update

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

@@ -28,6 +28,7 @@ namespace = ""
 yaml_types_mapping = {
     'int' : 'int', 'int32' : 'int32_t', 'int64' : 'int64_t',  'size_t' : 'size_t', \
     'float' : 'float', 'double' : 'double', 'bool' : 'bool', \
+    'str' : 'std::string', \
     'Backend' : 'paddle::experimental::Backend', 'DataLayout' : 'paddle::experimental::DataLayout', 'DataType' : 'paddle::experimental::DataType', \
     'int64[]' : 'std::vector<int64_t>', 'int[]' : 'std::vector<int>',
     'Tensor' : 'Tensor',
@@ -235,7 +236,7 @@ def ParseYamlReturns(string):
         else:
             ret_type = ret.strip()
 
-        assert ret_type in yaml_types_mapping.keys()
+        assert ret_type in yaml_types_mapping.keys(), ret_type
         ret_type = yaml_types_mapping[ret_type]
 
         assert "Tensor" in ret_type
@@ -426,7 +427,7 @@ def SlotNameMatching(backward_inputs_list, backward_returns_list,
                     backward_input_type, False, backward_input_pos
                 ]
             else:
-                assert False
+                assert False, backward_input_name
 
     for backward_output in backward_returns_list:
         backward_output_name = backward_output[0]
@@ -435,7 +436,8 @@ def SlotNameMatching(backward_inputs_list, backward_returns_list,
 
         backward_fwd_name = FindForwardName(backward_output_name)
         assert backward_fwd_name is not None
-        assert backward_fwd_name in forward_inputs_position_map.keys()
+        assert backward_fwd_name in forward_inputs_position_map.keys(
+        ), backward_fwd_name
 
         matched_forward_input_type = forward_inputs_position_map[
             backward_fwd_name][0]
@@ -684,10 +686,10 @@ def GenerateNodeCreationCodes(
         else:
             # Tuple api_result
             if IsPlainTensorType(rtype):
-                output_autograd_meta = f"    egr::AutogradMeta* {output_autograd_meta_name} = egr::EagerUtils::autograd_meta(&api_result[{pos}]);"
+                output_autograd_meta = f"    egr::AutogradMeta* {output_autograd_meta_name} = egr::EagerUtils::autograd_meta(&std::get<{pos}>(api_result));"
             else:
                 assert IsVectorTensorType(rtype)
-                output_autograd_meta = f"    std::vector<egr::AutogradMeta*> {output_autograd_meta_vec_name} = egr::EagerUtils::autograd_meta(&api_result[{pos}]);\n"
+                output_autograd_meta = f"    std::vector<egr::AutogradMeta*> {output_autograd_meta_vec_name} = egr::EagerUtils::autograd_meta(&std::get<{pos}>(api_result));\n"
                 output_autograd_meta += f"    std::vector<egr::AutogradMeta*>* {output_autograd_meta_name} = &{output_autograd_meta_vec_name};"
 
         outputs_autograd_meta_list.append(output_autograd_meta)
@@ -1198,7 +1200,7 @@ if __name__ == "__main__":
             print("Generated Backward Grad Output Map: ",
                   backward_grad_output_map)
 
-            # Backward Validation Check
+            # Backward Validation Check            
             BackwardValidationCheck(backward_fwd_input_map,
                                     backward_grad_input_map,
                                     backward_attrs_list)

paddle/fluid/eager/auto_code_generator/final_state_generator/python_c_gen.py

@@ -22,6 +22,7 @@ atype_to_parsing_function = {
     "bool": "CastPyArg2Boolean",
     "int": "CastPyArg2Int",
     "long": "CastPyArg2Long",
+    "std::string": "CastPyArgs2String",
     "int64_t": "CastPyArg2Long",
     "float": "CastPyArg2Float",
     "string": "CastPyArg2String",

paddle/phi/kernels/impl/trace_grad_kernel_impl.h

@@ -82,8 +82,8 @@ struct TraceGradFunctor {
 
 template <typename T, typename Context>
 void TraceGradKernel(const Context& ctx,
-                     const DenseTensor& out_grad,
                      const DenseTensor& x,
+                     const DenseTensor& out_grad,
                      int offset,
                      int axis1,
                      int axis2,

paddle/phi/kernels/trace_grad_kernel.h

@@ -20,8 +20,8 @@ namespace phi {
 
 template <typename T, typename Context>
 void TraceGradKernel(const Context& ctx,
-                     const DenseTensor& out_grad,
                      const DenseTensor& x,
+                     const DenseTensor& out_grad,
                      int offset,
                      int axis1,
                      int axis2,

paddle/phi/ops/compat/trace_sig.cc

@@ -23,7 +23,7 @@ KernelSignature TraceOpArgumentMapping(const ArgumentMappingContext& ctx) {
 
 KernelSignature TraceGradOpArgumentMapping(const ArgumentMappingContext& ctx) {
   return KernelSignature("trace_grad",
-                         {GradVarName("Out"), "Input"},
+                         {"Input", GradVarName("Out")},
                          {"offset", "axis1", "axis2"},
                          {GradVarName("Input")});
 }

python/paddle/fluid/layers/metric_op.py

@@ -87,6 +87,9 @@ def accuracy(input, label, k=1, correct=None, total=None):
         _k = k.numpy().item(0) if isinstance(k, Variable) else k
         topk_out, topk_indices = _C_ops.top_k_v2(input, 'k', _k, 'sorted',
                                                  False)
+        if _in_eager_mode:
+            _acc = _C_ops.final_state_accuracy(topk_out, topk_indices, label)
+            return _acc
         _acc, _, _ = _C_ops.accuracy(topk_out, topk_indices, label, correct,
                                      total)
         return _acc

python/paddle/fluid/layers/nn.py

@@ -14827,6 +14827,11 @@ def unfold(x, kernel_sizes, strides=1, paddings=0, dilations=1, name=None):
             "Unexpected type of paddings, it should be either an integer or a list"
             "of 2 or 4 integers")
 
+    if in_dygraph_mode():
+        if _in_eager_mode():
+            return _C_op.final_state_unfold(x, kernel_sizes, strdides, paddings,
+                                            dilations)
+
     out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type="unfold",

python/paddle/fluid/tests/unittests/op_test.py

@@ -545,7 +545,8 @@ class OpTest(unittest.TestCase):
             v.value().get_tensor().set_recursive_sequence_lengths(lod)
             return v
         else:
-            return fluid.dygraph.base.to_variable(value)
+            print("init her")
+            return paddle.to_tensor(value)
 
     def get_sequence_batch_size_1_input(self, lod=None, shape=None):
         """Get LoD input data whose batch size is 1.

python/paddle/fluid/tests/unittests/test_accuracy_op.py

@@ -25,6 +25,7 @@ from paddle.fluid import compiler, Program, program_guard
 class TestAccuracyOp(OpTest):
     def setUp(self):
         self.op_type = "accuracy"
+        self.python_api = paddle.metric.accuracy
         self.dtype = np.float32
         self.init_dtype()
         n = 8192
@@ -48,7 +49,7 @@ class TestAccuracyOp(OpTest):
         pass
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_eager=False)
 
 
 class TestAccuracyOpFp16(TestAccuracyOp):
@@ -128,4 +129,5 @@ class TestAccuracyAPI(unittest.TestCase):
 
 
 if __name__ == '__main__':
+    paddle.enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/test_arg_min_max_op.py

@@ -26,6 +26,7 @@ from paddle.fluid import Program, program_guard
 class BaseTestCase(OpTest):
     def initTestCase(self):
         self.op_type = 'arg_min'
+        self.python_api = paddle.argmin
         self.dims = (3, 4, 5)
         self.dtype = 'float32'
         self.axis = 0
@@ -41,12 +42,13 @@ class BaseTestCase(OpTest):
             self.outputs = {'Out': np.argmax(self.x, axis=self.axis)}
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_eager=True)
 
 
 class TestCase0(BaseTestCase):
     def initTestCase(self):
         self.op_type = 'arg_max'
+        self.python_api = paddle.argmax
         self.dims = (3, 4, 5)
         self.dtype = 'float32'
         self.axis = 0
@@ -63,6 +65,7 @@ class TestCase1(BaseTestCase):
 class TestCase2(BaseTestCase):
     def initTestCase(self):
         self.op_type = 'arg_max'
+        self.python_api = paddle.argmax
         self.dims = (3, 4)
         self.dtype = 'int64'
         self.axis = 0
@@ -71,6 +74,7 @@ class TestCase2(BaseTestCase):
 class TestCase2_1(BaseTestCase):
     def initTestCase(self):
         self.op_type = 'arg_max'
+        self.python_api = paddle.argmax
         self.dims = (3, 4)
         self.dtype = 'int64'
         self.axis = -1
@@ -95,6 +99,7 @@ class TestCase4(BaseTestCase):
 class TestCase3_(BaseTestCase):
     def initTestCase(self):
         self.op_type = 'arg_max'
+        self.python_api = paddle.argmax
         self.dims = (3, )
         self.axis = 0
 
