[API 2.0] adaptive expand op to use shape instead of expand_times (#26206)

* adaptive expand op to 2.0 (align to torch.expand) , test=develop

paddle/fluid/operators/expand_v2_op.cu

@@ -0,0 +1,32 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#include "paddle/fluid/operators/expand_v2_op.h"
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_CUDA_KERNEL(
+    expand_v2, ops::ExpandV2Kernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ExpandV2Kernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ExpandV2Kernel<paddle::platform::CUDADeviceContext, plat::float16>,
+    ops::ExpandV2Kernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ExpandV2Kernel<paddle::platform::CUDADeviceContext, int64_t>,
+    ops::ExpandV2Kernel<paddle::platform::CUDADeviceContext, bool>);
+REGISTER_OP_CUDA_KERNEL(
+    expand_v2_grad,
+    ops::ExpandV2GradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ExpandV2GradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ExpandV2GradKernel<paddle::platform::CUDADeviceContext, plat::float16>,
+    ops::ExpandV2GradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ExpandV2GradKernel<paddle::platform::CUDADeviceContext, int64_t>);

python/paddle/fluid/tests/unittests/test_expand_v2_op.py

@@ -0,0 +1,234 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+from op_test import OpTest
+import paddle.fluid as fluid
+from paddle.fluid import compiler, Program, program_guard
+import paddle
+
+
+# Situation 1: shape is a list(without tensor)
+class TestExpandV2OpRank1(OpTest):
+    def setUp(self):
+        self.op_type = "expand_v2"
+        self.init_data()
+
+        self.inputs = {'X': np.random.random(self.ori_shape).astype("float64")}
+        self.attrs = {'shape': self.shape}
+        output = np.tile(self.inputs['X'], self.expand_times)
+        self.outputs = {'Out': output}
+
+    def init_data(self):
+        self.ori_shape = [100]
+        self.shape = [100]
+        self.expand_times = [1]
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestExpandV2OpRank2_DimExpanding(TestExpandV2OpRank1):
+    def init_data(self):
+        self.ori_shape = [120]
+        self.shape = [2, 120]
+        self.expand_times = [2, 1]
+
+
+class TestExpandV2OpRank2(TestExpandV2OpRank1):
+    def init_data(self):
+        self.ori_shape = [1, 140]
+        self.shape = [12, 140]
+        self.expand_times = [12, 1]
+
+
+class TestExpandV2OpRank3_Corner(TestExpandV2OpRank1):
+    def init_data(self):
+        self.ori_shape = (2, 10, 5)
+        self.shape = (2, 10, 5)
+        self.expand_times = (1, 1, 1)
+
+
+class TestExpandV2OpRank4(TestExpandV2OpRank1):
+    def init_data(self):
+        self.ori_shape = (2, 4, 5, 7)
+        self.shape = (-1, -1, -1, -1)
+        self.expand_times = (1, 1, 1, 1)
+
+
+# Situation 2: shape is a list(with tensor)
+class TestExpandV2OpRank1_tensor_attr(OpTest):
+    def setUp(self):
+        self.op_type = "expand_v2"
+        self.init_data()
+        expand_shapes_tensor = []
+        for index, ele in enumerate(self.expand_shape):
+            expand_shapes_tensor.append(("x" + str(index), np.ones(
+                (1)).astype('int32') * ele))
+
+        self.inputs = {
+            'X': np.random.random(self.ori_shape).astype("float64"),
+            'expand_shapes_tensor': expand_shapes_tensor,
+        }
+        self.attrs = {"shape": self.infer_expand_shape}
+        output = np.tile(self.inputs['X'], self.expand_times)
+        self.outputs = {'Out': output}
+
+    def init_data(self):
+        self.ori_shape = [100]
+        self.expand_times = [1]
+        self.expand_shape = [100]
+        self.infer_expand_shape = [-1]
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestExpandV2OpRank2_Corner_tensor_attr(TestExpandV2OpRank1_tensor_attr):
+    def init_data(self):
+        self.ori_shape = [12, 14]
+        self.expand_times = [1, 1]
+        self.expand_shape = [12, 14]
+        self.infer_expand_shape = [12, -1]
+
+
+# Situation 3: shape is a tensor
+class TestExpandV2OpRank1_tensor(OpTest):
+    def setUp(self):
+        self.op_type = "expand_v2"
+        self.init_data()
+
+        self.inputs = {
+            'X': np.random.random(self.ori_shape).astype("float64"),
+            'Shape': np.array(self.expand_shape).astype("int32"),
