add fill_constant input(shape) dtype check. test=develop (#20620)

python/paddle/fluid/layers/tensor.py

@@ -524,7 +524,10 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None):
     The attribute `stop_gradient` of the created Tensor is setted to True.
 
     Args:
-        shape(tuple|list): Shape of the Tensor to be created.
+        shape(list|tuple|Variable): Shape of the Tensor to be created.
+                The data type is ``int32`` or ``int64`` . If ``shape`` is a list or tuple,
+                the elements of it should be integers or Tensors with shape [1].
+                If ``shape`` is an Variable, it should be an 1-D Tensor .
         dtype(np.dtype|core.VarDesc.VarType|str): Data type of the output tensor which can
             be float16, float32, float64, int32, int64.
         value(float): The constant value used to initialize the Tensor to be created.
@@ -544,9 +547,18 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None):
         .. code-block:: python
 
           import paddle.fluid as fluid
-          data1 = fluid.layers.fill_constant(shape=[2,1], value=0, dtype='int64') #data1=[[0],[0]]
+          # attr shape is a list which doesn't contain Variable Tensor.
+          data1 = fluid.layers.fill_constant(shape=[2,1], value=0, dtype='int64') # data1=[[0],[0]]
           data2 = fluid.layers.fill_constant(shape=[2,1], value=5, dtype='int64', out=data1)
-          #data1=[[5], [5]] data2=[[5], [5]]
+          # data1=[[0], [0]] data2=[[5], [5]]
+
+          # attr shape is a list which contains Variable Tensor.
+          positive_2 = fluid.layers.fill_constant([1], "int32", 2)
+          data3 = fluid.layers.fill_constant(shape=[1, positive_2], dtype='float32', value=1.5) # data3=[1.5, 1.5]
+
+          # attr shape is an Variable Tensor.
+          shape = fluid.layers.fill_constant([1,2], "int32", 2) # shape=[2,2]
+          data4 = fluid.layers.fill_constant(shape=shape, dtype='bool', value=True) # data4=[[True,True],[True,True]]
     """
     helper = LayerHelper("fill_constant", **locals())
     if convert_dtype(dtype) not in [
@@ -585,9 +597,17 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None):
 
     def _get_shape_tensor(list_shape):
         new_shape_tensor = []
-        for dim in list_shape:
+        for idx, dim in enumerate(list_shape):
             if isinstance(dim, Variable):
                 dim.stop_gradient = True
+                if convert_dtype(dim.dtype) not in ['int32', 'int64']:
+                    raise TypeError(
+                        "When type of 'shape' in fill_constant is list or tuple, "
+                        "the data type of the element with type Variable must be int32 or int64, "
+                        "but received the data type of shape[%d] is %s." %
+                        (idx, convert_dtype(dim.dtype)))
+                if convert_dtype(dim.dtype) == 'int64':
+                    dim = cast(x=dim, dtype='int32')
                 new_shape_tensor.append(dim)
             else:
                 temp_out = helper.create_variable_for_type_inference('int32')
@@ -597,6 +617,12 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None):
 
     if isinstance(shape, Variable):
         shape.stop_gradient = True
+        if convert_dtype(shape.dtype) not in ['int32', 'int64']:
+            raise TypeError(
+                "When type of 'shape' in fill_constant is Variable, the data type of 'shape' must be int32 or int64, "
+                "but received %s." % (convert_dtype(shape.dtype)))
+        if (convert_dtype(shape.dtype) == 'int64'):
+            shape = cast(shape, 'int32')
         inputs["ShapeTensor"] = shape
     elif isinstance(shape, (list, tuple)):
         assert len(shape) > 0, (

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

@@ -189,33 +189,47 @@ class TestFillConstantOp1_ShapeTensor(OpTest):
         self.check_output()
 