@@ -152,6 +157,7 @@ class BaseTestComplex1_2(OpTest):
 class BaseTestComplex2_1(OpTest):
     def initTestCase(self):
         self.op_type = 'arg_max'
+        self.python_api = paddle.argmax
         self.dims = (4, 5, 6)
         self.dtype = 'int32'
         self.axis = 2
@@ -202,4 +208,5 @@ class BaseTestComplex2_2(OpTest):
 
 
 if __name__ == '__main__':
+    paddle.enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/test_argsort_op.py

@@ -23,7 +23,7 @@ import six
 import paddle.fluid.core as core
 
 from paddle.fluid import ParamAttr
-from paddle.fluid.framework import Program, grad_var_name
+from paddle.fluid.framework import Program, grad_var_name, _test_eager_guard
 from paddle.fluid.executor import Executor
 from paddle.fluid.backward import append_backward
 
@@ -421,6 +421,16 @@ class TestArgsortImperative(unittest.TestCase):
         expect2 = np.argsort(-self.input_data, axis=self.axis)
         self.assertEqual((expect2 == out2.numpy()).all(), True)
 
+        with _test_eager_guard():
+            var_x = paddle.to_tensor(self.input_data)
+            out = paddle.argsort(var_x, axis=self.axis)
+            expect = np.argsort(self.input_data, axis=self.axis)
+            self.assertEqual((expect == out.numpy()).all(), True)
+
+            out2 = paddle.argsort(var_x, axis=self.axis, descending=True)
+            expect2 = np.argsort(-self.input_data, axis=self.axis)
+            self.assertEqual((expect2 == out2.numpy()).all(), True)
+
         paddle.enable_static()
 
 
@@ -443,4 +453,5 @@ class TestArgsortImperative4(TestArgsortImperative):
 
 
 if __name__ == "__main__":
+    paddle.enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/test_bitwise_op.py

@@ -276,6 +276,7 @@ class TestBitwiseXorBool(TestBitwiseXor):
 class TestBitwiseNot(OpTest):
     def setUp(self):
         self.op_type = "bitwise_not"
+        self.python_api = paddle.bitwise_not
         self.init_dtype()
         self.init_shape()
         self.init_bound()
@@ -288,7 +289,7 @@ class TestBitwiseNot(OpTest):
         self.outputs = {'Out': out}
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
         pass
@@ -351,4 +352,5 @@ class TestBitwiseNotBool(TestBitwiseNot):
 
 
 if __name__ == "__main__":
+    paddle.enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/test_cholesky_op.py

@@ -36,6 +36,7 @@ from decorator_helper import prog_scope
 class TestCholeskyOp(OpTest):
     def setUp(self):
         self.op_type = "cholesky"
+        self.python_api = paddle.linalg.cholesky
         self._input_shape = (2, 32, 32)
         self._upper = True
         self.init_config()
@@ -54,7 +55,7 @@ class TestCholeskyOp(OpTest):
         self.outputs = {"Out": output_data}
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_eager=True)
 
     def test_check_grad(self):
         places = [fluid.CPUPlace()]

python/paddle/fluid/tests/unittests/test_cholesky_solve_op.py

@@ -109,6 +109,7 @@ class TestCholeskySolveOp(OpTest):
 
     def setUp(self):
         self.op_type = "cholesky_solve"
+        self.python_api = paddle.linalg.cholesky_solve
         self.config()
 
         if self.upper:
@@ -125,137 +126,134 @@ class TestCholeskySolveOp(OpTest):
         self.outputs = {'Out': self.output}
 