+        }
+        self.attrs = {}
+        output = np.tile(self.inputs['X'], self.expand_times)
+        self.outputs = {'Out': output}
+
+    def init_data(self):
+        self.ori_shape = [100]
+        self.expand_times = [2, 1]
+        self.expand_shape = [2, 100]
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+# Situation 4: input x is Integer
+class TestExpandV2OpInteger(OpTest):
+    def setUp(self):
+        self.op_type = "expand_v2"
+        self.inputs = {
+            'X': np.random.randint(
+                10, size=(2, 4, 5)).astype("int32")
+        }
+        self.attrs = {'shape': [2, 4, 5]}
+        output = np.tile(self.inputs['X'], (1, 1, 1))
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+# Situation 5: input x is Bool
+class TestExpandV2OpBoolean(OpTest):
+    def setUp(self):
+        self.op_type = "expand_v2"
+        self.inputs = {'X': np.random.randint(2, size=(2, 4, 5)).astype("bool")}
+        self.attrs = {'shape': [2, 4, 5]}
+        output = np.tile(self.inputs['X'], (1, 1, 1))
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+# Situation 56: input x is Integer
+class TestExpandV2OpInt64_t(OpTest):
+    def setUp(self):
+        self.op_type = "expand_v2"
+        self.inputs = {
+            'X': np.random.randint(
+                10, size=(2, 4, 5)).astype("int64")
+        }
+        self.attrs = {'shape': [2, 4, 5]}
+        output = np.tile(self.inputs['X'], (1, 1, 1))
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestExpandV2Error(unittest.TestCase):
+    def test_errors(self):
+        with program_guard(Program(), Program()):
+            x1 = fluid.create_lod_tensor(
+                np.array([[-1]]), [[1]], fluid.CPUPlace())
+            shape = [2, 2]
+            self.assertRaises(TypeError, paddle.tensor.expand, x1, shape)
+            x2 = fluid.layers.data(name='x2', shape=[4], dtype="uint8")
+            self.assertRaises(TypeError, paddle.tensor.expand, x2, shape)
+            x3 = fluid.layers.data(name='x3', shape=[4], dtype="bool")
+            x3.stop_gradient = True
+            self.assertRaises(ValueError, paddle.tensor.expand, x3, shape)
+
+
+# Test python API
+class TestExpandV2API(unittest.TestCase):
+    def test_api(self):
+        input = np.random.random([12, 14]).astype("float32")
+        x = fluid.layers.data(
+            name='x', shape=[12, 14], append_batch_size=False, dtype="float32")
+
+        positive_2 = fluid.layers.fill_constant([1], "int32", 12)
+        expand_shape = fluid.layers.data(
+            name="expand_shape",
+            shape=[2],
+            append_batch_size=False,
+            dtype="int32")
+
+        out_1 = paddle.expand(x, shape=[12, 14])
+        out_2 = paddle.expand(x, shape=[positive_2, 14])
+        out_3 = paddle.expand(x, shape=expand_shape)
+
+        g0 = fluid.backward.calc_gradient(out_2, x)
+
+        exe = fluid.Executor(place=fluid.CPUPlace())
+        res_1, res_2, res_3 = exe.run(fluid.default_main_program(),
+                                      feed={
+                                          "x": input,
+                                          "expand_shape":
+                                          np.array([12, 14]).astype("int32")
+                                      },
+                                      fetch_list=[out_1, out_2, out_3])
+        assert np.array_equal(res_1, np.tile(input, (1, 1)))
+        assert np.array_equal(res_2, np.tile(input, (1, 1)))
+        assert np.array_equal(res_3, np.tile(input, (1, 1)))
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -60,8 +60,10 @@ class TestRetainGraph(unittest.TestCase):
                 interpolatesv = fake_data
             elif type == 'mixed':
                 alpha = paddle.rand((real_data.shape[0], 1))
-                alpha = paddle.expand(
-                    alpha, [1, np.prod(real_data.shape) // real_data.shape[0]])
+                alpha = paddle.expand(alpha, [
+                    real_data.shape[0],
+                    np.prod(real_data.shape) // real_data.shape[0]
+                ])
                 alpha = paddle.reshape(alpha, real_data.shape)
                 interpolatesv = alpha * real_data + ((1 - alpha) * fake_data)
             else:

