel : public framework::OpKernel {
         context.Output<Tensor>(framework::GradVarName("Input"));
     Tensor* filter_grad_ =
         context.Output<Tensor>(framework::GradVarName("Filter"));
-    input_grad->mutable_data<T>(context.GetPlace());
-    filter_grad_->mutable_data<T>(context.GetPlace());
 
     // The filter and filter_grad will be reshaped in the calculations,
     // so here use an assignment operation,
     // that avoids modifying the variable in the Scope.
     Tensor filter = *context.Input<Tensor>("Filter");
-    Tensor filter_grad = *filter_grad_;
+    Tensor filter_grad;
+    if (filter_grad_) {
+      filter_grad_->mutable_data<T>(context.GetPlace());
+      filter_grad = *filter_grad_;
+    }
 
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
@@ -162,12 +164,20 @@ class GemmConvGrad2DKernel : public framework::OpKernel {
     framework::DDim filter_matrix_shape = {filter.dims()[0],
                                            filter.numel() / filter.dims()[0]};
     filter.Resize(filter_matrix_shape);
-    filter_grad.Resize(filter_matrix_shape);
 
+    if (filter_grad_) {
+      filter_grad.Resize(filter_matrix_shape);
       auto t1 = framework::EigenVector<T>::Flatten(filter_grad);
-    t1.device(context.GetEigenDevice<Place>()) = t1.constant(static_cast<T>(0));
+      t1.device(context.GetEigenDevice<Place>()) =
+          t1.constant(static_cast<T>(0));
+    }
+
+    if (input_grad) {
+      input_grad->mutable_data<T>(context.GetPlace());
       auto t2 = framework::EigenVector<T>::Flatten(*input_grad);
-    t2.device(context.GetEigenDevice<Place>()) = t2.constant(static_cast<T>(0));
+      t2.device(context.GetEigenDevice<Place>()) =
+          t2.constant(static_cast<T>(0));
+    }
 
     auto* device_context =
         const_cast<platform::DeviceContext*>(context.device_context_);
@@ -176,16 +186,24 @@ class GemmConvGrad2DKernel : public framework::OpKernel {
     // convolution backward weight operator: im2col + gemm
     int in_step = input_channels / groups;
     int out_step = output_channels / groups;
+    Tensor in_grad_batch;
+    Tensor in_batch;
     for (int i = 0; i < batch_size; i++) {
       Tensor out_grad_batch =
           output_grad->Slice<T>(i, i + 1).Resize(output_matrix_shape);
-      Tensor in_grad_batch = input_grad->Slice<T>(i, i + 1).Resize(input_shape);
-      Tensor in_batch = input->Slice<T>(i, i + 1).Resize(input_shape);
+      if (input_grad) {
+        in_grad_batch = input_grad->Slice<T>(i, i + 1).Resize(input_shape);
+      }
+      if (filter_grad_) {
+        in_batch = input->Slice<T>(i, i + 1).Resize(input_shape);
+      }
       for (int g = 0; g < groups; g++) {
-        // gemm
         Tensor out_grad_slice =
             out_grad_batch.Slice<T>(g * out_step, (g + 1) * out_step);
-        Tensor filter_slice = filter.Slice<T>(g * out_step, (g + 1) * out_step);
+        if (input_grad) {
+          // gemm
+          Tensor filter_slice =
+              filter.Slice<T>(g * out_step, (g + 1) * out_step);
           math::matmul<Place, T>(filter_slice, true, out_grad_slice, false,
                                  T(1.0), &col_matrix, T(0.0), device_context);
 
@@ -194,17 +212,21 @@ class GemmConvGrad2DKernel : public framework::OpKernel {
               in_grad_batch.Slice<T>(g * in_step, (g + 1) * in_step);
           col2im(in_grad_slice, col, strides[0], strides[1], paddings[0],
                  paddings[1], device_context);
+        }
 
+        if (filter_grad_) {
           // im2col
           Tensor in_slice = in_batch.Slice<T>(g * in_step, (g + 1) * in_step);
-        im2col(in_slice, col, strides[0], strides[1], paddings[0], paddings[1],
-               device_context);
+          im2col(in_slice, col, strides[0], strides[1], paddings[0],
+                 paddings[1], device_context);
 
           // gemm
           Tensor filter_grad_slice =
               filter_grad.Slice<T>(g * out_step, (g + 1) * out_step);
-        math::matmul<Place, T>(out_grad_slice, false, col_matrix, true, T(1.0),
-                               &filter_grad_slice, T(1.0), device_context);
+          math::matmul<Place, T>(out_grad_slice, false, col_matrix, true,
+                                 T(1.0), &filter_grad_slice, T(1.0),
+                                 device_context);
+        }
       }
     }
   }

python/paddle/v2/framework/tests/test_conv2d_op.py

@@ -75,6 +75,12 @@ class TestConv2dOp(OpTest):
     def test_check_grad(self):
         self.check_grad(set(['Input', 'Filter']), 'Output')
 
+    def test_check_grad_no_filter(self):
+        self.check_grad(['Input'], 'Output', no_grad_set=set(['Filter']))
+
+    def test_check_grad_no_input(self):
+        self.check_grad(['Filter'], 'Output', no_grad_set=set(['Input']))
+
     def init_groups(self):
         self.groups = 1