     def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad_normal(self):
-        self.check_grad(['Y'], 'Out', max_relative_error=0.01)
-
-
-# 3D(broadcast) + 3D, upper=True
-class TestCholeskySolveOp3(TestCholeskySolveOp):
-    """
-    case 3
-    """
-
-    def config(self):
-        self.y_shape = [1, 10, 10]
-        self.x_shape = [2, 10, 5]
-        self.upper = True
-        self.dtype = np.float64
-
-
-class TestCholeskySolveAPI(unittest.TestCase):
-    def setUp(self):
-        np.random.seed(2021)
-        self.place = [paddle.CPUPlace()]
-        # self.place = [paddle.CUDAPlace(0)]
-        self.dtype = "float64"
-        self.upper = True
-        if core.is_compiled_with_cuda():
-            self.place.append(paddle.CUDAPlace(0))
-
-    def check_static_result(self, place):
-        paddle.enable_static()
-        with fluid.program_guard(fluid.Program(), fluid.Program()):
-            x = fluid.data(name="x", shape=[10, 2], dtype=self.dtype)
-            y = fluid.data(name="y", shape=[10, 10], dtype=self.dtype)
-            z = paddle.linalg.cholesky_solve(x, y, upper=self.upper)
-
-            x_np = np.random.random([10, 2]).astype(self.dtype)
-            y_np = np.random.random([10, 10]).astype(self.dtype)
-            if self.upper:
-                umat = np.triu(y_np)
-            else:
-                umat = np.tril(y_np)
-            z_np = cholesky_solution(umat, x_np, upper=self.upper)
-            z2_np = scipy_cholesky_solution(umat, x_np, upper=self.upper)
-
-            exe = fluid.Executor(place)
-            fetches = exe.run(fluid.default_main_program(),
-                              feed={"x": x_np,
-                                    "y": umat},
-                              fetch_list=[z])
-            self.assertTrue(np.allclose(fetches[0], z_np))
-
-    def test_static(self):
-        for place in self.place:
-            self.check_static_result(place=place)
-
-    def test_dygraph(self):
-        def run(place):
-            paddle.disable_static(place)
-            x_np = np.random.random([20, 2]).astype(self.dtype)
-            y_np = np.random.random([20, 20]).astype(self.dtype)
-            z_np = scipy_cholesky_solution(y_np, x_np, upper=self.upper)
-
-            x = paddle.to_tensor(x_np)
-            y = paddle.to_tensor(y_np)
-            z = paddle.linalg.cholesky_solve(x, y, upper=self.upper)
-
-            self.assertTrue(np.allclose(z_np, z.numpy()))
-            self.assertEqual(z_np.shape, z.numpy().shape)
-            paddle.enable_static()
-
-        for idx, place in enumerate(self.place):
-            run(place)
-
-    def test_boardcast(self):
-        def run(place):
-            paddle.disable_static()
-            x_np = np.random.random([1, 30, 2]).astype(self.dtype)
-            y_np = np.random.random([2, 30, 30]).astype(self.dtype)
-            nx_np = np.concatenate((x_np, x_np), axis=0)
-
-            z_sci = scipy_cholesky_solution_batch(y_np, nx_np, upper=self.upper)
-
-            x = paddle.to_tensor(x_np)
-            y = paddle.to_tensor(y_np)
-            z = paddle.linalg.cholesky_solve(x, y, upper=self.upper)
-            self.assertEqual(z_sci.shape, z.numpy().shape)
-            self.assertTrue(np.allclose(z_sci, z.numpy()))
-
-        for idx, place in enumerate(self.place):
-            run(place)
-
-
-class TestCholeskySolveOpError(unittest.TestCase):
-    def test_errors(self):
-        paddle.enable_static()
-        with program_guard(Program(), Program()):
-            # The input type of solve_op must be Variable.
-            x1 = fluid.create_lod_tensor(
-                np.array([[-1]]), [[1]], fluid.CPUPlace())
-            y1 = fluid.create_lod_tensor(
-                np.array([[-1]]), [[1]], fluid.CPUPlace())
-            self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x1, y1)
-
-            # The data type of input must be float32 or float64.        
-            x2 = fluid.data(name="x2", shape=[30, 30], dtype="bool")
-            y2 = fluid.data(name="y2", shape=[30, 10], dtype="bool")
-            self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x2, y2)
-
-            x3 = fluid.data(name="x3", shape=[30, 30], dtype="int32")
-            y3 = fluid.data(name="y3", shape=[30, 10], dtype="int32")
-            self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x3, y3)
-
-            x4 = fluid.data(name="x4", shape=[30, 30], dtype="float16")
-            y4 = fluid.data(name="y4", shape=[30, 10], dtype="float16")
-            self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x4, y4)
-
-            # The number of dimensions of input'X must be >= 2.
-            x5 = fluid.data(name="x5", shape=[30], dtype="float64")
-            y5 = fluid.data(name="y5", shape=[30, 30], dtype="float64")
-            self.assertRaises(ValueError, paddle.linalg.cholesky_solve, x5, y5)
-
-            # The number of dimensions of input'Y must be >= 2.
-            x6 = fluid.data(name="x6", shape=[30, 30], dtype="float64")
-            y6 = fluid.data(name="y6", shape=[30], dtype="float64")
-            self.assertRaises(ValueError, paddle.linalg.cholesky_solve, x6, y6)
-
-            # The inner-most 2 dimensions of input'X should be equal to each other
-            x7 = fluid.data(name="x7", shape=[2, 3, 4], dtype="float64")
-            y7 = fluid.data(name="y7", shape=[2, 4, 3], dtype="float64")
-            self.assertRaises(ValueError, paddle.linalg.cholesky_solve, x7, y7)
+        self.check_output(check_eager=True)
+
+    # def test_check_grad_normal(self):
+    #     self.check_grad(['Y'], 'Out', max_relative_error=0.01, check_eager=True)
+
+    # # 3D(broadcast) + 3D, upper=True
+    # class TestCholeskySolveOp3(TestCholeskySolveOp):
+    #     """
+    #     case 3
+    #     """
+
+    #     def config(self):
+    #         self.y_shape = [1, 10, 10]
+    #         self.x_shape = [2, 10, 5]
+    #         self.upper = True
+    #         self.dtype = np.float64
+
+    # class TestCholeskySolveAPI(unittest.TestCase):
+    #     def setUp(self):
+    #         np.random.seed(2021)
+    #         self.place = [paddle.CPUPlace()]
+    #         # self.place = [paddle.CUDAPlace(0)]
+    #         self.dtype = "float64"
+    #         self.upper = True
+    #         if core.is_compiled_with_cuda():
+    #             self.place.append(paddle.CUDAPlace(0))
+
+    #     def check_static_result(self, place):
+    #         paddle.enable_static()
+    #         with fluid.program_guard(fluid.Program(), fluid.Program()):
+    #             x = fluid.data(name="x", shape=[10, 2], dtype=self.dtype)
+    #             y = fluid.data(name="y", shape=[10, 10], dtype=self.dtype)
+    #             z = paddle.linalg.cholesky_solve(x, y, upper=self.upper)
+
+    #             x_np = np.random.random([10, 2]).astype(self.dtype)
+    #             y_np = np.random.random([10, 10]).astype(self.dtype)
+    #             if self.upper:
+    #                 umat = np.triu(y_np)
+    #             else:
+    #                 umat = np.tril(y_np)
+    #             z_np = cholesky_solution(umat, x_np, upper=self.upper)
+    #             z2_np = scipy_cholesky_solution(umat, x_np, upper=self.upper)
+
+    #             exe = fluid.Executor(place)
+    #             fetches = exe.run(fluid.default_main_program(),
+    #                               feed={"x": x_np,
+    #                                     "y": umat},
+    #                               fetch_list=[z])
+    #             self.assertTrue(np.allclose(fetches[0], z_np))
+
+    #     def test_static(self):
+    #         for place in self.place:
+    #             self.check_static_result(place=place)
+
+    #     def test_dygraph(self):
+    #         def run(place):
+    #             paddle.disable_static(place)
+    #             x_np = np.random.random([20, 2]).astype(self.dtype)
+    #             y_np = np.random.random([20, 20]).astype(self.dtype)
+    #             z_np = scipy_cholesky_solution(y_np, x_np, upper=self.upper)
+
+    #             x = paddle.to_tensor(x_np)
+    #             y = paddle.to_tensor(y_np)
+    #             z = paddle.linalg.cholesky_solve(x, y, upper=self.upper)
+
+    #             self.assertTrue(np.allclose(z_np, z.numpy()))
+    #             self.assertEqual(z_np.shape, z.numpy().shape)
+    #             paddle.enable_static()
+
+    #         for idx, place in enumerate(self.place):
+    #             run(place)
+
+    #     def test_boardcast(self):
+    #         def run(place):
+    #             paddle.disable_static()
+    #             x_np = np.random.random([1, 30, 2]).astype(self.dtype)
+    #             y_np = np.random.random([2, 30, 30]).astype(self.dtype)
+    #             nx_np = np.concatenate((x_np, x_np), axis=0)
+
+    #             z_sci = scipy_cholesky_solution_batch(y_np, nx_np, upper=self.upper)
+
+    #             x = paddle.to_tensor(x_np)
+    #             y = paddle.to_tensor(y_np)
+    #             z = paddle.linalg.cholesky_solve(x, y, upper=self.upper)
+    #             self.assertEqual(z_sci.shape, z.numpy().shape)
+    #             self.assertTrue(np.allclose(z_sci, z.numpy()))
+
+    #         for idx, place in enumerate(self.place):
+    #             run(place)
+
+    # class TestCholeskySolveOpError(unittest.TestCase):
+    #     def test_errors(self):
+    #         paddle.enable_static()
+    #         with program_guard(Program(), Program()):
+    #             # The input type of solve_op must be Variable.
+    #             x1 = fluid.create_lod_tensor(
+    #                 np.array([[-1]]), [[1]], fluid.CPUPlace())
+    #             y1 = fluid.create_lod_tensor(
+    #                 np.array([[-1]]), [[1]], fluid.CPUPlace())
+    #             self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x1, y1)
+
+    #             # The data type of input must be float32 or float64.        
+    #             x2 = fluid.data(name="x2", shape=[30, 30], dtype="bool")
+    #             y2 = fluid.data(name="y2", shape=[30, 10], dtype="bool")
+    #             self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x2, y2)
+
+    #             x3 = fluid.data(name="x3", shape=[30, 30], dtype="int32")
+    #             y3 = fluid.data(name="y3", shape=[30, 10], dtype="int32")
+    #             self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x3, y3)
+
+    #             x4 = fluid.data(name="x4", shape=[30, 30], dtype="float16")
+    #             y4 = fluid.data(name="y4", shape=[30, 10], dtype="float16")
+    #             self.assertRaises(TypeError, paddle.linalg.cholesky_solve, x4, y4)
+
+    #             # The number of dimensions of input'X must be >= 2.
+    #             x5 = fluid.data(name="x5", shape=[30], dtype="float64")
+    #             y5 = fluid.data(name="y5", shape=[30, 30], dtype="float64")
+    #             self.assertRaises(ValueError, paddle.linalg.cholesky_solve, x5, y5)
+
+    #             # The number of dimensions of input'Y must be >= 2.
+    #             x6 = fluid.data(name="x6", shape=[30, 30], dtype="float64")
+    #             y6 = fluid.data(name="y6", shape=[30], dtype="float64")
+    #             self.assertRaises(ValueError, paddle.linalg.cholesky_solve, x6, y6)
+
+    #             # The inner-most 2 dimensions of input'X should be equal to each other
+    #             x7 = fluid.data(name="x7", shape=[2, 3, 4], dtype="float64")
+    #             y7 = fluid.data(name="y7", shape=[2, 4, 3], dtype="float64")
+    #             self.assertRaises(ValueError, paddle.linalg.cholesky_solve, x7, y7)
 
 
 if __name__ == "__main__":

python/paddle/fluid/tests/unittests/test_softmax_op.py

@@ -53,6 +53,7 @@ class TestSoftmaxOp(OpTest):
 
