diff --git a/ppcls/modeling/architectures/ghostnet.py b/ppcls/modeling/architectures/ghostnet.py index 50978b785..b37f2230f 100644 --- a/ppcls/modeling/architectures/ghostnet.py +++ b/ppcls/modeling/architectures/ghostnet.py @@ -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"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at +# 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. +# 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 math import paddle @@ -23,16 +23,14 @@ from paddle.nn.initializer import Uniform class ConvBNLayer(nn.Layer): - def __init__( - self, - in_channels, - out_channels, - kernel_size, - stride=1, - groups=1, - act="relu", - name=None - ): + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride=1, + groups=1, + act="relu", + name=None): super(ConvBNLayer, self).__init__() self._conv = Conv2d( in_channels=in_channels, @@ -41,19 +39,24 @@ class ConvBNLayer(nn.Layer): stride=stride, padding=(kernel_size - 1) // 2, groups=groups, - weight_attr=ParamAttr(initializer=nn.initializer.MSRA(), name=name + "_weights"), - bias_attr=False - ) + weight_attr=ParamAttr( + initializer=nn.initializer.MSRA(), name=name + "_weights"), + bias_attr=False) bn_name = name + "_bn" # In the old version, moving_variance_name was name + "_variance" self._batch_norm = BatchNorm( num_channels=out_channels, act=act, - param_attr=ParamAttr(name=bn_name + "_scale", regularizer=L2DecayRegularizer(regularization_coeff=0.0)), - bias_attr=ParamAttr(name=bn_name + "_offset", regularizer=L2DecayRegularizer(regularization_coeff=0.0)), + param_attr=ParamAttr( + 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_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): @@ -63,12 +66,7 @@ class ConvBNLayer(nn.Layer): class SEBlock(nn.Layer): - def __init__( - self, - num_channels, - reduction_ratio=4, - name=None - ): + def __init__(self, num_channels, reduction_ratio=4, name=None): super(SEBlock, self).__init__() self.pool2d_gap = AdaptiveAvgPool2d(1) self._num_channels = num_channels @@ -77,16 +75,16 @@ class SEBlock(nn.Layer): self.squeeze = Linear( num_channels, med_ch, - weight_attr=ParamAttr(initializer=Uniform(-stdv, stdv), name=name + "_1_weights"), - bias_attr=ParamAttr(name=name + "_1_offset") - ) + weight_attr=ParamAttr( + initializer=Uniform(-stdv, stdv), name=name + "_1_weights"), + bias_attr=ParamAttr(name=name + "_1_offset")) stdv = 1.0 / math.sqrt(med_ch * 1.0) self.excitation = Linear( med_ch, num_channels, - weight_attr=ParamAttr(initializer=Uniform(-stdv, stdv), name=name+"_2_weights"), - bias_attr=ParamAttr(name=name+"_2_offset") - ) + weight_attr=ParamAttr( + initializer=Uniform(-stdv, stdv), name=name + "_2_weights"), + bias_attr=ParamAttr(name=name + "_2_offset")) def forward(self, inputs): pool = self.pool2d_gap(inputs) @@ -95,23 +93,22 @@ class SEBlock(nn.Layer): squeeze = F.relu(squeeze) excitation = self.excitation(squeeze) 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 return out class GhostModule(nn.Layer): - def __init__( - self, - in_channels, - output_channels, - kernel_size=1, - ratio=2, - dw_size=3, - stride=1, - relu=True, - name=None - ): + def __init__(self, + in_channels, + output_channels, + kernel_size=1, + ratio=2, + dw_size=3, + stride=1, + relu=True, + name=None): super(GhostModule, self).