mirrors/PaddleClas
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/PaddleClas.git synced 2025-06-03 21:55:06 +08:00
Code Issues Projects Releases Wiki Activity

formatting

Browse Source
This commit is contained in:
weishengyu 2020-10-22 17:45:14 +08:00
parent e0859f7498
commit 2ec1d73ece
1 changed files with 71 additions and 87 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

120
ppcls/modeling/architectures/ghostnet.py
View File

@ -1,16 +1,16 @@
#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve. # copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# #
#Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
#You may obtain a copy of the License at # You may obtain a copy of the License at
# #
# http://www.apache.org/licenses/LICENSE-2.0 # http://www.apache.org/licenses/LICENSE-2.0
# #
#Unless required by applicable law or agreed to in writing, software # Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS, # distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
#limitations under the License. # limitations under the License.
import math import math
import paddle import paddle
@ -23,16 +23,14 @@ from paddle.nn.initializer import Uniform
class ConvBNLayer(nn.Layer): class ConvBNLayer(nn.Layer):
def __init__( def __init__(self,
self,
in_channels, in_channels,
out_channels, out_channels,
kernel_size, kernel_size,
stride=1, stride=1,
groups=1, groups=1,
act="relu", act="relu",
name=None name=None):
):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2d( self._conv = Conv2d(
in_channels=in_channels, in_channels=in_channels,
@ -41,19 +39,24 @@ class ConvBNLayer(nn.Layer):
stride=stride, stride=stride,
padding=(kernel_size - 1) // 2, padding=(kernel_size - 1) // 2,
groups=groups, groups=groups,
weight_attr=ParamAttr(initializer=nn.initializer.MSRA(), name=name + "_weights"), weight_attr=ParamAttr(
bias_attr=False initializer=nn.initializer.MSRA(), name=name + "_weights"),
) bias_attr=False)
bn_name = name + "_bn" bn_name = name + "_bn"
# In the old version, moving_variance_name was name + "_variance" # In the old version, moving_variance_name was name + "_variance"
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
num_channels=out_channels, num_channels=out_channels,
act=act, act=act,
param_attr=ParamAttr(name=bn_name + "_scale", regularizer=L2DecayRegularizer(regularization_coeff=0.0)), param_attr=ParamAttr(
bias_attr=ParamAttr(name=bn_name + "_offset", regularizer=L2DecayRegularizer(regularization_coeff=0.0)), name=bn_name + "_scale",
regularizer=L2DecayRegularizer(regularization_coeff=0.0)),
bias_attr=ParamAttr(
name=bn_name + "_offset",
regularizer=L2DecayRegularizer(regularization_coeff=0.0)),
moving_mean_name=bn_name + "_mean", moving_mean_name=bn_name + "_mean",
moving_variance_name=name + "_variance" # wrong due to an old typo, will be fixed later. moving_variance_name=name +
"_variance" # wrong due to an old typo, will be fixed later.
) )
def forward(self, inputs): def forward(self, inputs):
@ -63,12 +66,7 @@ class ConvBNLayer(nn.Layer):
class SEBlock(nn.Layer): class SEBlock(nn.Layer):
def __init__( def __init__(self, num_channels, reduction_ratio=4, name=None):
self,
num_channels,
reduction_ratio=4,
name=None
):
super(SEBlock, self).__init__() super(SEBlock, self).__init__()
self.pool2d_gap = AdaptiveAvgPool2d(1) self.pool2d_gap = AdaptiveAvgPool2d(1)
self._num_channels = num_channels self._num_channels = num_channels
@ -77,16 +75,16 @@ class SEBlock(nn.Layer):
self.squeeze = Linear( self.squeeze = Linear(
num_channels, num_channels,
med_ch, med_ch,
weight_attr=ParamAttr(initializer=Uniform(-stdv, stdv), name=name + "_1_weights"), weight_attr=ParamAttr(