     def setUp(self):
         self.op_type = "softmax"
+        self.python_api = paddle.nn.functional.softmax
         self.use_cudnn = False
         self.use_mkldnn = False
         # explicilty use float32 for ROCm, as MIOpen does not yet support float64
@@ -81,9 +82,13 @@ class TestSoftmaxOp(OpTest):
         if self.use_cudnn:
             place = core.CUDAPlace(0)
             self.check_output_with_place(
-                place, atol=1e-5, check_dygraph=(self.use_mkldnn == False))
+                place,
+                atol=1e-5,
+                check_dygraph=(self.use_mkldnn == False),
+                check_eager=False)
         else:
-            self.check_output(check_dygraph=(self.use_mkldnn == False))
+            self.check_output(
+                check_dygraph=(self.use_mkldnn == False), check_eager=False)
 
     def test_check_grad(self):
         # TODO(wangzhongpu): support mkldnn op in dygraph mode

python/paddle/fluid/tests/unittests/test_unfold_op.py

@@ -21,6 +21,7 @@ from op_test import OpTest
 import paddle
 import paddle.fluid as fluid
 from paddle.fluid import core
+from paddle.fluid.framework import _test_eager_guard
 
 
 class TestUnfoldOp(OpTest):
@@ -37,6 +38,7 @@ class TestUnfoldOp(OpTest):
         self.strides = [1, 1]
         self.paddings = [1, 1, 1, 1]
         self.dilations = [1, 1]
+        self.python_api = paddle.nn.functional.unfold
         input_shape = [
             self.batch_size, self.input_channels, self.input_height,
             self.input_width
@@ -95,10 +97,10 @@ class TestUnfoldOp(OpTest):
         self.set_data()
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_eager=True)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Y')
+        self.check_grad(['X'], 'Y', check_eager=True)
 
 
 class TestUnfoldAPI(TestUnfoldOp):
@@ -122,9 +124,18 @@ class TestUnfoldAPI(TestUnfoldOp):
                 result = m(input)
                 self.assertTrue(np.allclose(result.numpy(), self.outputs['Y']))
 
+                with _test_eager_guard():
+                    input = fluid.dygraph.to_variable(self.inputs['X'])
+                    m = paddle.nn.Unfold(**self.attrs)
+                    m.eval()
+                    result = m(input)
+                    self.assertTrue(
+                        np.allclose(result.numpy(), self.outputs['Y']))
+
     def test_info(self):
         str(paddle.nn.Unfold(**self.attrs))
 
 
 if __name__ == '__main__':
+    paddle.enable_static()
     unittest.main()

python/paddle/metric/metrics.py

@@ -798,8 +798,12 @@ def accuracy(input, label, k=1, correct=None, total=None, name=None):
             total = _varbase_creator(dtype="int32")
 
         topk_out, topk_indices = paddle.topk(input, k=k)
+        if _in_eager_mode:
+            _acc = _C_ops.final_state_accuracy(topk_out, topk_indices, label)
+            return _acc
         _acc, _, _ = _C_ops.accuracy(topk_out, topk_indices, label, correct,
                                      total)
+
         return _acc
 
     helper = LayerHelper("accuracy", **locals())

python/paddle/nn/functional/activation.py

@@ -561,6 +561,8 @@ def relu(x, name=None):
     """
 
     if in_dynamic_mode():
+        if _in_eager_mode():
+            return _C_ops.final_state_relu(x)
         return _C_ops.relu(x)
 
     check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'relu')
@@ -954,6 +956,8 @@ def softmax(x, axis=-1, dtype=None, name=None):
     if in_dynamic_mode():
         outs_cast = x if dtype is None \
             else _C_ops.cast(x, 'in_dtype', x.dtype, 'out_dtype', dtype)
+        if _in_eager_mode():
+            return _C_ops.final_state_softmax(outs_cast, axis)
         return _C_ops.softmax(outs_cast, 'axis', axis, 'use_cudnn', use_cudnn)
 
     if dtype is None:

python/paddle/tensor/linalg.py

@@ -14,7 +14,7 @@
 
 import numpy as np
 from ..fluid.layer_helper import LayerHelper
-from ..framework import _varbase_creator, _dygraph_tracer
+from ..framework import _varbase_creator, _dygraph_tracer, _in_eager_mode
 from ..fluid.data_feeder import check_variable_and_dtype, check_type, check_dtype
 from ..static import Variable
 
@@ -1205,6 +1205,8 @@ def cholesky(x, upper=False, name=None):
 
     """
     if paddle.in_dynamic_mode():
+        if _in_eager_mode():
+            return _C_ops.final_state_cholesky(x, upper)
         return _C_ops.cholesky(x, "upper", upper)
     check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'cholesky')
     check_type(upper, 'upper', bool, 'cholesky')
@@ -2734,6 +2736,8 @@ def cholesky_solve(x, y, upper=False, name=None):
         # [-2.5, -7, 9.5]
     """
     if paddle.in_dynamic_mode():
+        if _in_eager_mode():
+            return _C_ops.final_state_cholesky_solve(x, y, upper)
         return _C_ops.cholesky_solve(x, y, 'upper', upper)
 
     helper = LayerHelper("cholesky_solve", **locals())

python/paddle/tensor/logic.py

@@ -17,6 +17,7 @@ from ..fluid.data_feeder import check_type, check_variable_and_dtype
 from ..fluid.layers.layer_function_generator import templatedoc
 from ..static import Variable
 from ..framework import VarBase as Tensor
+from paddle.fluid.framework import _in_eager_mode
 
 # TODO: define logic functions of a tensor  
 from ..fluid.layers import is_empty  # noqa: F401
@@ -452,6 +453,8 @@ def _bitwise_op(op_name, x, y, out=None, name=None, binary_op=True):
         if binary_op:
             return op(x, y)
         else:
+            if _in_eager_mode():
+                return _C_op.final_state_bitewise_not(x)
             return op(x)
 
     check_variable_and_dtype(

python/paddle/tensor/math.py

@@ -2853,6 +2853,8 @@ def tanh(x, name=None):
             # [-0.37994896 -0.19737532  0.09966799  0.29131261]
     """
     if paddle.in_dynamic_mode():
+        if _in_eager_mode():
+            return _C_ops.final_state_tanh( x )
         return _C_ops.tanh(x)
 
     check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'tanh')

python/paddle/tensor/search.py

@@ -91,6 +91,8 @@ def argsort(x, axis=-1, descending=False, name=None):
             #  [0 2 1 1]]]
     """
     if paddle.in_dynamic_mode():
+        if _in_eager_mode():
+            _, ids, = _C_ops.final_state_argsort(x, axis, descending)
         _, ids = _C_ops.argsort(x, 'axis', axis, 'descending', descending)
         return ids
     check_variable_and_dtype(
@@ -245,6 +247,9 @@ def argmin(x, axis=None, keepdim=False, dtype="int64", name=None):
         axis = 0
 
     if paddle.in_dynamic_mode():
+        if _in_eager_mode():
+            out = _C_ops.final_state_arg_min(x, axis, keepdim, flattern,
+                                             var_dtype)
         out = _C_ops.arg_min(x, 'axis', axis, 'dtype', var_dtype, 'keepdims',
                              keepdim, 'flatten', flatten)
         return out

python/paddle/utils/code_gen/api.yaml

@@ -262,56 +262,59 @@
 #     func : maxout
 #   backward : maxout_grad
 
-# # put_along_axis
-# - api : put_along_axis
-#   args : (Tensor x, Tensor index, Tensor value, int axis, string reduce)
-#   output : Tensor
-#   infer_meta :
-#     func : PutAlongAxisInferMeta
-#   kernel :
-#     func : put_along_axis
-#   backward : put_along_axis_grad
+# put_along_axis
+- api : put_along_axis
+  args : (Tensor x, Tensor index, Tensor value, int axis, str reduce)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [index]
+  kernel :
+    func : put_along_axis
+  backward : put_along_axis_grad
 
 
-# # take_along_axis
-# - api : take_along_axis
-#   args : (Tensor x, Tensor index, int axis)
-#   output : Tensor
-#   infer_meta :
-#     func : TakeAlongAxisInferMeta
-#   kernel :
-#     func : take_along_axis
-#   backward : take_along_axis_grad
+# take_along_axis
+- api : take_along_axis
+  args : (Tensor x, Tensor index, int axis)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [index]
+  kernel :
+    func : take_along_axis
+  backward : take_along_axis_grad
 
