From c5f957ae387f6193a399c8749f923f1635df19e3 Mon Sep 17 00:00:00 2001
From: lilong12 <lilong12@baidu.com>
Date: Thu, 10 Sep 2020 15:55:09 +0800
Subject: [PATCH] add double grad for tile op and expand_v2 op (#27114)

* add double grad for tile, test=develop

* add double grad for expand_v2 op, test=develop
---
 paddle/fluid/operators/expand_v2_op.cc        | 22 +++++++++
 paddle/fluid/operators/tile_op.cc             | 22 +++++++++
 .../fluid/tests/unittests/test_nn_grad.py     | 49 +++++++++++++++++++
 3 files changed, 93 insertions(+)

diff --git a/paddle/fluid/operators/expand_v2_op.cc b/paddle/fluid/operators/expand_v2_op.cc
index 359d512c34..a1ee47b7f9 100644
--- a/paddle/fluid/operators/expand_v2_op.cc
+++ b/paddle/fluid/operators/expand_v2_op.cc
@@ -230,6 +230,26 @@ class ExpandV2GradOpMaker : public framework::SingleGradOpMaker<T> {
   }
 };
 
+template <typename T>
+class ExpandV2DoubleGradOpMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("expand_v2");
+    op->SetInput("X", this->OutputGrad(framework::GradVarName("X")));
+    op->SetOutput("Out", this->InputGrad(framework::GradVarName("Out")));
+    if (this->HasInput("expand_shapes_tensor")) {
+      op->SetInput("expand_shapes_tensor", this->Input("expand_shapes_tensor"));
+    }
+    if (this->HasInput("Shape")) {
+      op->SetInput("Shape", this->Input("Shape"));
+    }
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
 DECLARE_NO_NEED_BUFFER_VARS_INFERER(ExpandV2GradNoNeedBufVarsInferer, "X");
 
 }  // namespace operators
@@ -240,6 +260,8 @@ REGISTER_OPERATOR(expand_v2, ops::ExpandV2Op, ops::ExpandV2OpMaker,
                   ops::ExpandV2GradOpMaker<paddle::framework::OpDesc>,
                   ops::ExpandV2GradOpMaker<paddle::imperative::OpBase>);
 REGISTER_OPERATOR(expand_v2_grad, ops::ExpandV2GradOp,
+                  ops::ExpandV2DoubleGradOpMaker<paddle::framework::OpDesc>,
+                  ops::ExpandV2DoubleGradOpMaker<paddle::imperative::OpBase>,
                   ops::ExpandV2GradNoNeedBufVarsInferer);
 REGISTER_OP_CPU_KERNEL(
     expand_v2, ops::ExpandV2Kernel<paddle::platform::CPUDeviceContext, float>,
diff --git a/paddle/fluid/operators/tile_op.cc b/paddle/fluid/operators/tile_op.cc
index da4ca87296..bc1cb3b4aa 100644
--- a/paddle/fluid/operators/tile_op.cc
+++ b/paddle/fluid/operators/tile_op.cc
@@ -241,6 +241,26 @@ class TileGradOpMaker : public framework::SingleGradOpMaker<T> {
   }
 };
 
+template <typename T>
+class TileDoubleGradOpMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("tile");
+    op->SetInput("X", this->OutputGrad(framework::GradVarName("X")));
+    op->SetOutput("Out", this->InputGrad(framework::GradVarName("Out")));
+    if (this->HasInput("repeat_times_tensor")) {
+      op->SetInput("repeat_times_tensor", this->Input("repeat_times_tensor"));
+    }
+    if (this->HasInput("RepeatTimes")) {
+      op->SetInput("RepeatTimes", this->Input("RepeatTimes"));
+    }
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
 DECLARE_NO_NEED_BUFFER_VARS_INFERER(TileGradNoNeedBufVarsInferer, "X");
 
 }  // namespace operators
@@ -251,6 +271,8 @@ REGISTER_OPERATOR(tile, ops::TileOp, ops::TileOpMaker,
                   ops::TileGradOpMaker<paddle::framework::OpDesc>,
                   ops::TileGradOpMaker<paddle::imperative::OpBase>);
 REGISTER_OPERATOR(tile_grad, ops::TileGradOp,
+                  ops::TileDoubleGradOpMaker<paddle::framework::OpDesc>,
+                  ops::TileDoubleGradOpMaker<paddle::imperative::OpBase>,
                   ops::TileGradNoNeedBufVarsInferer);
 REGISTER_OP_CPU_KERNEL(
     tile, ops::TileKernel<paddle::platform::CPUDeviceContext, float>,
diff --git a/python/paddle/fluid/tests/unittests/test_nn_grad.py b/python/paddle/fluid/tests/unittests/test_nn_grad.py
index 0e4f89f602..5d1e016287 100644
--- a/python/paddle/fluid/tests/unittests/test_nn_grad.py
+++ b/python/paddle/fluid/tests/unittests/test_nn_grad.py
@@ -17,6 +17,7 @@ from __future__ import print_function
 import unittest
 import numpy as np
 
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.layers as layers
 import paddle.fluid.core as core
@@ -200,5 +201,53 @@ class TestExpandDoubleGradCheck(unittest.TestCase):
             self.func(p)
 
 
+class TestTileDoubleGradCheck(unittest.TestCase):
+    @prog_scope()
+    def func(self, place):
+        x_shape = [3, 12]
+        repeat_times = [4, 9]
+        eps = 0.005
+        dtype = np.float64
+
+        x = layers.data('x', x_shape, False, dtype)
+        x.persistable = True
+        out = paddle.tile(x, repeat_times)
+        x_arr = np.random.uniform(-1, 1, x_shape).astype(dtype)
+
+        gradient_checker.double_grad_check(
+            [x], out, x_init=x_arr, place=place, eps=eps)
+
+    def test_grad(self):
+        places = [fluid.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(fluid.CUDAPlace(0))
+        for p in places:
+            self.func(p)
+
+
+class TestExpandV2DoubleGradCheck(unittest.TestCase):
+    @prog_scope()
+    def func(self, place):
+        x_shape = [1, 12]
+        new_shape = [4, 12]
+        eps = 0.005
+        dtype = np.float64
+
+        x = layers.data('x', x_shape, False, dtype)
+        x.persistable = True
+        out = paddle.expand(x, new_shape)
+        x_arr = np.random.uniform(-1, 1, x_shape).astype(dtype)
+
+        gradient_checker.double_grad_check(
+            [x], out, x_init=x_arr, place=place, eps=eps)
+
+    def test_grad(self):
+        places = [fluid.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(fluid.CUDAPlace(0))
+        for p in places:
+            self.func(p)
+
+
 if __name__ == "__main__":
     unittest.main()
-- 
GitLab