From ff1da1880ece95db066de737d117a1a671385286 Mon Sep 17 00:00:00 2001
From: Charles-hit <56987902+Charles-hit@users.noreply.github.com>
Date: Tue, 13 Sep 2022 20:40:12 +0800
Subject: [PATCH] support concat backward refuse forward (#45940)

---
 paddle/phi/api/yaml/legacy_backward.yaml      |  6 +-
 .../fluid/tests/unittests/test_concat_op.py   | 81 +++++++++++++++++++
 2 files changed, 82 insertions(+), 5 deletions(-)

diff --git a/paddle/phi/api/yaml/legacy_backward.yaml b/paddle/phi/api/yaml/legacy_backward.yaml
index a284f4304e6..090283877ff 100755
--- a/paddle/phi/api/yaml/legacy_backward.yaml
+++ b/paddle/phi/api/yaml/legacy_backward.yaml
@@ -430,11 +430,7 @@
   forward : concat_grad (Tensor[] x, Tensor grad_out, Scalar axis) -> Tensor[](grad_x)
   args : (Tensor[] grad_x_grad, Scalar axis = 0)
   output : Tensor(grad_out_grad)
-  infer_meta :
-    func : ConcatInferMeta
-    param : [grad_x_grad, axis]
-  kernel :
-    func : concat
+  invoke : concat(grad_x_grad, axis)
 
 - backward_api : concat_grad
   forward : concat (Tensor[] x, Scalar axis) -> Tensor(out)
diff --git a/python/paddle/fluid/tests/unittests/test_concat_op.py b/python/paddle/fluid/tests/unittests/test_concat_op.py
index 0bf3d6230d8..10c74107083 100644
--- a/python/paddle/fluid/tests/unittests/test_concat_op.py
+++ b/python/paddle/fluid/tests/unittests/test_concat_op.py
@@ -21,6 +21,9 @@ import paddle.fluid as fluid
 from paddle.fluid import compiler, Program, program_guard, core
 from paddle.fluid.framework import _test_eager_guard
 import paddle
+import gradient_checker
+from decorator_helper import prog_scope
+import paddle.fluid.layers as layers
 
 
 class TestConcatOp(OpTest):
@@ -451,5 +454,83 @@ class TestConcatAPIWithLoDTensorArray(unittest.TestCase):
             res[0], np.concatenate([self.x] * self.iter_num, axis=self.axis))
 
 
+class TestConcatDoubleGradCheck(unittest.TestCase):
+
+    def concat_wrapper(self, x):
+        return paddle.concat(x)
+
+    @prog_scope()
+    def func(self, place):
+        # the shape of input variable should be clearly specified, not inlcude -1.
+        eps = 0.005
+        dtype = np.float32
+
+        data1 = layers.data('data1', [2, 3], False, dtype)
+        data1.persistable = True
+        data2 = layers.data('data2', [2, 3], False, dtype)
+        data2.persistable = True
+        out = paddle.concat([data1, data2])
+        data1_arr = np.random.uniform(-1, 1, data1.shape).astype(dtype)
+        data2_arr = np.random.uniform(-1, 1, data2.shape).astype(dtype)
+        gradient_checker.double_grad_check([data1, data2],
+                                           out,
+                                           x_init=[data1_arr, data2_arr],
+                                           place=place,
+                                           eps=eps)
+        fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
+        gradient_checker.double_grad_check_for_dygraph(
+            self.concat_wrapper, [data1, data2],
+            out,
+            x_init=[data1_arr, data2_arr],
+            place=place)
+
+    def test_grad(self):
+        paddle.enable_static()
+        places = [fluid.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(fluid.CUDAPlace(0))
+        for p in places:
+            self.func(p)
+
+
+class TestConcatTripleGradCheck(unittest.TestCase):
+
+    def concat_wrapper(self, x):
+        return paddle.concat(x, 1)
+
+    @prog_scope()
+    def func(self, place):
+        # the shape of input variable should be clearly specified, not inlcude -1.
+        eps = 0.005
+        dtype = np.float32
+
+        data1 = layers.data('data1', [2, 3, 4], False, dtype)
+        data1.persistable = True
+        data2 = layers.data('data2', [2, 3, 4], False, dtype)
+        data2.persistable = True
+        out = paddle.concat([data1, data2], 1)
+        data1_arr = np.random.uniform(-1, 1, data1.shape).astype(dtype)
+        data2_arr = np.random.uniform(-1, 1, data2.shape).astype(dtype)
+        gradient_checker.double_grad_check([data1, data2],
+                                           out,
+                                           x_init=[data1_arr, data2_arr],
+                                           place=place,
+                                           eps=eps)
+        fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
+        gradient_checker.double_grad_check_for_dygraph(
+            self.concat_wrapper, [data1, data2],
+            out,
+            x_init=[data1_arr, data2_arr],
+            place=place)
+
+    def test_grad(self):
+        paddle.enable_static()
+        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