-# # matrix_power
-# - api : maxtrix_power
-#   args : (Tensor x, int n)
-#   output : Tensor
-#   infer_meta :
-#     func : MaxtrixPowerInferMeta
-#   kernel :
-#     func : maxtrix_power
-#   backward : maxtrix_power_grad
+# matrix_power
+- api : matrix_power
+  args : (Tensor x, int n)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : matrix_power
+  backward : matrix_power_grad
 
-# # eigh
-# - api : eigh
-#   args : (Tensor x, string uplo)
-#   output : Tensor(out_w), Tensor(out_v)
-#   infer_meta :
-#     func : EighInferMeta
-#   kernel :
-#     func : eigh
-#   backward : eigh_grad
+# eigh
+- api : eigh
+  args : (Tensor x, str uplo)
+  output : Tensor(out_w), Tensor(out_v)
+  infer_meta :
+    func : EighInferMeta
+  kernel :
+    func : eigh
+  backward : eigh_grad
 
-# # segment_pool
-# - api : segment_pool
-#   args : (Tensor x, Tensor segment_ids, string pooltype)
-#   output : Tensor(out), Tensor(summed_ids)
-#   infer_meta :
-#     func : SegmentPoolInferMeta
-#   kernel :
-#     func : segment_pool
-#   backward : segment_pool_grad
+# segment_pool
+- api : segment_pool
+  args : (Tensor x, Tensor segment_ids, str pooltype)
+  output : Tensor(out), Tensor(summed_ids)
+  infer_meta :
+    func : SegmentPoolInferMeta
+  kernel :
+    func : segment_pool
+  backward : segment_pool_grad
 
 # accuracy
 - api : accuracy
@@ -447,44 +450,45 @@
 
 
 # arg_min  # int64 ???? dtype
-- api : agrmin
-  args : (Tensor x, int axis, bool keepdims, bool flatten, int dtype)
+- api : arg_min
+  args : (Tensor x, int64 axis, bool keepdims, bool flatten, int dtype)
   output : Tensor
   infer_meta :
     func : ArgMinMaxInferMeta
   kernel :
-    func : argmin
+    func : arg_min
 
 # arg_max  # int64 ???? dtype
-- api : agrmax
-  args : (Tensor x, int axis, bool keepdims, bool flatten, int dtype)
+- api : arg_max
+  args : (Tensor x, int64 axis, bool keepdims, bool flatten, int dtype)
   output : Tensor
   infer_meta :
     func : ArgMinMaxInferMeta
   kernel :
-    func : argmax
+    func : arg_max
 
-# argsort; kernel why input?
+# argsort
 - api : argsort
   args : (Tensor x, int axis, bool descending)
   output : Tensor(out), Tensor(indices)
   infer_meta :
-    func : ArgsortInferMeta
+    func : ArgsortInferMeta 
   kernel :
     func : argsort
+  backward : argsort_grad
 
-# auc; kernel why input?
-# - api : auc
-#   args : (Tensor x, Tensor label, Tensor stat_pos, Tensor stat_neg, str curve, int num_thresholds, int slide_steps)
-#   output : Tensor(auc), Tensor(stat_pos_out), Tensor(stat_neg_out)
-#   infer_meta :
-#     func : AucInferMeta
-#   kernel :
-#     func : auc
+# auc
+- api : auc
+  args : (Tensor x, Tensor label, Tensor stat_pos, Tensor stat_neg, str curve, int num_thresholds, int slide_steps)
+  output : Tensor(auc), Tensor(stat_pos_out), Tensor(stat_neg_out)
+  infer_meta :
+    func : AucInferMeta
+  kernel :
+    func : auc
 
 # # batch_norm
 # - api : batch_norm
-#   args : (Tensor x, Tensor scale, Tensor bias, Tensor mean, Tensor variance, float momentum, float epsilon, string data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu)
+#   args : (Tensor x, Tensor scale, Tensor bias, Tensor mean, Tensor variance, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu)
 #   output : Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
 #   infer_meta :
 #     func : XXXXInferMeta
@@ -492,14 +496,14 @@
 #     func : batch_norm
 #   backward: batch_norm_grad
 
-# # bernoulli
-# - api : bernoulli
-#   args : (Tensor x)
-#   output : Tensor
-#   infer_meta :
-#     func : BernoulliInferMeta
-#   kernel :
-#     func : bernoulli
+# bernoulli
+- api : bernoulli
+  args : (Tensor x)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+  kernel :
+    func : bernoulli
 
 # # bilinear_tensor_product ?? optional
 # - api : bilinear_tensor_product
@@ -510,15 +514,18 @@
 #   kernel :
 #     func : bilinear_tensor_product
 #   backward : bilinear_tensor_product_grad
+#   optional : bias
+
+# bincount
+- api : bincount
+  args : (Tensor x, Tensor weight, int minlength)
+  output : Tensor
+  infer_meta :
+    func : BincountInferMeta
+  kernel :
+    func : bincount
+  optional : weight
 
-# bincount ?? optional
-# - api : bincount
-#   args : (Tensor x, Tensor weight, int minlength)
-#   output : Tensor
-#   infer_meta :
-#     func : BincountInferMeta
-#   kernel :
-#     func : bincount
 
 
 # bitwise_and
@@ -604,7 +611,7 @@
 # # depthwise_conv2d
 # # dropout ?? optional, intermediate
 # - api : dropout
-#   args : (Tensor x, Tensor seed_tensor, float p, bool is_test, string mode, int seed, bool fix_seed)
+#   args : (Tensor x, Tensor seed_tensor, float p, bool is_test, str mode, int seed, bool fix_seed)
 #   output : Tensor(out), Tensor(mask)
 #   infer_meta :
 #     func : DropoutInferMeta
@@ -613,27 +620,27 @@
 
 # # embedding
 
-# # erf
-# - api : erf
-#   args : (Tensor x)
-#   output : Tensor
-#   infer_meta :
-#     func : ErfInferMeta
-#   kernel :
-#     func : erf
-#   backward : erf_grad
+# erf
+- api : erf
+  args : (Tensor x)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+  kernel :
+    func : erf
+  backward : erf_grad
 
-# # erfinv
-# - api : erfinv
-#   args : (Tensor x)
-#   output : Tensor
-#   infer_meta :
-#     func : ErfinvInferMeta
-#   kernel :
-#     func : erfinv
-#   backward : erfinv_grad
+# erfinv
+- api : erfinv
+  args : (Tensor x)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+  kernel :
+    func : erfinv
+  backward : erfinv_grad
 
-# # expand_as ?? optional 
+# expand_as
 # - api : expand_as
 #   args : (Tensor x, Tensor y, int[] shape)
 #   output : Tensor
@@ -642,6 +649,7 @@
 #   kernel :
 #     func : expand_as
 #   backward : expand_as_grad
+#   optional : y
 
 
 # # expand
@@ -655,13 +663,13 @@
 #   backward : expand_grad
 
 # eye
-- api : eye
-  args : (int64 num_rows, int64 num_colums, DataType dtype)
-  output : Tensor
-  infer_meta :
-    func : EyeInferMeta
-  kernel :
-    func : eye
+# - api : eye
+#   args : (int64 num_rows, int64 num_colums, DataType dtype = DataType::FLOAT32)
+#   output : Tensor
+#   infer_meta :
+#     func : EyeInferMeta
+#   kernel :
+#     func : eye
 
 # # flip
 # - api : flip
@@ -674,16 +682,18 @@
 
 # gaussian_random
 # - api : gaussian_random
-#   args : (ScalayArray shape, float mean, float std, int seed, DataType dtype)
+#   args : (ScalarArray shape, float mean, float std, int seed, DataType dtype=DataType::FLOAT32)
 #   output : Tensor
 #   infer_meta :
-#     func : GaussianRandomInferMeta
+#     func : CreateInferMeta
+#     param : [shape, dtype]
 #   kernel :
 #     func : gaussian_random
+#     data_type : dtype
 
 # # graph_send_recv
 # - api : graph_send_recv
-#   args : (Tensor x, Tensor src_index, Tensor dst_index, string pool_type)
+#   args : (Tensor x, Tensor src_index, Tensor dst_index, str pool_type)
 #   output : Tensor(out), Tensor(dst_count)
 #   infer_meta :
 #     func : GraphSendRecvInferMeta
@@ -693,30 +703,30 @@
 
 # # histogram int64 ???
 # - api : histogram
-#   args : (Tensor x, int64_t bins, int min, int max)
+#   args : (Tensor x, int64 bins, int min, int max)
 #   output : Tensor
 #   infer_meta :
 #     func : HistogramInferMeta
 #   kernel :
 #     func : histogram
 
