m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add paddle.nn.initializer API, including: Normal, TruncatedNormal, Uniform, XavierNormal, XavierUniform, Assign (#27769)

Browse Source
my_2.0rc
furnace 4 years ago committed by GitHub
parent a4f850748a
commit 8e70b18e6c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
7 changed files with 892 additions and 12 deletions
Show Stats Download Patch File Download Diff File

3
python/paddle/fluid/tests/unittests/dygraph_to_static/simnet_dygraph_model_v2.py
Unescape View File

@ -41,7 +41,8 @@ class EmbeddingLayer(object):
is_sparse=True,
padding_idx=self.padding_idx,
param_attr=paddle.ParamAttr(
name=self.name, initializer=paddle.nn.initializer.Xavier()))
name=self.name,
initializer=paddle.nn.initializer.XavierUniform()))
return emb

486
python/paddle/fluid/tests/unittests/test_initializer_nn.py
View File

File diff suppressed because it is too large Load Diff

31
python/paddle/nn/initializer/__init__.py
Unescape View File

@ -14,10 +14,6 @@
# TODO: define the initializers to create a Parameter in neural network
from ...fluid.initializer import Bilinear #DEFINE_ALIAS
from ...fluid.initializer import Normal #DEFINE_ALIAS
from ...fluid.initializer import TruncatedNormal #DEFINE_ALIAS
from ...fluid.initializer import Uniform #DEFINE_ALIAS
from ...fluid.initializer import Xavier #DEFINE_ALIAS
from . import constant
from .constant import Constant #DEFINE_ALIAS
@ -26,13 +22,26 @@ from . import kaiming
from .kaiming import KaimingNormal #DEFINE_ALIAS
from .kaiming import KaimingUniform #DEFINE_ALIAS
__all__ = [
'Bilinear',
'Normal',
'TruncatedNormal',
'Uniform',
'Xavier',
]
__all__ = ['Bilinear', ]
__all__ += constant.__all__
__all__ += kaiming.__all__
from . import xavier
from .xavier import XavierNormal #DEFINE_ALIAS
from .xavier import XavierUniform #DEFINE_ALIAS
from . import assign
from .assign import Assign #DEFINE_ALIAS
from . import normal
from .normal import Normal #DEFINE_ALIAS
from .normal import TruncatedNormal #DEFINE_ALIAS
from . import uniform
from .uniform import Uniform #DEFINE_ALIAS
__all__ += xavier.__all__
__all__ += assign.__all__
__all__ += normal.__all__
__all__ += uniform.__all__

100
python/paddle/nn/initializer/assign.py
Unescape View File

@ -0,0 +1,100 @@
# Copyright (c) 2020 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 ...fluid import framework
from ...fluid import core
from ...fluid import unique_name
from ...fluid.core import VarDesc
from ...fluid.data_feeder import check_type
from ...fluid.initializer import NumpyArrayInitializer
__all__ = ['Assign']
class Assign(NumpyArrayInitializer):
"""Init an parameter with a numpy array, list, or tensor.
Args:
value (Tensor|numpy.ndarray|list): numpy array, list, or tensor to initialize the parameter.
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:
A parameter initialized by the input numpy array, list, or tensor.
Examples:
.. code-block:: python
import paddle
import numpy as np
# numpy array
data_1 = paddle.ones(shape=[1, 2], dtype='float32')
weight_attr_1 = paddle.framework.ParamAttr(
name="linear_weight_1",
initializer=paddle.nn.initializer.Assign(np.array([2, 2])))
bias_attr_1 = paddle.framework.ParamAttr(
name="linear_bias_1",
initializer=paddle.nn.initializer.Assign(np.array([2])))
linear_1 = paddle.nn.Linear(2, 2, weight_attr=weight_attr_1, bias_attr=bias_attr_1)
# linear_1.weight: [2. 2.]
# linear_1.bias: [2.]
res_1 = linear(data_1)
# res_1: [6.]
# python list
data_2 = paddle.ones(shape=[1, 2], dtype='float32')
weight_attr_2 = paddle.framework.ParamAttr(
name="linear_weight_2",
initializer=paddle.nn.initializer.Assign([2, 2]))
bias_attr_2 = paddle.framework.ParamAttr(
name="linear_bias_2",
initializer=paddle.nn.initializer.Assign([2]))
linear_2 = paddle.nn.Linear(2, 2, weight_attr=weight_attr_2, bias_attr=bias_attr_2)
# linear_2.weight: [2. 2.]
# linear_2.bias: [2.]
res_2 = linear(data_2)
# res_2: [6.]
# tensor
data_3 = paddle.ones(shape=[1, 2], dtype='float32')
weight_attr_3 = paddle.framework.ParamAttr(
name="linear_weight_3",
initializer=paddle.nn.initializer.Assign(paddle.full([2], 2)))
bias_attr_3 = paddle.framework.ParamAttr(
name="linear_bias_3",
initializer=paddle.nn.initializer.Assign(paddle.full([1], 2)))
linear_3 = paddle.nn.Linear(2, 2, weight_attr=weight_attr_3, bias_attr=bias_attr_3)
# linear_3.weight: [2. 2.]
# linear_3.bias: [2.]
res_3 = linear(data_3)
# res_3: [6.]
"""
def __init__(self, value, name=None):
import numpy
check_type(value, 'value', (numpy.ndarray, list, framework.Variable),
'Assign')
if (isinstance(value, list)):
value = numpy.array(value)
# TODO: value is already is a tensor, accounting efficiency maybe it does not need to convert tensor to numpy data and then initialized.
if (isinstance(value, framework.Variable)):
value = value.numpy()
super(Assign, self).__init__(value)