__init__() init_channels = int(math.ceil(output_channels / ratio)) new_channels = int(init_channels * (ratio - 1)) @@ -122,8 +119,7 @@ class GhostModule(nn.Layer): stride=stride, groups=1, act="relu" if relu else None, - name=name + "_primary_conv" - ) + name=name + "_primary_conv") self.cheap_operation = ConvBNLayer( in_channels=init_channels, out_channels=new_channels, @@ -131,8 +127,7 @@ class GhostModule(nn.Layer): stride=1, groups=init_channels, act="relu" if relu else None, - name=name + "_cheap_operation" - ) + name=name + "_cheap_operation") def forward(self, inputs): x = self.primary_conv(inputs) @@ -142,16 +137,14 @@ class GhostModule(nn.Layer): class GhostBottleneck(nn.Layer): - def __init__( - self, - in_channels, - hidden_dim, - output_channels, - kernel_size, - stride, - use_se, - name=None - ): + def __init__(self, + in_channels, + hidden_dim, + output_channels, + kernel_size, + stride, + use_se, + name=None): super(GhostBottleneck, self).__init__() self._stride = stride self._use_se = use_se @@ -163,8 +156,7 @@ class GhostBottleneck(nn.Layer): kernel_size=1, stride=1, relu=True, - name=name+"_ghost_module_1" - ) + name=name + "_ghost_module_1") if stride == 2: self.depthwise_conv = ConvBNLayer( in_channels=hidden_dim, @@ -173,20 +165,17 @@ class GhostBottleneck(nn.Layer): stride=stride, groups=hidden_dim, 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: - self.se_block = SEBlock( - num_channels=hidden_dim, - name=name + "_se" - ) + self.se_block = SEBlock(num_channels=hidden_dim, name=name + "_se") self.ghost_module_2 = GhostModule( in_channels=hidden_dim, output_channels=output_channels, kernel_size=1, relu=False, - name=name + "_ghost_module_2" - ) + name=name + "_ghost_module_2") if stride != 1 or in_channels != output_channels: self.shortcut_depthwise = ConvBNLayer( in_channels=in_channels, @@ -195,7 +184,8 @@ class GhostBottleneck(nn.Layer): stride=stride, groups=in_channels, 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( in_channels=in_channels, @@ -204,8 +194,7 @@ class GhostBottleneck(nn.Layer): stride=1, groups=1, act=None, - name=name + "_shortcut_conv" - ) + name=name + "_shortcut_conv") def forward(self, inputs): x = self.ghost_module_1(inputs) @@ -253,8 +242,7 @@ class GhostNet(nn.Layer): stride=2, groups=1, act="relu", - name="conv1" - ) + name="conv1") # build inverted residual blocks idx = 0 self.ghost_bottleneck_list = [] @@ -271,9 +259,7 @@ class GhostNet(nn.Layer): kernel_size=k, stride=s, use_se=use_se, - name="_ghostbottleneck_" + str(idx) - ) - ) + name="_ghostbottleneck_" + str(idx))) self.ghost_bottleneck_list.append(ghost_bottleneck) idx += 1 # build last several layers @@ -286,8 +272,7 @@ class GhostNet(nn.Layer): stride=1, groups=1, act="relu", - name="conv_last" - ) + name="conv_last") self.pool2d_gap = AdaptiveAvgPool2d(1) in_channels = output_channels self._fc0_output_channels = 1280 @@ -297,16 +282,15 @@ class GhostNet(nn.Layer): kernel_size=1, stride=1, act="relu", - name="fc_0" - ) + name="fc_0") self.dropout = nn.Dropout(p=0.2) stdv = 1.0 / math.sqrt(self._fc0_output_channels * 1.0) self.fc_1 = Linear( self._fc0_output_channels, class_dim, - weight_attr=ParamAttr(name="fc_1_weights", initializer=Uniform(-stdv, stdv)), - bias_attr=ParamAttr(name="fc_1_offset") - ) + weight_attr=ParamAttr( + name="fc_1_weights", initializer=Uniform(-stdv, stdv)), + bias_attr=ParamAttr(name="fc_1_offset")) def forward(self, inputs): x = self.conv1(inputs)