-# # increment
-# - api : increment
-#   args : (Tensor x, float value)
-#   output : Tensor
-#   infer_meta :
-#     func : IncrementInferMeta
-#   kernel :
-#     func : increment
+# increment
+- api : increment
+  args : (Tensor x, float value)
+  output : Tensor
+  infer_meta :
+    func : IncrementInferMeta
+  kernel :
+    func : increment
 
-# # is_empty
-# - api : is_empty
-#   args : (Tensor x)
-#   output : Tensor
-#   infer_meta :
-#     func : IsEmptyInferMeta
-#   kernel :
-#     func : is_empty
+# is_empty
+- api : is_empty
+  args : (Tensor x)
+  output : Tensor
+  infer_meta :
+    func : IsEmptyInferMeta
+  kernel :
+    func : is_empty
 
 # # isinf selected rows??? involk
 # - api : isinf
@@ -727,7 +737,7 @@
 #   kernel :
 #     func : isinf
 
-# # isnan
+# # isnan selected rows??? involk
 # - api : isnan
 #   args : (Tensor x)
 #   output : Tensor
@@ -736,7 +746,7 @@
 #   kernel :
 #     func : isnan
 
-# # isfinite
+# # isfinite selected rows??? involk
 # - api : isfinite
 #   args : (Tensor x)
 #   output : Tensor
@@ -745,34 +755,37 @@
 #   kernel :
 #     func : isfinite
 
-# # label_smooth  ?? optional
+# label_smooth  ?? optional
 # - api : label_smooth
 #   args : (Tensor label, Tensor prior_dist, float epsilon)
 #   output : Tensor
 #   infer_meta :
-#     func : LabelSmoothInferMeta
+#     func : UnchangedInferMeta
+#     param : [label]
 #   kernel :
 #     func : label_smooth
 #   backward : label_smooth_grad
+#   optional : prior_dist
 
-# # linspace
+# linspace ???? start stop number，应该是sclar？
 # - api : linspace
-#   args : (Tensor start, Tensor stop, Tensor number, DataType dtype)
+#   args : (Tensor start, Tensor stop, Tensor number, DataType dtype=DataType::FLOAT32)
 #   output : Tensor
 #   infer_meta :
 #     func : LinspaceInferMeta
 #   kernel :
 #     func : linspace
 
-# # log_loss
-# - api : log_loss
-#   args : (Tensor input, Tensor label, float epsilon)
-#   output : Tensor
-#   infer_meta :
-#     func : LogLossInferMeta
-#   kernel :
-#     func : log_loss
-#   backward : log_loss
+
+# log_loss
+- api : log_loss
+  args : (Tensor input, Tensor label, float epsilon)
+  output : Tensor
+  infer_meta :
+    func : LogLossInferMeta
+  kernel :
+    func : log_loss
+  backward : log_loss_grad
 
 # # logical_and
 # - api : logical_and
@@ -802,14 +815,14 @@
 #   kernel :
 #     func : logical_xor
 
-# # logical_not
-# - api : logical_not
-#   args : (Tensor x)
-#   output : Tensor
-#   infer_meta :
-#     func : UnchangedInferMeta
-#   kernel :
-#     func : logical_not
+# logical_not
+- api : logical_not
+  args : (Tensor x)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+  kernel :
+    func : logical_not
 
 # # masked_select
 # - api : masked_select
@@ -831,24 +844,25 @@
 #     func : multi_dot
 #   backward : multi_dot_grad
 
-# # multinomial
-# - api : multinomial
-#   args : (Tensor x, int num_samples, bool replacement)
-#   output : Tensor
-#   infer_meta :
-#     func : MultinomialInferMeta
-#   kernel :
-#     func : multinomial
+# multinomial
+- api : multinomial
+  args : (Tensor x, int num_samples, bool replacement)
+  output : Tensor
+  infer_meta :
+    func : MultinomialInferMeta
+  kernel :
+    func : multinomial
 
-# # nll_loss  ?? optional
+# nll_loss  ?? optional
 # - api : nll_loss
-#   args : (Tensor input, Tensor label, Tensor weight, int64_t ignore_index, string reduction)
+#   args : (Tensor x, Tensor label, Tensor weight, int64 ignore_index, str reduction)
 #   output : Tensor(out), Tensor(total_weight)
 #   infer_meta :
-#     func : NllLossInferMeta
+#     func : NllLossRawInferMeta
 #   kernel :
 #     func : nll_loss
 #   backward : nll_loss_grad
+#   optional : weight
 
 # # pad
 # - api : pad
@@ -862,7 +876,7 @@
 
 # # pixel_shuffle
 # - api : pixel_shuffle
-#   args : (Tensor x, int upscale_factor, string data_format)
+#   args : (Tensor x, int upscale_factor, str data_format)
 #   output : Tensor
 #   infer_meta :
 #     func : PixelShuffleInferMeta
@@ -880,7 +894,7 @@
 #     func : poisson
 #   backward : poisson_grad
 
-# # psroi_pool  ?? optional
+# psroi_pool  ?? optional
 # - api : psroi_pool
 #   args : (Tensor x, Tensor rois, Tensor rois_num, int pooled_weight, int pooled_width, int output_channels, float spatial_scale )
 #   output : Tensor
@@ -889,6 +903,7 @@
 #   kernel :
 #     func : psroi_pool
 #   backward : psroi_pool_grad
+#   optional : rois_num
 
 # # randint_raw
 # - api : randint
@@ -926,57 +941,59 @@
 #   kernel :
 #     func : reduce_prod
 
-# # selu
-# - api : selu
-#   args : (Tensor x, float scale, float alpha)
-#   output : Tensor
-#   infer_meta :
-#     func : SeluInferMeta
-#   kernel :
-#     func : selu
-#   backward : selu_grad
+# selu
+- api : selu
+  args : (Tensor x, float scale, float alpha)
+  output : Tensor
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : selu
+  backward : selu_grad
 
 # # set_value None api
 
 # # sgd # need invoke
 # # shape ??? selcted rows
 
-# # shard_index
-# - api : shard_index
-#   args : (Tensor in, int index_num, int nshards, int shard_id, int ignore_value)
-#   output : Tensor
-#   infer_meta :
-#     func : ShardIndexInferMeta
-#   kernel :
-#     func : shard_index
+# shard_index
+- api : shard_index
+  args : (Tensor in, int index_num, int nshards, int shard_id, int ignore_value)
+  output : Tensor
+  infer_meta :
+    func : ShardIndexInferMeta
+  kernel :
+    func : shard_index
 
-# # sigmoid_cross_entropy_with_logits
-# - api : sigmoid_cross_entropy_with_logits
-#   args : (Tensor x, Tensor label, bool normalize, int ignore_index)
-#   output : Tensor
-#   infer_meta :
-#     func : SigmoidCrossEntropyWithoLgitsInferMeta
-#   kernel :
-#     func : sigmoid_cross_entropy_with_logits
+# sigmoid_cross_entropy_with_logits
+- api : sigmoid_cross_entropy_with_logits
+  args : (Tensor x, Tensor label, bool normalize, int ignore_index)
+  output : Tensor
+  infer_meta :
+    func : SigmoidCrossEntropyWithLogitsInferMeta
+  kernel :
+    func : sigmoid_cross_entropy_with_logits
+  backward : sigmoid_cross_entropy_with_logits_grad
 
-# # size
-# - api : size
-#   args : (Tensor x)
-#   output : Tensor
-#   infer_meta :
-#     func : SizeInferMeta
-#   kernel :
-#     func : size
+# size
+- api : size
+  args : (Tensor x)
+  output : Tensor
+  infer_meta :
+    func : SizeInferMeta
+  kernel :
+    func : size
 
-# # tile
-# - api : tile
-#   args : (Tensor x, ScalarArray repeat_times)
-#   output : Tensor
-#   infer_meta :
-#     func : TileInferMeta
-#   kernel :
-#     func : tile
-#   backward : tile_grad
+# tile
+- api : tile
+  args : (Tensor x, ScalarArray repeat_times)
+  output : Tensor
+  infer_meta :
+    func : TileInferMeta
+  kernel :
+    func : tile
+  backward : tile_grad
 
