primapi.py 6.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
# Copyright (c) 2022 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.

import typing

17 18
from paddle.fluid import backward, framework
from paddle.incubate.autograd import primx, utils
19 20 21


@framework.static_only
22
def forward_grad(outputs, inputs, grad_inputs=None):
23 24 25 26 27 28
    """Forward mode of automatic differentiation.

    .. note::
        **ONLY available in the static mode and primitive operators.**

    Args:
29
        outputs: The output tensor or tensors
30
        inputs: The input tensor or tensors
31
        grad_inputs: The gradient Tensor or Tensors of inputs which has 
32 33 34 35
            the same shape with inputs, Defaults to None, in this case is 
            equivalent to all ones .

    Returns:
36
        grad_outputs (Tensor|Sequence[Tensor]): The gradients for outputs.
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53

    Examples:

        .. code-block:: python

            import numpy as np
            import paddle

            paddle.enable_static()
            paddle.incubate.autograd.enable_prim()

            startup_program = paddle.static.Program()
            main_program = paddle.static.Program()

            with paddle.static.program_guard(main_program, startup_program):
                x = paddle.static.data('x', shape=[1], dtype='float32')
                y = x * x 
54
                y_grad = paddle.incubate.autograd.forward_grad(y, x)
55 56 57 58 59 60 61 62 63 64 65
                paddle.incubate.autograd.prim2orig()

            exe = paddle.static.Executor()
            exe.run(startup_program)
            y_grad = exe.run(main_program, feed={'x': np.array([2.]).astype('float32')}, fetch_list=[y_grad])
            print(y_grad)
            # [array([4.], dtype=float32)]

            paddle.incubate.autograd.disable_prim()
            paddle.disable_static()
    """
66 67
    if not utils.prim_enabled():
        raise RuntimeError('forward_grad must be running on primitive'
68 69
                           'operators, use enable_prim to turn it on.')

70 71 72
    if not isinstance(outputs, (framework.Variable, typing.Sequence)):
        raise TypeError(f'Expected outputs is Tensor|Sequence[Tesnor], '
                        f'but got {type(outputs)}.')
73 74 75 76 77

    if not isinstance(inputs, (framework.Variable, typing.Sequence)):
        raise TypeError(f'Expected inputs is Tensor|Sequence[Tesnor], '
                        f'but got {type(inputs)}.')

78 79
    ys, xs, xs_dot = utils.as_tensors(outputs), utils.as_tensors(
        inputs), utils.as_tensors(grad_inputs)
80 81 82 83 84 85 86 87 88 89 90

    block = framework.default_main_program().current_block()
    if any(x.block != block for x in xs + ys):
        raise RuntimeError(
            'Variable in inputs and targets should exist in current block of '
            'main program.')

    primx.orig2prim(block)
    ad = primx.Transform(ys[0].block)
    _, ys_dot = ad.linearize(xs, ys, xs_dot)

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
    return ys_dot[0] if isinstance(outputs, framework.Variable) else ys_dot


@framework.static_only
def grad(outputs, inputs, grad_outputs=None):
    """Reverse mode of automatic differentiation.

    .. note::
        **ONLY available in the static mode and primitive operators**

    Args:
        outputs (Tensor|Sequence[Tensor]): The output Tensor or Tensors.
        inputs (Tensor|Sequence[Tensor]): The input Tensor or Tensors.
        grad_outputs (Tensor|Sequence[Tensor]): The gradient Tensor or 
            Tensors of outputs which has the same shape with outputs, Defaults 
            to None, in this case is equivalent to all ones .

    Returns:
        grad_inputs (Tensor|Tensors): The gradients for inputs. 

    Examples:

        .. code-block:: python

            import numpy as np
            import paddle
            paddle.enable_static()
            paddle.incubate.autograd.enable_prim()
            startup_program = paddle.static.Program()
            main_program = paddle.static.Program()
            with paddle.static.program_guard(main_program, startup_program):
                x = paddle.static.data('x', shape=[1], dtype='float32')
                x.stop_gradients = False
                y = x * x 
                x_grad = paddle.incubate.autograd.grad(y, x)
                paddle.incubate.autograd.prim2orig()
            exe = paddle.static.Executor()
            exe.run(startup_program)
            x_grad = exe.run(main_program, feed={'x': np.array([2.]).astype('float32')}, fetch_list=[x_grad])
            print(x_grad)
            # [array([4.], dtype=float32)]
            paddle.incubate.autograd.disable_prim()
            paddle.disable_static()
    """

    if not utils.prim_enabled():
        return backward.gradients(outputs, inputs, grad_outputs)

    if not isinstance(outputs, (framework.Variable, typing.Sequence)):
        raise TypeError(f'Expected outputs is Tensor|Sequence[Tesnor], '
                        f'but got {type(outputs)}.')

    if not isinstance(inputs, (framework.Variable, typing.Sequence)):
        raise TypeError(f'Expected inputs is Tensor|Sequence[Tesnor], '
                        f'but got {type(inputs)}.')

    ys, xs, ys_bar = utils.as_tensors(outputs), utils.as_tensors(
        inputs), utils.as_tensors(grad_outputs)
    block = framework.default_main_program().current_block()
    if any((x is not None and x.block != block) for x in xs + ys):
        raise RuntimeError(
            'Variable in inputs and outputs should be None or in current block of main program'
        )

    # TODO(Tongxin) without any prior knowledge about whether the program
    # is completely lowered to primitive ops, it's mandatory to run the lowering
    # pass once and again. This is obviously inefficient and needs to be
    # optimized.
    primx.orig2prim(block)
    ad = primx.Transform(block)
    xs_dot, ys_dot = ad.linearize(xs, ys)
    if any(var is None for var in ys_dot):
        raise RuntimeError(
            'Grads cannot be computed. The given outputs does not depend on inputs'
        )
    ys_bar, xs_bar = ad.transpose(ys_dot, xs_dot, ys_bar)

    # remove xs_dot and their constructor ops
    op_indexes = []
    for var in xs_dot:
        if var is not None:
            op_index = block.ops.index(var.op)
            if op_index < 0:
                raise ValueError(
                    f'op_index should be greater than or equal to 0, but op_index={op_index}.'
                )
            op_indexes.append(op_index)

    ad.erase_ops(sorted(op_indexes))
    ad.erase_dots(xs_dot)

    return xs_bar[0] if isinstance(inputs, framework.Variable) else xs_bar