From badc6f2233f7b4ede9ce267845f3368947d4d63c Mon Sep 17 00:00:00 2001
From: lijianshe02 <48898730+lijianshe02@users.noreply.github.com>
Date: Fri, 15 Jan 2021 12:57:50 +0800
Subject: [PATCH]  add transpose double grad , cherry-pick from #29600 (#30435)

* add transpose double grad test=develop (#29600)

* add transpose double grad test=develop

* cherry-pick test=develop
---
 paddle/fluid/operators/transpose_op.cc        | 18 +++++++-
 .../fluid/tests/unittests/test_nn_grad.py     | 44 +++++++++++++++++++
 2 files changed, 61 insertions(+), 1 deletion(-)

diff --git a/paddle/fluid/operators/transpose_op.cc b/paddle/fluid/operators/transpose_op.cc
index 42f4a819baa..d9940ddca3e 100644
--- a/paddle/fluid/operators/transpose_op.cc
+++ b/paddle/fluid/operators/transpose_op.cc
@@ -272,6 +272,20 @@ class Transpose2GradMaker : public framework::SingleGradOpMaker<T> {
   }
 };
 
+template <typename T>
+class Transpose2DoubleGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> grad_op) const override {
+    grad_op->SetType("transpose2");
+    grad_op->SetInput("X", this->OutputGrad(framework::GradVarName("X")));
+    grad_op->SetOutput("Out", this->InputGrad(framework::GradVarName("Out")));
+    grad_op->SetOutput("XShape", this->Input("XShape"));
+    grad_op->SetAttrMap(this->Attrs());
+  }
+};
+
 class Transpose2OpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -338,7 +352,9 @@ REGISTER_OP_CPU_KERNEL(
 REGISTER_OPERATOR(transpose2, ops::Transpose2Op, ops::Transpose2OpMaker,
                   ops::Transpose2GradMaker<paddle::framework::OpDesc>,
                   ops::Transpose2GradMaker<paddle::imperative::OpBase>);
-REGISTER_OPERATOR(transpose2_grad, ops::Transpose2OpGrad);
+REGISTER_OPERATOR(transpose2_grad, ops::Transpose2OpGrad,
+                  ops::Transpose2DoubleGradMaker<paddle::framework::OpDesc>,
+                  ops::Transpose2DoubleGradMaker<paddle::imperative::OpBase>);
 
 REGISTER_OP_CPU_KERNEL(
     transpose2, ops::TransposeKernel<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 899c1f798e6..60b95e9e470 100644
--- a/python/paddle/fluid/tests/unittests/test_nn_grad.py
+++ b/python/paddle/fluid/tests/unittests/test_nn_grad.py
@@ -329,5 +329,49 @@ class TestUnsqueezeDoubleGradCheck(unittest.TestCase):
             self.func(p)
 
 
+class TestTransposeDoubleGradCheck(unittest.TestCase):
+    @prog_scope()
+    def func(self, place):
+        x_shape = [3, 40]
+        perm = [1, 0]
+        dtype = np.float64
+
+        x = layers.data('x', x_shape, False, dtype)
+        x.persistable = True
+        out = paddle.transpose(x, perm)
+        x_arr = np.random.uniform(-1, 1, x_shape).astype(dtype)
+
+        gradient_checker.double_grad_check([x], out, x_init=x_arr, place=place)
+
+    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 TestTransposeDoubleGradCheckCase1(unittest.TestCase):
+    @prog_scope()
+    def func(self, place):
+        x_shape = [2, 3, 4, 5]
+        perm = [0, 2, 3, 1]
+        dtype = np.float64
+
+        x = layers.data('x', x_shape, False, dtype)
+        x.persistable = True
+        out = paddle.transpose(x, perm)
+        x_arr = np.random.uniform(-1, 1, x_shape).astype(dtype)
+
+        gradient_checker.double_grad_check([x], out, x_init=x_arr, place=place)
+
+    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