Skip to content

P
PaddleClas

PaddlePaddle / PaddleClas
大约 1 年 前同步成功

通知 115
Star 4999
Fork 1114
P
PaddleClas
提交 8ff9d3fb 编写于 作者: W wqz960
浏览文件
操作

fix weight name

上级 ac50d65d
隐藏空白更改
内联 并排
Showing with 145 addition and 170 deletion
ppcls/modeling/architectures/__init__.py
浏览文件 @ 8ff9d3fb
......@@ -47,4 +47,4 @@ from .ghostnet import GhostNet_x0_5, GhostNet_x1_0, GhostNet_x1_3
# distillation model
from .distillation_models import ResNet50_vd_distill_MobileNetV3_large_x1_0, ResNeXt101_32x16d_wsl_distill_ResNet50_vd
from .csp_resnet import CSPResNet50_leaky
\ No newline at end of file
from .csp_resnet import CSPResNet50_leaky
ppcls/modeling/architectures/ghostnet.py
浏览文件 @ 8ff9d3fb
......@@ -37,65 +37,55 @@ class GhostNet():
def net(self, input, class_dim=1000):
# build first layer:
output_channel = int(self._make_divisible(16 * self.scale, 4))
x = self.conv_bn_layer(
input=input,
num_filters=output_channel,
filter_size=3,
stride=2,
groups=1,
act="relu",
name="conv1")
x = self.conv_bn_layer(input=input,
num_filters=output_channel,
filter_size=3,
stride=2,
groups=1,
act="relu",
name="conv1")
# build inverted residual blocks
idx = 0
for k, exp_size, c, use_se, s in self.cfgs:
output_channel = int(self._make_divisible(c * self.scale, 4))
hidden_channel = int(
self._make_divisible(exp_size * self.scale, 4))
x = self.ghost_bottleneck(
inp=x,
hidden_dim=hidden_channel,
oup=output_channel,
kernel_size=k,
stride=s,
use_se=use_se,
name="ghost_bottle_" + str(idx))
hidden_channel = int(self._make_divisible(exp_size * self.scale, 4))
x = self.ghost_bottleneck(input=x,
hidden_dim=hidden_channel,
output=output_channel,
kernel_size=k,
stride=s,
use_se=use_se,
name="_ghostbottleneck_" + str(idx))
idx += 1
# build last several layers
output_channel = int(
self._make_divisible(exp_size * self.scale, 4))
x = self.conv_bn_layer(
input=x,
num_filters=output_channel,
filter_size=1,
stride=1,
groups=1,
act="relu",
name="conv2")
x = fluid.layers.pool2d(
input=x, pool_type='avg', global_pooling=True)
output_channel = int(self._make_divisible(exp_size * self.scale, 4))
x = self.conv_bn_layer(input=x,
num_filters=output_channel,
filter_size=1,
stride=1,
groups=1,
act="relu",
name="conv_last")
x = fluid.layers.pool2d(input=x, pool_type='avg', global_pooling=True)
output_channel = 1280
stdv = 1.0 / math.sqrt(x.shape[1] * 1.0)
out = self.conv_bn_layer(
input=x,
num_filters=output_channel,
filter_size=1,
stride=1,
groups=1,
act="relu",
name="fc_0")
out = self.conv_bn_layer(input=x,
num_filters=output_channel,
filter_size=1,
stride=1,
act="relu",
name="fc_0")
out = fluid.layers.dropout(x=out, dropout_prob=0.2)
stdv = 1.0 / math.sqrt(out.shape[1] * 1.0)
out = fluid.layers.fc(
input=out,
size=class_dim,
param_attr=ParamAttr(
name="fc_1_weight",
initializer=fluid.initializer.Uniform(-stdv, stdv)),
bias_attr=ParamAttr(name="fc_1_offset"))
return out
out = fluid.layers.fc(input=out,
size=class_dim,
param_attr=ParamAttr(name="fc_1_weights",
initializer=fluid.initializer.Uniform(-stdv, stdv)),
bias_attr=ParamAttr(name="fc_1_offset"))
return out
def _make_divisible(self, v, divisor, min_value=None):
"""
This function is taken from the original tf repo.
......@@ -119,160 +109,145 @@ class GhostNet():
groups=1,
act=None,
name=None):
x = fluid.layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) // 2,
groups=groups,
act=None,
param_attr=ParamAttr(
initializer=fluid.initializer.MSRA(), name=name + "_weights"),
bias_attr=False)
x = fluid.layers.batch_norm(
input=x,
act=act,
param_attr=ParamAttr(
name=name + "_bn_scale",
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0)),
bias_attr=ParamAttr(
name=name + "_bn_offset",
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0)),
moving_mean_name=name + "_bn_mean",
moving_variance_name=name + "_bn_variance")
x = fluid.layers.conv2d(input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) // 2,
groups=groups,
act=None,
param_attr=ParamAttr(
initializer=fluid.initializer.MSRA(), name=name + "_weights"),
bias_attr=False)
bn_name = name+"_bn"
x = fluid.layers.batch_norm(input=x,
act=act,
param_attr=ParamAttr(
name=bn_name+"_scale",
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0)),
bias_attr=ParamAttr(
name=bn_name+"_offset",
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0)),
moving_mean_name=bn_name+"_mean",
moving_variance_name=name+"_variance")
return x
def se_layer(self, input, num_channels, reduction_ratio=4, name=None):
pool = fluid.layers.pool2d(
input=input, pool_size=0, pool_type='avg', global_pooling=True)
def se_block(self, input, num_channels, reduction_ratio=4, name=None):
pool = fluid.layers.pool2d(input=input, pool_type='avg', global_pooling=True, use_cudnn=False)
stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
squeeze = fluid.layers.fc(
input=pool,
size=num_channels // reduction_ratio,
act='relu',
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + '_sqz_weights'),
bias_attr=ParamAttr(name=name + '_sqz_offset'))
squeeze = fluid.layers.fc(input=pool,
size=num_channels // reduction_ratio,
act='relu',
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + '_1_weights'),
bias_attr=ParamAttr(name=name + '_1_offset'))
stdv = 1.0 / math.sqrt(squeeze.shape[1] * 1.0)
excitation = fluid.layers.fc(
input=squeeze,
size=num_channels,
act=None,
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + '_exc_weights'),
bias_attr=ParamAttr(name=name + '_exc_offset'))
excitation = fluid.layers.clip(
x=excitation, min=0, max=1)
excitation = fluid.layers.fc(input=squeeze,
size=num_channels,
act=None,
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + '_2_weights'),
bias_attr=ParamAttr(name=name + '_2_offset'))
excitation = fluid.layers.clip(x=excitation, min=0, max=1)
se_scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0)
return se_scale
def depthwise_conv(self,
inp,
oup,
input,
output,
kernel_size,
stride=1,
relu=False,
name=None):
return self.conv_bn_layer(
input=inp,
num_filters=oup,
filter_size=kernel_size,
stride=stride,
groups=inp.shape[1],
act="relu" if relu else None,
name=name + "_dw")
return self.conv_bn_layer(input=input,
num_filters=output,
filter_size=kernel_size,
stride=stride,
groups=input.shape[1],
act="relu" if relu else None,
name=name + "_depthwise")
def ghost_module(self,
inp,
oup,
input,
output,
kernel_size=1,
ratio=2,
dw_size=3,
stride=1,
relu=True,
name=None):
self.oup = oup
init_channels = int(math.ceil(oup / ratio))
self.output = output
init_channels = int(math.ceil(output / ratio))
new_channels = int(init_channels * (ratio - 1))
primary_conv = self.conv_bn_layer(
input=inp,
num_filters=init_channels,
filter_size=kernel_size,
stride=stride,
groups=1,
act="relu" if relu else None,
name=name + "_primary_conv")
cheap_operation = self.conv_bn_layer(
input=primary_conv,
num_filters=new_channels,
filter_size=dw_size,
stride=1,
groups=init_channels,
act="relu" if relu else None,
name=name + "_cheap_operation")
out = fluid.layers.concat(
[primary_conv, cheap_operation], axis=1)
primary_conv = self.conv_bn_layer(input=input,
num_filters=init_channels,
filter_size=kernel_size,
stride=stride,
groups=1,
act="relu" if relu else None,
name=name + "_primary_conv")
cheap_operation = self.conv_bn_layer(input=primary_conv,
num_filters=new_channels,
filter_size=dw_size,
stride=1,
groups=init_channels,
act="relu" if relu else None,
name=name + "_cheap_operation")
out = fluid.layers.concat([primary_conv, cheap_operation], axis=1)
return out
def ghost_bottleneck(self,
inp,
input,
hidden_dim,
oup,
output,
kernel_size,
stride,
use_se,
name=None):
inp_channels = inp.shape[1]
x = self.ghost_module(
inp=inp,
oup=hidden_dim,
kernel_size=1,
stride=1,
relu=True,
name=name + "ghost_module_1")
inp_channels = input.shape[1]
x = self.ghost_module(input=input,
output=hidden_dim,
kernel_size=1,
stride=1,
relu=True,
name=name + "_ghost_module_1")
if stride == 2:
x = self.depthwise_conv(
inp=x,
oup=hidden_dim,
kernel_size=kernel_size,
stride=stride,
relu=False,
name=name + "_dw2")
x = self.depthwise_conv(input=x,
output=hidden_dim,
kernel_size=kernel_size,
stride=stride,
relu=False,
name=name + "_depthwise")
if use_se:
x = self.se_layer(
input=x, num_channels=hidden_dim, name=name + "se_layer")
x = self.ghost_module(
inp=x,
oup=oup,
kernel_size=1,
relu=False,
name=name + "ghost_module_2")
if stride == 1 and inp_channels == oup:
shortcut = inp
x = self.se_block(input=x, num_channels=hidden_dim, name=name+"_se")
x = self.ghost_module(input=x,
output=output,
kernel_size=1,
relu=False,
name=name + "_ghost_module_2")
if stride == 1 and inp_channels == output:
shortcut = input
else:
shortcut = self.depthwise_conv(
inp=inp,
oup=inp_channels,
kernel_size=kernel_size,
stride=stride,
relu=False,
name=name + "shortcut_depthwise_conv")
shortcut = self.conv_bn_layer(
input=shortcut,
num_filters=oup,
filter_size=1,
stride=1,
groups=1,
act=None,
name=name + "shortcut_conv_bn")
return fluid.layers.elementwise_add(
x=x, y=shortcut, axis=-1, act=None)
shortcut = self.depthwise_conv(input=input,
output=inp_channels,
kernel_size=kernel_size,
stride=stride,
relu=False,
name=name + "_shortcut_depthwise")
shortcut = self.conv_bn_layer(input=shortcut,
num_filters=output,
filter_size=1,
stride=1,
groups=1,
act=None,
name=name + "_shortcut_conv")
return fluid.layers.elementwise_add(x=x,
y=shortcut,
axis=-1)
def GhostNet_x0_5():
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或