param_attr.py 3.5 KB
Newer Older
D
dzhwinter 已提交
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
D
dzhwinter 已提交
2
#
F
fengjiayi 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
F
fengjiayi 已提交
9 10 11 12 13
# 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.
F
update  
fengjiayi 已提交
14

Y
Yu Yang 已提交
15 16 17
from initializer import Initializer, Xavier, Constant
from regularizer import WeightDecayRegularizer

18 19 20 21
__all__ = [
    'ParamAttr',
    'WeightNormParamAttr',
]
Y
Yu Yang 已提交
22

Y
Yu Yang 已提交
23 24 25 26 27 28 29

class ParamAttr(object):
    def __init__(self,
                 name=None,
                 initializer=None,
                 learning_rate=1.0,
                 regularizer=None,
Y
Yu Yang 已提交
30
                 trainable=True,
F
fengjiayi 已提交
31
                 gradient_clip=None):
Y
Yu Yang 已提交
32 33 34 35 36
        self.name = name
        self.initializer = initializer
        self.learning_rate = learning_rate
        self.regularizer = regularizer
        self.trainable = trainable
F
fengjiayi 已提交
37
        self.gradient_clip = gradient_clip
Y
Yu Yang 已提交
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59

    def set_default_initializer(self, initializer):
        if initializer is None:
            if self.initializer is None:
                raise ValueError("ParamAttr.initializer is not set")
            return

        if self.initializer is not None:
            return

        self.initializer = initializer

    def set_default_param_initializer(self):
        self.set_default_initializer(Xavier())

    def set_default_bias_initializer(self):
        self.set_default_initializer(Constant(0.0))

    @staticmethod
    def to_attr(arg):
        if arg is None:
            return ParamAttr()
60 61
        elif isinstance(arg, list) or isinstance(arg, tuple):
            return [ParamAttr.to_attr(a) for a in arg]
Y
Yu Yang 已提交
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77
        elif isinstance(arg, ParamAttr):
            return arg
        elif isinstance(arg, str) or isinstance(arg, unicode):
            return ParamAttr(name=arg)
        elif isinstance(arg, Initializer):
            return ParamAttr(initializer=arg)
        elif isinstance(arg, WeightDecayRegularizer):
            return ParamAttr(regularizer=arg)
        elif isinstance(arg, bool):
            return ParamAttr.to_attr(None) if arg else False
        else:
            raise TypeError("{0} cast to ParamAttr".format(type(arg)))

    def to_kwargs(self, with_initializer=False):
        kwargs = {
            'name': self.name,
G
guosheng 已提交
78 79 80
            'optimize_attr': {
                'learning_rate': self.learning_rate
            },
Y
Yu Yang 已提交
81
            'regularizer': self.regularizer,
Y
Yu Yang 已提交
82
            'trainable': self.trainable,
F
fengjiayi 已提交
83
            'gradient_clip_attr': self.gradient_clip
Y
Yu Yang 已提交
84 85 86 87
        }
        if with_initializer:
            kwargs['initializer'] = self.initializer
        return kwargs
G
guosheng 已提交
88 89 90 91 92 93 94 95 96 97


class WeightNormParamAttr(ParamAttr):
    """
    Used for weight normalization. Any field in ParamAttr can also be set here.
    Besides, an extra field dim can be set to indicate the dimension except 
    which to normalize.
    """
    # List to record the parameters reparameterized by weight normalization.
    # If these parameters are treated as Variable rather than Parameter,
98
    # it can be used to discriminate these parameters and help to serialize
G
guosheng 已提交
99 100 101 102 103 104
    # these paramters for inference.
    params_with_weight_norm = []

    def __init__(self, dim=None, **kwargs):
        super(WeightNormParamAttr, self).__init__(**kwargs)
        self.dim = dim