layers.py 4.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
# 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.

import contextlib
import sys
import numpy as np
M
minqiyang 已提交
18
import collections
19 20 21 22 23

from paddle.fluid import core
from paddle.fluid import framework
from paddle.fluid.imperative import base

X
Xin Pan 已提交
24
__all__ = ['Layer', 'PyLayer']
25 26


X
Xin Pan 已提交
27 28 29
class Layer(core.Layer):
    """Layers composed of operators."""

M
minqiyang 已提交
30
    def __init__(self, dtype=core.VarDesc.VarType.FP32, name=None):
X
Xin Pan 已提交
31
        self._built = False
M
minqiyang 已提交
32
        self._dtype = dtype
33

X
Xin Pan 已提交
34
    def parameters(self):
M
minqiyang 已提交
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
        params = []
        for key in self.__dict__.keys():
            value = self.__dict__[key]
            if isinstance(value, framework.Parameter):
                params.append(value)
            elif isinstance(value, core.Layer):
                params.extend(value.parameters())
            elif isinstance(value, collections.Container):
                if len(value) == 0:
                    continue
                if isinstance(value[0], framework.Parameter):
                    params.extend(value)
                elif isinstance(value[0], core.Layer):
                    for v in value:
                        params.extend(v.parameters())

        return params
X
Xin Pan 已提交
52

X
Xin Pan 已提交
53 54
    def clear_gradients(self):
        for p in self.parameters():
M
minqiyang 已提交
55
            p._clear_gradient()
X
Xin Pan 已提交
56

57 58 59
    def _build_once(self, inputs):
        pass

60
    def __call__(self, *inputs):
X
Xin Pan 已提交
61
        if not self._built:
62 63
            self._build_once(*inputs)

64
        outputs = self.forward(*inputs)
X
Xin Pan 已提交
65
        self._built = True
M
minqiyang 已提交
66
        return outputs
M
minqiyang 已提交
67

68 69
    def forward(self, *inputs):
        raise NotImplementedError
X
Xin Pan 已提交
70 71 72 73 74

    def backward(self, *inputs):
        raise ValueError("Layer shouldn't implement backward")


X
Xin Pan 已提交
75
class PyLayer(core.PyLayer):
X
Xin Pan 已提交
76 77
    """Layers composed of user-defined python codes."""

X
Xin Pan 已提交
78 79
    def __init__(self):
        super(PyLayer, self).__init__()
X
Xin Pan 已提交
80

X
Xin Pan 已提交
81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99
    @classmethod
    def _do_forward(cls, inputs):
        return cls._to_tuple(cls.forward(inputs))

    @classmethod
    def _do_backward(cls, inputs):
        return cls._to_tuple(cls.backward(inputs))

    @staticmethod
    def _to_tuple(inputs):
        if not isinstance(inputs, list) and not isinstance(inputs, tuple):
            inputs = [inputs]
        ret = []
        for inp in inputs:
            tensor = core.LoDTensor()
            tensor.set(inp, core.CPUPlace())
            ret.append(tensor)
        return tuple(ret)

X
Xin Pan 已提交
100
    @staticmethod
M
minqiyang 已提交
101
    def forward(*inputs):
X
Xin Pan 已提交
102 103
        raise NotImplementedError

X
Xin Pan 已提交
104
    @staticmethod
M
minqiyang 已提交
105
    def backward(*douts):
X
Xin Pan 已提交
106
        raise NotImplementedError
X
Xin Pan 已提交
107 108

    @classmethod
M
minqiyang 已提交
109
    def __call__(cls, *inputs):
X
Xin Pan 已提交
110 111
        tracer = framework._imperative_tracer()
        block = framework.default_main_program().current_block()
M
minqiyang 已提交
112
        ivar_inputs = [x._ivar for x in inputs]
X
Xin Pan 已提交
113

X
polish  
Xin Pan 已提交
114 115
        if not hasattr(cls, 'forward_id'):
            cls.forward_id = core.PyLayer.num_funcs() + 1
X
Xin Pan 已提交
116
            PyLayer.register_func(cls.forward_id, cls._do_forward)
X
polish  
Xin Pan 已提交
117
            cls.backward_id = core.PyLayer.num_funcs() + 1
X
Xin Pan 已提交
118
            PyLayer.register_func(cls.backward_id, cls._do_backward)
X
Xin Pan 已提交
119 120

        iop = core.OpBase()
X
polish  
Xin Pan 已提交
121 122
        iop.forward_id = cls.forward_id
        iop.backward_id = cls.backward_id
X
Xin Pan 已提交
123
        block.ops.append(iop)
M
minqiyang 已提交
124
        ivars = tracer.py_trace(iop, ivar_inputs, False)
X
Xin Pan 已提交
125 126
        ret = []
        for ivar in ivars:
M
minqiyang 已提交
127
            tensor = ivar.value().get_tensor()
X
Xin Pan 已提交
128 129 130 131 132 133 134 135 136
            py_var = framework.Variable(
                block,
                type=core.VarDesc.VarType.LOD_TENSOR,
                name=None,
                shape=tensor.shape(),
                dtype=tensor._dtype(),
                ivar=ivar)
            ret.append(py_var)
        return ret