varbase_patch_methods.py 10.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
# Copyright (c) 2019 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 .. import framework
from .. import core
from . import BackwardStrategy
from ..framework import Variable, _getitem_impl_
from .. import unique_name
import numpy as np
21
from .math_op_patch import monkey_patch_math_varbase
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210


def monkey_patch_varbase():
    # TODO(jiabin): move this to cplusplus end if we find some performance issue on it
    @framework.dygraph_only
    def set_value(self, value):
        """
        **Notes**:
            **This API is ONLY avaliable in Dygraph mode**

        Set a new value for this Variable.

        Args:
            value (Variable|np.ndarray): the new value.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
                from paddle.fluid.dygraph import FC
                import numpy as np

                data = np.ones([3, 32, 32], dtype='float32')
                with fluid.dygraph.guard():
                    fc = fluid.dygraph.FC("fc", 4)
                    t = to_variable(data)
                    fc(t)  # call with default weight
                    custom_weight = np.random.randn(1024, 4).astype("float32")
                    fc.weight.set_value(custom_weight)  # change existing weight
                    out = fc(t)  # call with different weight

        """
        assert isinstance(value, (np.ndarray, core.VarBase)), \
            "Variable set_value function, arguments type only support Variable, numpy, VarBase"

        value_np = value
        if isinstance(value, core.VarBase):
            value_np = value.numpy()

        self_tensor_np = self.numpy()

        assert self_tensor_np.shape == value_np.shape, \
            "Variable Shape not match, Variable [ {} ] need tensor with shape {} but load set tensor with shape {}".format(
                self.name, self_tensor_np.shape, value_np.shape)

        assert self_tensor_np.dtype == value_np.dtype, \
            "Variable dtype not match, Variable [ {} ] need tensor with dtype {}  but load tensor with dtype {}".format(
                self.name, self_tensor_np.dtype, value_np.dtype)

        self.value().get_tensor().set(value_np,
                                      framework._current_expected_place())

    @framework.dygraph_only
    def backward(self, backward_strategy=None):
        """
        **Notes**:
            **This API is ONLY avaliable in Dygraph mode**

        Run backward of current Graph which starts from current Variable

        Args:
            backward_strategy( :ref:`api_fluid_dygraph_BackwardStrategy` ): The Backward Strategy to run backward

        Returns:
            NoneType: None

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

                x = np.ones([2, 2], np.float32)
                with fluid.dygraph.guard():
                    inputs2 = []
                    for _ in range(10):
                        tmp = fluid.dygraph.base.to_variable(x)
                        # if we don't set tmp's stop_gradient as False then, all path to loss will has no gradient since
                        # there is no one need gradient on it.
                        tmp.stop_gradient=False
                        inputs2.append(tmp)
                    ret2 = fluid.layers.sums(inputs2)
                    loss2 = fluid.layers.reduce_sum(ret2)
                    backward_strategy = fluid.dygraph.BackwardStrategy()
                    backward_strategy.sort_sum_gradient = True
                    loss2.backward(backward_strategy)

        """
        if framework.in_dygraph_mode():
            if backward_strategy is None:
                backward_strategy = BackwardStrategy()
                backward_strategy.sort_sum_gradient = False

            self._run_backward(backward_strategy, framework._dygraph_tracer())
        else:
            raise ValueError(
                "Variable.backward() is only avaliable in DyGraph mode")

    @framework.dygraph_only
    def gradient(self):
        """
        **Notes**:
            **This API is ONLY avaliable in Dygraph mode**

        Get the Gradient of Current Variable

        Returns:
            ndarray: Numpy value of the gradient of current Variable

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

                x = np.ones([2, 2], np.float32)
                with fluid.dygraph.guard():
                    inputs2 = []
                    for _ in range(10):
                        tmp = fluid.dygraph.base.to_variable(x)
                        tmp.stop_gradient=False
                        inputs2.append(tmp)
                    ret2 = fluid.layers.sums(inputs2)
                    loss2 = fluid.layers.reduce_sum(ret2)
                    backward_strategy = fluid.dygraph.BackwardStrategy()
                    backward_strategy.sort_sum_gradient = True
                    loss2.backward(backward_strategy)
                    print(loss2.gradient())