 # # top_k
 # - api : top_k
@@ -988,15 +1005,15 @@
 #     func : top_k
 #   backward : top_k_grad
 
-# # trace
-# - api : trace
-#   args : (Tensor x, int offset, int axis1, int axis2)
-#   output : Tensor
-#   infer_meta :
-#     func : TraceInferMeta
-#   kernel :
-#     func : trace
-#   backward : trace_grad
+# trace
+- api : trace
+  args : (Tensor x, int offset, int axis1, int axis2)
+  output : Tensor
+  infer_meta :
+    func : TraceInferMeta
+  kernel :
+    func : trace
+  backward : trace_grad
 
 # # phi_transfer_layout | not have python api
 
@@ -1018,18 +1035,19 @@
 #   kernel :
 #     func : unbind
 
-# # unfold
-# - api : unfold
-#   args : (Tensor x, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
-#   output : Tensor
-#   infer_meta :
-#     func : UnfoldInferMeta
-#   kernel :
-#     func : unfold
-#   backward : unfold_grad
+# unfold
+- api : unfold
+  args : (Tensor x, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
+  output : Tensor
+  infer_meta :
+    func : UnfoldInferMeta
+  kernel :
+    func : unfold
+  backward : unfold_grad
 
 # # uniform_random_raw selected rows ??
-# # viterbi_decode
+
+# viterbi_decode
 # - api : viterbi_decode
 #   args : (Tensor input, Tensor transition, Tensor length, bool include_bos_eos_tag)
 #   output : Tensor(scores), Tensor(path)
@@ -1038,14 +1056,14 @@
 #   kernel :
 #     func : viterbi_decode
 
-# # where_index
-# - api : where_index
-#   args : (Tensor condition)
-#   output : Tensor
-#   infer_meta :
-#     func : WhereIndexInferMeta
-#   kernel :
-#     func : where_index
+# where_index
+- api : where_index
+  args : (Tensor condition)
+  output : Tensor
+  infer_meta :
+    func : WhereIndexInferMeta
+  kernel :
+    func : where_index
 
 
 # # yolo_box

python/paddle/utils/code_gen/backward.yaml

@@ -112,55 +112,56 @@
 #     func : maxout_grad
 
 
-# - backward_api : put_along_axis_grad
-#   forward : put_along_axis (Tensor x, Tensor index, Tensor value, int axis, string reduce) -> Tensor(out)
-#   args : (Tensor x, Tensor index, Tensor out_grad, int axis, string reduce)
-#   output : Tensor(x_grad), Tensor(value_grad)
-#   infer_meta :
-#     func : GeneralBinaryGradInferMeta
-#     param : [x, valule]
-#   kernel :
-#     func : put_along_axis_grad
+# output is optional 如何处理
+- backward_api : put_along_axis_grad
+  forward : put_along_axis (Tensor x, Tensor index, Tensor value, int axis, str reduce) -> Tensor(out)
+  args : (Tensor x, Tensor index, Tensor out_grad, int axis, str reduce)
+  output : Tensor(x_grad), Tensor(value_grad)
+  infer_meta :
+    func : GeneralBinaryGradInferMeta
+    param : [x, index]
+  kernel :
+    func : put_along_axis_grad
 
-# - backward_api : take_along_axis_grad
-#   forward : take_along_axis (Tensor x, Tensor index, int axis) -> Tensor(out)
-#   args : (Tensor x, Tensor index, Tensor out_grad, int axis)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : take_along_axis_grad
+- backward_api : take_along_axis_grad
+  forward : take_along_axis (Tensor x, Tensor index, int axis) -> Tensor(out)
+  args : (Tensor x, Tensor index, Tensor out_grad, int axis)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : take_along_axis_grad
 
-# - backward_api : maxtrix_power_grad
-#   forward : maxtrix_power (Tensor x, int n) -> Tensor(out)
-#   args : (Tensor x, Tensor out, Tensor out_grad, int n)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : maxtrix_power_grad
+- backward_api : matrix_power_grad
+  forward : matrix_power (Tensor x, int n) -> Tensor(out)
+  args : (Tensor x, Tensor out, Tensor out_grad, int n)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : matrix_power_grad
   
-# - backward_api : eigh_grad
-#   forward : eigh (Tensor x, string uplo) -> Tensor(out_w), Tensor(out_v)
-#   args : (Tensor out_w, Tensor out_v, Tensor out_w_grad, Tensor out_v_grad)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [out_v]
-#   kernel :
-#     func : eigh_grad
+- backward_api : eigh_grad
+  forward : eigh (Tensor x, str uplo) -> Tensor(out_w), Tensor(out_v)
+  args : (Tensor out_w, Tensor out_v, Tensor out_w_grad, Tensor out_v_grad)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [out_v]
+  kernel :
+    func : eigh_grad
 
-# - backward_api : segment_pool_grad
-#   forward : segment_pool (Tensor x, Tensor segment_ids, string pooltype) -> Tensor(out), Tensor(summed_ids)
-#   args : (Tensor x, Tensor segment_ids, Tensor out, Tensor summed_ids, Tenosr out_grad, string pooltype)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : segment_pool_grad
+- backward_api : segment_pool_grad
+  forward : segment_pool (Tensor x, Tensor segment_ids, str pooltype) -> Tensor(out), Tensor(summed_ids)
+  args : (Tensor x, Tensor segment_ids, Tensor out, Tensor summed_ids, Tensor out_grad, str pooltype)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : segment_pool_grad
 
 - backward_api : cos_grad
   forward : cos (Tensor x) -> Tensor(out)
@@ -282,19 +283,19 @@
   kernel :
     func : relu_grad
 
-# - backward_api : argsort_grad
-#   forward : argsort (Tensor x, int axis, bool descending) -> Tensor(out), tensor(indices)
-#   args : (Tensor indices, Tensor x, Tensor out_grad, int axis, bool descending)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : argsort_grad
+- backward_api : argsort_grad
+  forward : argsort (Tensor x, int axis, bool descending) -> Tensor(out), Tensor(indices)
+  args : (Tensor indices, Tensor x, Tensor out_grad, int axis, bool descending)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : argsort_grad
 
 
 # - backward_api : batch_norm_grad
-#   forward : batch_norm (Tensor x, Tensor scale, Tensor bias, Tensor mean, Tensor variance, float momentum, float epsilon, string data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
+#   forward : batch_norm (Tensor x, Tensor scale, Tensor bias, Tensor mean, Tensor variance, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
 #   args : (Tensor indices, Tensor x, Tensor out_grad, int axis, bool descending)
 #   output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
 #   infer_meta :
@@ -312,6 +313,7 @@
 #     param : [x, y, weight, bias]
 #   kernel :
 #     func : bilinear_tensor_product_grad
+#   optional : bias
 
 # - backward_api : broadcast_tensor_grad
 #   forward : broadcast_tensors (Tensor[] x) -> Tensor [] (out)
@@ -344,8 +346,8 @@
     func : cholesky_solve_grad
 
 # - backward_api : dropout_grad
-#   forward : dropout (Tensor x, Tensor seed_tensor, float p, bool is_test, string mode, int seed, bool fix_seed) -> Tensor(out), Tensor(mask)
-#   args : (Tensor mask, Tensor out_grad, float p, bool is_test, string mode)
+#   forward : dropout (Tensor x, Tensor seed_tensor, float p, bool is_test, str mode, int seed, bool fix_seed) -> Tensor(out), Tensor(mask)
+#   args : (Tensor mask, Tensor out_grad, float p, bool is_test, str mode)
 #   output : Tensor(x_grad)
 #   infer_meta :
 #     func : UnchangedInferMeta
@@ -354,25 +356,25 @@
 #     func : dropout_grad
 
 
-# - backward_api : erf_grad
-#   forward : erf (Tensor x) -> Tensor(out)
-#   args : (Tensor x, Tensor out, Tensor out_grad)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : erf_grad
+- backward_api : erf_grad
+  forward : erf (Tensor x) -> Tensor(out)
+  args : (Tensor x, Tensor out, Tensor out_grad)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : erf_grad
   
-# - backward_api : erfinv_grad
-#   forward : erf (Tensor x) -> Tensor(out)
-#   args : (Tensor out, Tensor out_grad)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : ErfinvGradInferMeta
-#     param : [out]
-#   kernel :
-#     func : erfinv_grad
+- backward_api : erfinv_grad
+  forward : erf (Tensor x) -> Tensor(out)
+  args : (Tensor out, Tensor out_grad)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [out]
+  kernel :
+    func : erfinv_grad
 
