[View] Change view_shape_grad to use invoke view_shape (#71086)

HydrogenSulfate · web-flow · commit c4c05c49b473 · 2025-02-13T13:35:49.000+08:00
* change view_reshape_grad to use invoke view_reshape for high-order derivative

* remove unecessary attrs when generating nodes.cc

* Update dygraph_ops.yaml

Remove `paddle::dialect::InferSymbolicShapeInterface` for view_shape
diff --git a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
@@ -2710,7 +2710,12 @@ def GenerateNodeDefinition(
             )
 
             grad_api_args[grad_api_position] = name
-            get_grad_in_args_list.append(get_attr_str)
+            if (
+                not is_invoke_forward_api
+                or name in self.grad_api_contents['invoke']
+            ):
+                # NOTE: attr 'dims' is not necessary for 'invoke: view_shape(out_grad, input.shape())'
+                get_grad_in_args_list.append(get_attr_str)
 
         get_grad_in_args_str = "\n".join(get_grad_in_args_list)
 
diff --git a/paddle/phi/ops/yaml/backward.yaml b/paddle/phi/ops/yaml/backward.yaml
@@ -3625,16 +3625,6 @@
     func : view_dtype_grad
     data_type : out_grad
 
-- backward_op : view_shape_grad
-  forward : view_shape (Tensor input, int64_t[] dims = {}) -> Tensor(out)
-  args : (Tensor input, Tensor out_grad, int64_t[] dims = {})
-  output : Tensor(input_grad)
-  infer_meta :
-    func : StridedUnChangedInferMeta
-    param : [input]
-  kernel :
-    func : view_shape_grad
-
 - backward_op : warpctc_grad
   forward : warpctc (Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank = 0, bool norm_by_times = false) -> Tensor(loss), Tensor(warpctcgrad)
   args : (Tensor logits, Tensor logits_length, Tensor warpctcgrad, Tensor loss_grad, int blank, bool norm_by_times)
diff --git a/paddle/phi/ops/yaml/inconsistent/dygraph_backward.yaml b/paddle/phi/ops/yaml/inconsistent/dygraph_backward.yaml
@@ -358,6 +358,12 @@
   composite : tile_grad(x, out_grad, repeat_times, x_grad)
   backward : tile_double_grad
 
+- backward_op : view_shape_grad
+  forward : view_shape (Tensor input, int64_t[] dims = {}) -> Tensor(out)
+  args : (Tensor input, Tensor out_grad, int64_t[] dims = {})
+  output : Tensor(input_grad)
+  invoke: view_shape(out_grad, input.shape())
+
 - backward_op: fused_gemm_epilogue_grad
   forward : fused_gemm_epilogue(Tensor x, Tensor y, Tensor bias, bool trans_x, bool trans_y, str activation) -> Tensor(out), Tensor(reserve_space)
   args : (Tensor x, Tensor y, Tensor reserve_space, Tensor out_grad, bool trans_x, bool trans_y, str activation)
diff --git a/paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml b/paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml
@@ -385,3 +385,12 @@
     data_type : x
   optional : indices, inverse, counts
   traits : paddle::dialect::ForwardOnlyTrait
+
+- op : view_shape
+  args : (Tensor input, int64_t[] dims = {})
+  output : Tensor(out)
+  infer_meta :
+    func : ViewShapeInferMeta
+  kernel :
+    func : view_shape
+  backward : view_shape_grad
diff --git a/paddle/phi/ops/yaml/ops.yaml b/paddle/phi/ops/yaml/ops.yaml
@@ -5294,16 +5294,6 @@
   no_need_buffer : input
   interfaces : paddle::dialect::InferSymbolicShapeInterface
 
-- op : view_shape
-  args : (Tensor input, int64_t[] dims = {})
-  output : Tensor(out)
-  infer_meta :
-    func : ViewShapeInferMeta
-  kernel :
-    func : view_shape
-  backward : view_shape_grad
-  interfaces : paddle::dialect::InferSymbolicShapeInterface
-
 - op : view_slice
   args : (Tensor input, int64_t begin_idx, int64_t end_idx)
   output : Tensor
diff --git a/test/legacy_test/test_stride.py b/test/legacy_test/test_stride.py
@@ -664,6 +664,33 @@ def call_view7(self):
 
         self.assertTrue(out_c._is_shared_buffer_with(out))
 
+    # test invoke of view_reshape_grad for high order derivative
+    def call_view8(self):
+        x_np = np.random.random(size=[10, 10, 10, 20]).astype('float32')
+        x = paddle.to_tensor(x_np)
+        x.stop_gradient = False
+        y = x.view([10, -1]).tanh()
+        dx = paddle.grad(y, x, create_graph=True)[0]
+        np.testing.assert_allclose(
+            dx.numpy(), (1 - y**2).reshape(x.shape).numpy(), 1e-6, 1e-6
+        )
+        dxx = paddle.grad(dx, x, create_graph=True)[0]
+        np.testing.assert_allclose(
+            dxx.numpy(),
+            (-2 * y * (1 - y**2)).reshape(x.shape).numpy(),
+            1e-6,
+            1e-6,
+        )
+        dxxx = paddle.grad(dxx, x, create_graph=True)[0]
+        np.testing.assert_allclose(
+            dxxx.numpy(),
+            (-2 * (1 - y**2) ** 2 + -2 * y * (-2 * y) * (1 - y**2))
+            .reshape(x.shape)
+            .numpy(),
+            1e-6,
+            1e-6,
+        )
+
     def call_view_as(self):
         x_np = np.random.random(size=[10, 10, 10, 20]).astype('float32')
         x = paddle.to_tensor(x_np)