        """
        if self._grad_ivar() is None:
            raise ValueError(
                "%s has no grad, Please set Variable.stop_gradient=False, or "
                "check if this is the first and only variable need grad, if so, please set its pre-Variable's "
                "stop_gradient=False, to make sure it has gradient " %
                self.name)
        new_ivar = self._grad_ivar()._copy_to(core.CPUPlace(), True)
        if self._grad_ivar().type == core.VarDesc.VarType.SELECTED_ROWS:
            return (np.array(new_ivar.value().get_selected_rows().get_tensor()),
                    np.array(new_ivar.value().get_selected_rows().rows()))
        else:
            return np.array(new_ivar.value().get_tensor())

    def __str__(self):
        return self.to_string(True)

    @property
    def block(self):
        return framework.default_main_program().global_block()

    def to_string(self, throw_on_error, with_details=False):
        """
        Get debug string.

        Args:

            throw_on_error (bool): True if raise an exception when self is not initialized.

            with_details (bool): more details about variables and parameters (e.g. trainable, optimize_attr, ...) will be printed when with_details is True. Default value is False;

        Returns:
            str: The debug string.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                cur_program = fluid.Program()
                cur_block = cur_program.current_block()
                new_variable = cur_block.create_var(name="X",
                                                    shape=[-1, 23, 48],
                                                    dtype='float32')
                print(new_variable.to_string(True))
                print("=============with detail===============")
                print(new_variable.to_string(True, True))
        """
        if framework.in_dygraph_mode():
            # TODO(panyx0718): add more dygraph debug info.
            tensor = self.value().get_tensor()
            if tensor._is_initialized():
                return 'name %s, dtype: %s shape: %s %s' % (
                    self.name, self.dtype, self.shape, str(tensor))
            else:
                return 'name %s, shape: %s, not inited' % (self.name,
                                                           self.shape)

    def __getitem__(self, item):
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271
        if not isinstance(item, tuple):
            item = [item]

        decrease_axis = []
        slice_axis = []
        slice_start = []
        slice_end = []
        reverse_axis = []

        for dim, slice_item in enumerate(item):
            if isinstance(slice_item, slice):
                start = slice_item.start
                end = slice_item.stop
                step = slice_item.step if slice_item.step else 1

                assert (step == 1 or step == -1)

                if step == -1:
                    reverse_axis.append(dim)
                    assert (start is None and end is None)

                if start is None and end is None:
                    continue

                if start is None:
                    start = 0

                if end is None:
                    end = 10000000

                slice_axis.append(dim)
                slice_start.append(start)
                slice_end.append(end)
            else:
                # int
                decrease_axis.append(dim)
                slice_axis.append(dim)
                slice_start.append(slice_item)
                slice_end.append(slice_item + 1
                                 if slice_item != -1 else 10000000)

        out = self
        if len(slice_axis) > 0:
            # append slice_op here
            inputs = {'Input': [out]}
            attrs = {
                'axes': slice_axis,
                'starts': slice_start,
                'ends': slice_end,
                'decrease_axis': decrease_axis
            }
            outs = core.ops.slice(inputs, attrs)
            out = outs['Out'][0]

        if len(reverse_axis) > 0:
            inputs = {'X': [out]}
            attrs = {'axis': reverse_axis}
            outs = core.ops.reverse(inputs, attrs)
            out = outs['Out'][0]

        return out
272 273 274 275 276 277

    for method_name, method in (("set_value", set_value), ("block", block),
                                ("backward", backward), ("gradient", gradient),
                                ("__str__", __str__), ("to_string", to_string),
                                ("__getitem__", __getitem__)):
        setattr(core.VarBase, method_name, method)
278 279 280

    # patch math methods for varbase
    monkey_patch_math_varbase()