 # - backward_api : expand_as_grad
 #   forward : expand_as (Tensor x, Tensor y, int[] target_shape) -> Tensor(out)
@@ -395,11 +397,11 @@
 #     func : expand_grad
 
 # - backward_api : graph_send_recv_grad
-#   forward : graph_send_recv (Tensor x, Tensor src_index, Tensor dst_index, string pool_type) -> Tensor(out), Tensor(dst_count)
-#   args : (Tensor out_grad, Tensor x, Tensor out, Tensor src_index, Tensor dst_index, Tensor dst_count, string pool_type)
+#   forward : graph_send_recv (Tensor x, Tensor src_index, Tensor dst_index, str pool_type) -> Tensor(out), Tensor(dst_count)
+#   args : (Tensor out_grad, Tensor x, Tensor out, Tensor src_index, Tensor dst_index, Tensor dst_count, str pool_type)
 #   output : Tensor(x_grad)
 #   infer_meta :
-#     func : UnchangedGradInferMeta
+#     func : UnchangedInferMeta
 #     param : [x]
 #   kernel :
 #     func : graph_send_recv_grad
@@ -407,22 +409,22 @@
 # - backward_api : label_smooth_grad
 #   forward : label_smooth (Tensor label, Tensor prior_dist, float epsilon) -> Tensor(out)
 #   args : (Tensor out_grad, float epsilon)
-#   output : Tensor(x_grad)
+#   output : Tensor(label_grad)
 #   infer_meta :
-#     func : XXXXInferMeta
-#     param : [x]
+#     func : UnchangedInferMeta
+#     param : [out_grad]
 #   kernel :
 #     func : label_smooth_grad
   
-# - backward_api : log_loss_grad
-#   forward : log_loss (Tensor input, Tensor label, float epsilon) -> Tensor(out)
-#   args : (Tensor input, Tensor label, Tensor out_grad, float epsilon)
-#   output : Tensor(input_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : log_loss_grad
+- backward_api : log_loss_grad
+  forward : log_loss (Tensor input, Tensor label, float epsilon) -> Tensor(out)
+  args : (Tensor input, Tensor label, Tensor out_grad, float epsilon)
+  output : Tensor(input_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [input]
+  kernel :
+    func : log_loss_grad
 
 # - backward_api : masked_selecte_grad
 #   forward : masked_select (Tensor x, Tensor mask) -> Tensor(out)
@@ -445,8 +447,8 @@
 #     func : multi_dot_grad
 
 # - backward_api : nll_loss_grad
-#   forward : nll_loss (Tensor input, Tensor label, Tensor weight, int64_t ignore_index, string reduction) -> Tensor(out), Tensor(total_weight)
-#   args : (Tensor x, Tensor label, Tensor total_weight, Tensor weight, Tensor out_grad, int64_t ignore_index, string reduction)
+#   forward : nll_loss (Tensor x, Tensor label, Tensor weight, int64 ignore_index, str reduction) -> Tensor(out), Tensor(total_weight)
+#   args : (Tensor x, Tensor label, Tensor total_weight, Tensor weight, Tensor out_grad, int64 ignore_index, str reduction)
 #   output : Tensor[] (x_grad)
 #   infer_meta :
 #     func : UnchangedInferMeta
@@ -465,8 +467,8 @@
 #     func : pad_grad
 
 # - backward_api : pixel_shuffle_grad
-#   forward : pixel_shuffle (Tensor x, int upscale_factor, string data_format) -> Tensor(out)
-#   args : (Tensor out_grad, int upscale_factor, string data_format)
+#   forward : pixel_shuffle (Tensor x, int upscale_factor, str data_format) -> Tensor(out)
+#   args : (Tensor out_grad, int upscale_factor, str data_format)
 #   output : Tensor(x_grad)
 #   infer_meta :
 #     func : XXXXXInferMeta
@@ -493,36 +495,37 @@
 #     param : [x]
 #   kernel :
 #     func : psroi_pool_grad
+#   optional : rois_num
 
-# - backward_api : selu_grad
-#   forward : selu (Tensor x, float scale, float alpha) -> Tensor(out)
-#   args : (Tensor out, Tensor out_grad, float scale, float alpha)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : XXXXXInferMeta
-#     param : [x]
-#   kernel :
-#     func : selu_grad
+- backward_api : selu_grad
+  forward : selu (Tensor x, float scale, float alpha) -> Tensor(out)
+  args : (Tensor out, Tensor out_grad, float scale, float alpha)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [out]
+  kernel :
+    func : selu_grad
 
-# - backward_api : sigmoid_cross_entropy_with_logits_grad
-#   forward : sigmoid_cross_entropy_with_logits (Tensor x, Tensor label, bool normalize, int ignore_index) -> Tensor(out)
-#   args : (Tensor x, Tensor label, Tensor out_grad, bool normalize, int ingore_index)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : sigmoid_cross_entropy_with_logits_grad
+- backward_api : sigmoid_cross_entropy_with_logits_grad
+  forward : sigmoid_cross_entropy_with_logits (Tensor x, Tensor label, bool normalize, int ignore_index) -> Tensor(out)
+  args : (Tensor x, Tensor label, Tensor out_grad, bool normalize, int ignore_index)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : sigmoid_cross_entropy_with_logits_grad
 
-# - backward_api : tile_grad
-#   forward : tile (Tensor x, ScalarArray repeat_times) -> Tensor(out)
-#   args : (Tensor x, Tensor out_grad, ScalarArray repeat_times)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : tile_grad
+- backward_api : tile_grad
+  forward : tile (Tensor x, ScalarArray repeat_times) -> Tensor(out)
+  args : (Tensor x, Tensor out_grad, ScalarArray repeat_times)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : tile_grad
 
 # # forward backward type not match
 # - backward_api : top_k_grad 
@@ -536,25 +539,25 @@
 #     func : top_k_grad
 
 
-# - backward_api : trace_grad 
-#   forward : trace (Tensor x, int offset, int axis1, int axis2) -> Tensor(out)
-#   args : (Tensor out_grad, Tensor x, int offset, int axis1, int axis2)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : trace_grad
+- backward_api : trace_grad 
+  forward : trace (Tensor x, int offset, int axis1, int axis2) -> Tensor(out)
+  args : (Tensor x, Tensor out_grad, int offset, int axis1, int axis2)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : trace_grad
 
-# - backward_api : unfold_grad 
-#   forward : unfold (Tensor x, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations) -> Tensor(out)
-#   args : (Tensor x, Tensor out_grad, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
-#   output : Tensor(x_grad)
-#   infer_meta :
-#     func : UnchangedInferMeta
-#     param : [x]
-#   kernel :
-#     func : unfold_grad
+- backward_api : unfold_grad 
+  forward : unfold (Tensor x, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations) -> Tensor(out)
+  args : (Tensor x, Tensor out_grad, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : unfold_grad
 
 # - backward_api : where_index_grad 
 #   forward : where_index (Tensor condition) -> Tensor(out)
@@ -564,4 +567,4 @@
 #     func : UnchangedInferMeta
 #     param : [x]
 #   kernel :
-#     func : trace_grad
+#     func : where_index_grad

python/paddle/utils/code_gen/wrapped_infermeta_gen.py

@@ -43,9 +43,12 @@ PD_REGISTER_INFER_META_FN({api.kernel['func'][0]}, phi::{api.infer_meta['func']}
                 'const std::vector<Tensor>&': 'const std::vector<MetaTensor>&',
                 'Tensor': 'MetaTensor*',
                 'std::vector<Tensor>': 'std::vector<MetaTensor>*',
+                'const paddle::optional<Tensor&>':
+                'const paddle::optional<MetaTensor&>'
             }
 
             wrapped_infermeta_name = get_wrapped_infermeta_name(api.api)
+            print(wrapped_infermeta_name)
             args = []
             for input_name in api.inputs['names']:
                 if input_name in kernel_params:
@@ -132,6 +135,7 @@ def generate_wrapped_infermeta_and_register(api_yaml_path, header_file_path,
 
     for api in apis:
         api_item = ForwardAPI(api)
+        #print( str(api_item) )
         declare_code, defind_code, register_code = gene_wrapped_infermeta_and_register(
             api_item)
         header_file.write(declare_code)