Refine gradient proto maker and python API for affine_channel_op (#16340)

* Rewrite gradient ProtoMaker for affine_channel_op to remove the Output as the input.
* Add act in Python API to make the act can be in-place by layer_help.py

paddle/fluid/API.spec

@@ -205,7 +205,7 @@ paddle.fluid.layers.maxout (ArgSpec(args=['x', 'groups', 'name'], varargs=None,
 paddle.fluid.layers.space_to_depth (ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5f207ae10589ebe38a63575ef6ff8e1e'))
 paddle.fluid.layers.affine_grid (ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '51def402b8910e163cbace9d0c0526ed'))
 paddle.fluid.layers.sequence_reverse (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '77a6d80aa5551ca70324fc975c44507f'))
-paddle.fluid.layers.affine_channel (ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None)), ('document', '2f46f1ff39a13ab00857e7b9f44b2fa7'))
+paddle.fluid.layers.affine_channel (ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name', 'act'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None, None)), ('document', 'ab84fdc6dc60f3ad9aa397e6007e3bf9'))
 paddle.fluid.layers.similarity_focus (ArgSpec(args=['input', 'axis', 'indexes', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '70e3b5182a18b40b47ecabd7c8490a35'))
 paddle.fluid.layers.hash (ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None)), ('document', '9bb77f8dc002dd2ce75d4769eaaf5007'))
 paddle.fluid.layers.grid_sampler (ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd256cba1c41a5ed92ce3f31e24a2ca6d'))

paddle/fluid/operators/affine_channel_op.cc

@@ -113,6 +113,27 @@ class AffineChannelOpGrad : public framework::OperatorWithKernel {
   }
 };
 
+class AffineChannelGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* op = new framework::OpDesc();
+    op->SetType("affine_channel_grad");
+    op->SetInput("X", Input("X"));
+    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    op->SetInput("Scale", Input("Scale"));
+
+    op->SetAttrMap(Attrs());
+
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetOutput(framework::GradVarName("Scale"), InputGrad("Scale"));
+    op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias"));
+
+    return std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
 template <typename T>
 using EigenArrayMap =
     Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
@@ -260,8 +281,7 @@ namespace ops = paddle::operators;
 using CPU = paddle::platform::CPUDeviceContext;
 
 REGISTER_OPERATOR(affine_channel, ops::AffineChannelOp,
-                  ops::AffineChannelOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
+                  ops::AffineChannelOpMaker, ops::AffineChannelGradMaker);
 REGISTER_OPERATOR(affine_channel_grad, ops::AffineChannelOpGrad);
 
 REGISTER_OP_CPU_KERNEL(affine_channel, ops::AffineChannelKernel<CPU, float>,

python/paddle/fluid/layers/nn.py

@@ -9706,7 +9706,12 @@ def sequence_reverse(x, name=None):
     return out
 
 
-def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
+def affine_channel(x,
+                   scale=None,
+                   bias=None,
+                   data_layout='NCHW',
+                   name=None,
+                   act=None):
     """
     Applies a separate affine transformation to each channel of the input.
     Useful for replacing spatial batch norm with its equivalent fixed
@@ -9725,6 +9730,7 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
         data_layout (string, default NCHW): NCHW or NHWC. If input is 2D
             tensor, you can ignore data_layout.
         name (str, default None): The name of this layer.
+        act (str, default None): Activation to be applied to the output of this layer.
 
     Returns:
         out (Variable): A tensor of the same shape and data layout with x.
@@ -9744,7 +9750,7 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
                 'Bias': bias},
         attrs={"data_layout": data_layout},
         outputs={"Out": out})
-    return out
+    return helper.append_activation(pre_activation)
 
 
 def similarity_focus(input, axis, indexes, name=None):