bias_attr=ParamAttr(name=name + "_1_offset") initializer=Uniform(-stdv, stdv), name=name + "_1_weights"),
) bias_attr=ParamAttr(name=name + "_1_offset"))
stdv = 1.0 / math.sqrt(med_ch * 1.0) stdv = 1.0 / math.sqrt(med_ch * 1.0)
self.excitation = Linear( self.excitation = Linear(
med_ch, med_ch,
num_channels, num_channels,
weight_attr=ParamAttr(initializer=Uniform(-stdv, stdv), name=name+"_2_weights"), weight_attr=ParamAttr(
bias_attr=ParamAttr(name=name+"_2_offset") initializer=Uniform(-stdv, stdv), name=name + "_2_weights"),
) bias_attr=ParamAttr(name=name + "_2_offset"))
def forward(self, inputs): def forward(self, inputs):
pool = self.pool2d_gap(inputs) pool = self.pool2d_gap(inputs)
@ -95,14 +93,14 @@ class SEBlock(nn.Layer):
squeeze = F.relu(squeeze) squeeze = F.relu(squeeze)
excitation = self.excitation(squeeze) excitation = self.excitation(squeeze)
excitation = paddle.fluid.layers.clip(x=excitation, min=0, max=1) excitation = paddle.fluid.layers.clip(x=excitation, min=0, max=1)
excitation = paddle.reshape(excitation, shape=[-1, self._num_channels, 1, 1]) excitation = paddle.reshape(
excitation, shape=[-1, self._num_channels, 1, 1])
out = inputs * excitation out = inputs * excitation
return out return out
class GhostModule(nn.Layer): class GhostModule(nn.Layer):
def __init__( def __init__(self,
self,
in_channels, in_channels,
output_channels, output_channels,
kernel_size=1, kernel_size=1,
@ -110,8 +108,7 @@ class GhostModule(nn.Layer):
dw_size=3, dw_size=3,
stride=1, stride=1,
relu=True, relu=True,
name=None name=None):
):
super(GhostModule, self).__init__() super(GhostModule, self).__init__()
init_channels = int(math.ceil(output_channels / ratio)) init_channels = int(math.ceil(output_channels / ratio))
new_channels = int(init_channels * (ratio - 1)) new_channels = int(init_channels * (ratio - 1))
@ -122,8 +119,7 @@ class GhostModule(nn.Layer):
stride=stride, stride=stride,
groups=1, groups=1,
act="relu" if relu else None, act="relu" if relu else None,
name=name + "_primary_conv" name=name + "_primary_conv")
)
self.cheap_operation = ConvBNLayer( self.cheap_operation = ConvBNLayer(
in_channels=init_channels, in_channels=init_channels,
out_channels=new_channels, out_channels=new_channels,
@ -131,8 +127,7 @@ class GhostModule(nn.Layer):
stride=1, stride=1,
groups=init_channels, groups=init_channels,
act="relu" if relu else None, act="relu" if relu else None,
name=name + "_cheap_operation" name=name + "_cheap_operation")
)
def forward(self, inputs): def forward(self, inputs):
x = self.primary_conv(inputs) x = self.primary_conv(inputs)
@ -142,16 +137,14 @@ class GhostModule(nn.Layer):
class GhostBottleneck(nn.Layer): class GhostBottleneck(nn.Layer):
def __init__( def __init__(self,
self,
in_channels, in_channels,
hidden_dim, hidden_dim,
output_channels, output_channels,
kernel_size, kernel_size,
stride, stride,
use_se, use_se,
name=None name=None):
):
super(GhostBottleneck, self).__init__() super(GhostBottleneck, self).__init__()
self._stride = stride self._stride = stride
self._use_se = use_se self._use_se = use_se
@ -163,8 +156,7 @@ class GhostBottleneck(nn.Layer):
kernel_size=1, kernel_size=1,
stride=1, stride=1,
relu=True, relu=True,
name=name+"_ghost_module_1" name=name + "_ghost_module_1")
)
if stride == 2: if stride == 2:
self.depthwise_conv = ConvBNLayer( self.depthwise_conv = ConvBNLayer(
in_channels=hidden_dim, in_channels=hidden_dim,
@ -173,20 +165,17 @@ class GhostBottleneck(nn.Layer):
stride=stride, stride=stride,
groups=hidden_dim, groups=hidden_dim,
act=None, act=None,
name=name+"_depthwise_depthwise" # looks strange due to an old typo, will be fixed later. name=name +
"_depthwise_depthwise" # looks strange due to an old typo, will be fixed later.