 
-# # Test python API
+# Test python API
 class TestFillConstantAPI(OpTest):
     def test_api(self):
-        positive_2 = fluid.layers.fill_constant([1], "int32", 2)
-        shape_tensor = fluid.layers.data(
-            name="shape_tensor",
-            shape=[2],
-            append_batch_size=False,
-            dtype="int32")
+        positive_2_int32 = fluid.layers.fill_constant([1], "int32", 2)
+
+        positive_2_int64 = fluid.layers.fill_constant([1], "int64", 2)
+        shape_tensor_int32 = fluid.data(
+            name="shape_tensor_int32", shape=[2], dtype="int32")
+
+        shape_tensor_int64 = fluid.data(
+            name="shape_tensor_int64", shape=[2], dtype="int64")
 
         out_1 = fluid.layers.fill_constant(
             shape=[1, 2], dtype="float32", value=1.1)
+
         out_2 = fluid.layers.fill_constant(
-            shape=[1, positive_2], dtype="float32", value=1.1)
+            shape=[1, positive_2_int32], dtype="float32", value=1.1)
 
         out_3 = fluid.layers.fill_constant(
-            shape=shape_tensor, dtype="float32", value=1.1)
+            shape=[1, positive_2_int64], dtype="float32", value=1.1)
+
+        out_4 = fluid.layers.fill_constant(
+            shape=shape_tensor_int32, dtype="float32", value=1.1)
+
+        out_5 = fluid.layers.fill_constant(
+            shape=shape_tensor_int64, dtype="float32", value=1.1)
 
         exe = fluid.Executor(place=fluid.CPUPlace())
-        res_1, res_2, res_3 = exe.run(
+        res_1, res_2, res_3, res_4, res_5 = exe.run(
             fluid.default_main_program(),
-            feed={"shape_tensor": np.array([1, 2]).astype("int32")},
-            fetch_list=[out_1, out_2, out_3])
+            feed={
+                "shape_tensor_int32": np.array([1, 2]).astype("int32"),
+                "shape_tensor_int64": np.array([1, 2]).astype("int64"),
+            },
+            fetch_list=[out_1, out_2, out_3, out_4, out_5])
 
         assert np.array_equal(res_1, np.full([1, 2], 1.1, dtype="float32"))
         assert np.array_equal(res_2, np.full([1, 2], 1.1, dtype="float32"))
         assert np.array_equal(res_3, np.full([1, 2], 1.1, dtype="float32"))
+        assert np.array_equal(res_4, np.full([1, 2], 1.1, dtype="float32"))
+        assert np.array_equal(res_5, np.full([1, 2], 1.1, dtype="float32"))
 
 
 class TestFillConstantOpError(OpTest):
@@ -236,7 +250,8 @@ class TestFillConstantOpError(OpTest):
                 value=5,
                 dtype='int16',
                 out=x1)
-            # The input dtype of fill_constant must be one of bool, float16,
+
+            # The argument dtype of fill_constant_op must be one of bool, float16,
             #float32, float64, int32 or int64
             x2 = fluid.layers.data(name='x2', shape=[1], dtype="int32")
 
@@ -254,17 +269,35 @@ class TestFillConstantOpError(OpTest):
                 dtype='float64',
                 out=x2)
 
-            # test Error of Shape
+            # The argument shape's type of fill_constant_op must be list, tuple or Variable.
             def test_shape_type():
                 fluid.layers.fill_constant(shape=1, dtype="float32", value=1)
 
             self.assertRaises(TypeError, test_shape_type)
 
+            # The argument shape's size of fill_constant_op must not be 0.
             def test_shape_size():
                 fluid.layers.fill_constant(shape=[], dtype="float32", value=1)
 
             self.assertRaises(AssertionError, test_shape_size)
 
+            # The shape dtype of fill_constant_op must be int32 or int64.
+            def test_shape_tensor_dtype():
+                shape = fluid.data(
+                    name="shape_tensor", shape=[2], dtype="float32")
+                fluid.layers.fill_constant(
+                    shape=shape, dtype="float32", value=1)
+
+            self.assertRaises(TypeError, test_shape_tensor_dtype)
+
+            def test_shape_tensor_list_dtype():
+                shape = fluid.data(
+                    name="shape_tensor_list", shape=[1], dtype="bool")
+                fluid.layers.fill_constant(
+                    shape=[shape, 2], dtype="float32", value=1)
+
+            self.assertRaises(TypeError, test_shape_tensor_list_dtype)
+
 
 if __name__ == "__main__":
     unittest.main()