100
python/paddle/nn/initializer/normal.py
Unescape View File

@ -0,0 +1,100 @@
# Copyright (c) 2020 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 ...fluid.initializer import NormalInitializer
from ...fluid.initializer import TruncatedNormalInitializer
__all__ = ['Normal', 'TruncatedNormal']
class Normal(NormalInitializer):
"""The Random Normal (Gaussian) distribution initializer.
Args:
mean (float, optional): mean of the normal distribution. The default value is 0.0.
std (float, optional): standard deviation of the normal distribution. The default value is 1.0.
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:
A parameter initialized by Random Normal (Gaussian) distribution.
Examples:
.. code-block:: python
import paddle
data = paddle.ones(shape=[3, 1, 2], dtype='float32')
weight_attr = paddle.framework.ParamAttr(
name="linear_weight",
initializer=paddle.nn.initializer.Normal(mean=0.0, std=2.0))
bias_attr = paddle.framework.ParamAttr(
name="linear_bias",
initializer=paddle.nn.initializer.Normal(mean=0.0, std=2.0))
linear = paddle.nn.Linear(2, 2, weight_attr=weight_attr, bias_attr=bias_attr)
# linear.weight: [[ 2.1973135 -2.2697184]
# [-1.9104223 -1.0541488]]
# linear.bias: [ 0.7885926 -0.74719954]
res = linear(data)
# res: [[[ 1.0754838 -4.071067 ]]
# [[ 1.0754838 -4.071067 ]]
# [[ 1.0754838 -4.071067 ]]]
"""
def __init__(self, mean=0.0, std=1.0, name=None):
assert mean is not None, 'mean should not be None'
assert std is not None, 'std should not be None'
super(Normal, self).__init__(loc=mean, scale=std, seed=0)
class TruncatedNormal(TruncatedNormalInitializer):
"""The Random TruncatedNormal (Gaussian) distribution initializer.
Args:
mean (float, optional): mean of the normal distribution. The default value is 0.0.
std (float, optional): standard deviation of the normal distribution. The default value is 1.0.
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:
A parameter initialized by Random TruncatedNormal (Gaussian) distribution.
Examples:
.. code-block:: python
import paddle
data = paddle.ones(shape=[3, 1, 2], dtype='float32')
weight_attr = paddle.framework.ParamAttr(
name="linear_weight",
initializer=paddle.nn.initializer.TruncatedNormal(mean=0.0, std=2.0))
bias_attr = paddle.framework.ParamAttr(
name="linear_bias",
initializer=paddle.nn.initializer.TruncatedNormal(mean=0.0, std=2.0))
linear = paddle.nn.Linear(2, 2, weight_attr=weight_attr, bias_attr=bias_attr)
# linear.weight: [[-1.0981836 1.4140984]
# [ 3.1390522 -2.8266568]]
# linear.bias: [-2.1546738 -1.6570673]
res = linear(data)
# res: [[[-0.11380529 -3.0696259 ]]
# [[-0.11380529 -3.0696259 ]]
# [[-0.11380529 -3.0696259 ]]
"""
def __init__(self, mean=0.0, std=1.0, name=None):
assert mean is not None, 'mean should not be None'
assert std is not None, 'std should not be None'
super(TruncatedNormal, self).__init__(loc=mean, scale=std, seed=0)