) )
if use_se: if use_se:
self.se_block = SEBlock( self.se_block = SEBlock(num_channels=hidden_dim, name=name + "_se")
num_channels=hidden_dim,
name=name + "_se"
)
self.ghost_module_2 = GhostModule( self.ghost_module_2 = GhostModule(
in_channels=hidden_dim, in_channels=hidden_dim,
output_channels=output_channels, output_channels=output_channels,
kernel_size=1, kernel_size=1,
relu=False, relu=False,
name=name + "_ghost_module_2" name=name + "_ghost_module_2")
)
if stride != 1 or in_channels != output_channels: if stride != 1 or in_channels != output_channels:
self.shortcut_depthwise = ConvBNLayer( self.shortcut_depthwise = ConvBNLayer(
in_channels=in_channels, in_channels=in_channels,
@ -195,7 +184,8 @@ class GhostBottleneck(nn.Layer):
stride=stride, stride=stride,
groups=in_channels, groups=in_channels,
act=None, act=None,
name=name + "_shortcut_depthwise_depthwise" # looks strange due to an old typo, will be fixed later. name=name +
"_shortcut_depthwise_depthwise" # looks strange due to an old typo, will be fixed later.
) )
self.shortcut_conv = ConvBNLayer( self.shortcut_conv = ConvBNLayer(
in_channels=in_channels, in_channels=in_channels,
@ -204,8 +194,7 @@ class GhostBottleneck(nn.Layer):
stride=1, stride=1,
groups=1, groups=1,
act=None, act=None,
name=name + "_shortcut_conv" name=name + "_shortcut_conv")
)
def forward(self, inputs): def forward(self, inputs):
x = self.ghost_module_1(inputs) x = self.ghost_module_1(inputs)
@ -253,8 +242,7 @@ class GhostNet(nn.Layer):
stride=2, stride=2,
groups=1, groups=1,
act="relu", act="relu",
name="conv1" name="conv1")
)
# build inverted residual blocks # build inverted residual blocks
idx = 0 idx = 0
self.ghost_bottleneck_list = [] self.ghost_bottleneck_list = []
@ -271,9 +259,7 @@ class GhostNet(nn.Layer):
kernel_size=k, kernel_size=k,
stride=s, stride=s,
use_se=use_se, use_se=use_se,
name="_ghostbottleneck_" + str(idx) name="_ghostbottleneck_" + str(idx)))
)
)
self.ghost_bottleneck_list.append(ghost_bottleneck) self.ghost_bottleneck_list.append(ghost_bottleneck)
idx += 1 idx += 1
# build last several layers # build last several layers
@ -286,8 +272,7 @@ class GhostNet(nn.Layer):
stride=1, stride=1,
groups=1, groups=1,
act="relu", act="relu",
name="conv_last" name="conv_last")
)
self.pool2d_gap = AdaptiveAvgPool2d(1) self.pool2d_gap = AdaptiveAvgPool2d(1)
in_channels = output_channels in_channels = output_channels
self._fc0_output_channels = 1280 self._fc0_output_channels = 1280
@ -297,16 +282,15 @@ class GhostNet(nn.Layer):
kernel_size=1, kernel_size=1,
stride=1, stride=1,
act="relu", act="relu",
name="fc_0" name="fc_0")
)
self.dropout = nn.Dropout(p=0.2) self.dropout = nn.Dropout(p=0.2)
stdv = 1.0 / math.sqrt(self._fc0_output_channels * 1.0) stdv = 1.0 / math.sqrt(self._fc0_output_channels * 1.0)
self.fc_1 = Linear( self.fc_1 = Linear(
self._fc0_output_channels, self._fc0_output_channels,
class_dim, class_dim,
weight_attr=ParamAttr(name="fc_1_weights", initializer=Uniform(-stdv, stdv)), weight_attr=ParamAttr(
bias_attr=ParamAttr(name="fc_1_offset") name="fc_1_weights", initializer=Uniform(-stdv, stdv)),
) bias_attr=ParamAttr(name="fc_1_offset"))
def forward(self, inputs): def forward(self, inputs):
x = self.conv1(inputs) x = self.conv1(inputs)