python/paddle/tensor/manipulation.py

@@ -23,7 +23,6 @@ from ..fluid.layers import utils
 import numpy as np
 # TODO: define functions to manipulate a tensor  
 from ..fluid.layers import cast  #DEFINE_ALIAS
-from ..fluid.layers import expand  #DEFINE_ALIAS
 from ..fluid.layers import expand_as  #DEFINE_ALIAS
 from ..fluid.layers import reshape  #DEFINE_ALIAS
 from ..fluid.layers import scatter  #DEFINE_ALIAS
@@ -794,7 +793,6 @@ def tile(x, repeat_times, name=None):
     """
     :alias_main: paddle.tile
 	:alias: paddle.tile,paddle.tensor.tile,paddle.tensor.manipulation.tile
-
     Construct a new tensor by repeating ``x`` the number of times given by the parameter ``repeat_times``.
     The rank of ``x`` should be less than or equal to 6, and the size of the shape of ``repeat_times`` should
     be less than or equal to 6.
@@ -807,52 +805,38 @@ def tile(x, repeat_times, name=None):
     For example, if the tensor ``x`` is with a shape(4, 3, 2, 2) and ``repeat_times`` is a tuple (3, 2),
     ``repeat_times`` is first promoted to a tuple (1, 1, 3, 2).
     The following gives an using case:
-
-
     .. code-block:: text
-
         Input(x) is a 3-D tensor of shape (2, 3, 1):
-
                 [
                    [[1], [2], [3]],
                    [[4], [5], [6]]
                 ]
-
         Attr(repeat_times):  [1, 2, 2]
-
         Output(out) is a 3-D tensor of shape (2, 6, 2):
-
                 [
                     [[1, 1], [2, 2], [3, 3], [1, 1], [2, 2], [3, 3]],
                     [[4, 4], [5, 5], [6, 6], [4, 4], [5, 5], [6, 6]]
                 ]
-
     Args:
         x (Tensor): The input tensor, its data type should be bool, float32, float64, int32. The rank of ``x`` should be in [1, 6].
         repeat_times (Tensor|tuple|list): The number of repeating times for each dimension of the input ``x``. If repeat_times is a list or tuple, the elements of
                 it should be integers or Tensors with shape [1]. If repeat_times is Tensor, it should be an 1-D Tensor. The size of its shape should be in [1, 6].
         name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name` .
-
     Returns:
         N-D Tensor. The data type is the same as ``x``. After tiling, each dimension of the output is equal to the corresponding dimension of ``x`` multiplying the corresponding value given by ``repeat_times`` .
-
     Raises:
         TypeError: The type of ``repeat_times`` must be list, tuple or Tensor.
         ValueError: The elements of ``repeat_times`` cannot be negative.
-
     Examples:
         .. code-block:: python
-
             import paddle
             import numpy as np
             paddle.disable_static()
-
             # example 1:
             np_data_1 = np.array([1, 2, 3]).astype('int32')
             data_1 = paddle..to_variable(np_data_1)
             tiled_1 = paddle.tile(data_1, repeat_times=[2, 1])
             # [[1, 2, 3], [1, 2, 3]]
-
             # example 2:
             np_repeat_times = np.array([2, 1]).astype("int32")
             repeat_times = paddle.to_variable(np_repeat_times)
@@ -907,3 +891,95 @@ def tile(x, repeat_times, name=None):
     helper.append_op(
         type='tile', inputs=inputs, outputs={'Out': out}, attrs=attrs)
     return out
+
+
+def expand(x, shape, name=None):
+    """
+    :alias_main: paddle.expand
+    :alias: paddle.expand,paddle.tensor.expand,paddle.tensor.manipulation.expand
+
+    Expand the input tensor to a given shape.
+