60
python/paddle/nn/initializer/uniform.py
Unescape View File

@ -0,0 +1,60 @@
# Copyright (c) 2020 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 ...fluid.initializer import UniformInitializer
__all__ = ['Uniform']
class Uniform(UniformInitializer):
"""The random uniform distribution initializer.
Args:
low (float, optional): lower boundary of the uniform distribution. The default value is -1.0.
high (float, optional): upper boundary of the uniform distribution. The default value is 1.0.
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:
A parameter initialized by random uniform distribution.
Examples:
.. code-block:: python
import paddle
data = paddle.ones(shape=[3, 1, 2], dtype='float32')
weight_attr = paddle.framework.ParamAttr(
name="linear_weight",
initializer=paddle.nn.initializer.Uniform(low=-0.5, high=0.5))
bias_attr = paddle.framework.ParamAttr(
name="linear_bias",
initializer=paddle.nn.initializer.Uniform(low=-0.5, high=0.5))
linear = paddle.nn.Linear(2, 2, weight_attr=weight_attr, bias_attr=bias_attr)
# linear.weight: [[-0.46245047 0.05260676]
# [ 0.38054508 0.29169726]]
# linear.bias: [-0.2734719 0.23939109]
res = linear(data)
# res: [[[-0.3553773 0.5836951]]
# [[-0.3553773 0.5836951]]
# [[-0.3553773 0.5836951]]]
"""
def __init__(self, low=-1.0, high=1.0, name=None):
assert low is not None, 'low should not be None'
assert high is not None, 'high should not be None'
assert high >= low, 'high should greater or equal than low'
super(Uniform, self).__init__(
low=low, high=high, seed=0, diag_num=0, diag_step=0, diag_val=1.0)

124
python/paddle/nn/initializer/xavier.py
Unescape View File

@ -0,0 +1,124 @@
# Copyright (c) 2020 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 ...fluid.initializer import XavierInitializer
__all__ = ['XavierNormal', 'XavierUniform']
class XavierNormal(XavierInitializer):
"""
This class implements the Xavier weight initializer from the paper
`Understanding the difficulty of training deep feedforward neural
networks <http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf>`_
by Xavier Glorot and Yoshua Bengio, using a normal distribution.
The mean is 0 and the standard deviation is
.. math::
\sqrt{\\frac{2.0}{fan\_in + fan\_out}}
Args:
fan_in (float, optional): fan_in for Xavier initialization, It is
inferred from the tensor. The default value is None.
fan_out (float, optional): fan_out for Xavier initialization, it is
inferred from the tensor. The default value is None.
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:
A parameter initialized by Xavier weight, using a normal distribution.
Examples:
.. code-block:: python
import paddle
data = paddle.ones(shape=[3, 1, 2], dtype='float32')
weight_attr = paddle.framework.ParamAttr(
name="linear_weight",
initializer=paddle.nn.initializer.XavierNormal())
bias_attr = paddle.framework.ParamAttr(
name="linear_bias",
initializer=paddle.nn.initializer.XavierNormal())
linear = paddle.nn.Linear(2, 2, weight_attr=weight_attr, bias_attr=bias_attr)
# inear.weight: [[ 0.06910077 -0.18103665]
# [-0.02546741 -1.0402188 ]]
# linear.bias: [-0.5012929 0.12418364]
res = linear(data)
# res: [[[-0.4576595 -1.0970719]]
# [[-0.4576595 -1.0970719]]
# [[-0.4576595 -1.0970719]]]
"""
def __init__(self, fan_in=None, fan_out=None, name=None):
super(XavierNormal, self).__init__(
uniform=False, fan_in=fan_in, fan_out=fan_out, seed=0)
class XavierUniform(XavierInitializer):
"""
This class implements the Xavier weight initializer from the paper
`Understanding the difficulty of training deep feedforward neural
networks <http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf>`_
by Xavier Glorot and Yoshua Bengio.
This initializer is designed to keep the scale of the gradients
approximately same in all the layers. In case of Uniform distribution,
the range is [-x, x], where
.. math::
x = \sqrt{\\frac{6.0}{fan\_in + fan\_out}}
Args:
fan_in (float, optional): fan_in for Xavier initialization, it is
inferred from the tensor. The default value is None.
fan_out (float, optional): fan_out for Xavier initialization, it is
inferred from the tensor. The default value is None.
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:
A parameter initialized by Xavier weight, using a uniform distribution.
Examples:
.. code-block:: python
import paddle
data = paddle.ones(shape=[3, 1, 2], dtype='float32')
weight_attr = paddle.framework.ParamAttr(
name="linear_weight",
initializer=paddle.nn.initializer.XavierUniform())
bias_attr = paddle.framework.ParamAttr(
name="linear_bias",
initializer=paddle.nn.initializer.XavierUniform())
linear = paddle.nn.Linear(2, 2, weight_attr=weight_attr, bias_attr=bias_attr)
# linear.weight: [[-0.04229349 -1.1248565 ]
# [-0.10789523 -0.5938053 ]]
# linear.bias: [ 1.1983747 -0.40201235]
res = linear(data)
# res: [[[ 1.0481861 -2.1206741]]
# [[ 1.0481861 -2.1206741]]
# [[ 1.0481861 -2.1206741]]]
"""
def __init__(self, fan_in=None, fan_out=None, name=None):
super(XavierUniform, self).__init__(
uniform=True, fan_in=fan_in, fan_out=fan_out, seed=0)
Write Preview
Loading…
Cancel
Save