+    The rank of ``x`` should be less than or equal to 6, and the number of elements in ``shape`` should also be less than or equal to 6. The size of the dimension to expand must be 1. 
+
+
+    Args:
+        x (Tensor): The input Tensor with rank in [1, 6]. The data type is bool, float32, float64 or int32.
+        shape (list|tuple|Tensor): The result shape after expanding. The data type is int32. If shape is a list or tuple, all elements of
+                it should be integers or Tensors with shape (1,). If shape is a Tensor, it should be an 1-D Tensor. 
+                The value -1 in shape, means keeping the corresponding dimension unchanged.
+        name (str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name` .
+
+    Returns:
+        Tensor: A Tensor with the given shape. The data type is the same as ``x``.
+
+    Raises:
+        TypeError: The type of ``shape`` must be list, tuple or Variable.
+        ValueError: The elements of ``shape`` must be positive or -1.
+
+    Examples:
+        .. code-block:: python
+
+            import numpy as np
+            import paddle
+            paddle.disable_static()
+
+            # example 1:
+            np_data_1 = np.array([1, 2, 3]).astype=('int32)
+            data_1 = paddle.to_variable(np_data_1)
+            expanded_1 = paddle.expand(data_1, shape=[2, 3])
+            # [[1, 2, 3], [1, 2, 3]]
+
+            # example 2:
+            np_shape = np.array([2, 3]).astype=('int32)
+            shape = paddle.to_variable(np_shape)
+            expanded_2 = paddle.expand(data_1, shape=shape)
+            # [[1, 2, 3], [1, 2, 3]]
+    """
+    if in_dygraph_mode():
+        if isinstance(shape, (list, tuple)):
+            expand_shape = [
+                item.numpy()[0] if isinstance(item, Variable) else item
+                for item in shape
+            ]
+
+            return core.ops.expand_v2(x, 'shape', expand_shape)
+
+    inputs = {"X": [x]}
+    attrs = {}
+    check_variable_and_dtype(
+        x, 'x', ['bool', 'float32', 'float64', 'int32', 'int64'], 'expand')
+    check_type(shape, 'shape', (list, tuple, Variable), 'expand')
+    if convert_dtype(x.dtype) == 'bool' and x.stop_gradient == True:
+        raise ValueError("When the data type of input 'x' for expand is bool, "
+                         "you must set its stop_gradient to be False by "
+                         " some_var.stop_gradient = False, supporting "
+                         "some_var as the input.")
+
+    helper = LayerHelper('expand', input=x, **locals())
+
+    def get_attr_expand_shape(list_expand_shape):
+        attrs_expand_shape = []
+        for idx, shape in enumerate(list_expand_shape):
+            if isinstance(shape, Variable):
+                attrs_expand_shape.append(-1)
+            else:
+                attrs_expand_shape.append(shape)
+                assert shape > 0 or shape == -1, (
+                    "Every element in shape must be positive or -1.")
+        return attrs_expand_shape
+
+    if isinstance(shape, Variable):
+        shape.stop_gradient = True
+        inputs['Shape'] = shape
+    elif isinstance(shape, (list, tuple)):
+        attrs['shape'] = get_attr_expand_shape(shape)
+        if utils._contain_var(shape):
+            inputs['expand_shapes_tensor'] = utils._convert_to_tensor_list(
+                shape)
+
+    dtype = helper.input_dtype(input_param_name='x')
+    out = helper.create_variable_for_type_inference(dtype)
+    helper.append_op(
+        type='expand_v2', inputs=inputs, outputs={'Out': out}, attrs=attrs)
+    return out