Spaces:
Configuration error
Configuration error
File size: 11,105 Bytes
1ab1a09 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 |
# copyright (c) 2022 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
#
# 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.
# Code was based on https://github.com/huawei-noah/CV-Backbones/tree/master/ghostnet_pytorch
import math
import paddle
from paddle import ParamAttr
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.nn import Conv2D, BatchNorm, AdaptiveAvgPool2D, Linear
from paddle.regularizer import L2Decay
from paddle.nn.initializer import Uniform, KaimingNormal
from paddleseg.cvlibs import manager
from paddleseg.utils import utils, logger
__all__ = ["GhostNet_x0_5", "GhostNet_x1_0", "GhostNet_x1_3"]
class ConvBNLayer(nn.Layer):
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,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=(kernel_size - 1) // 2,
groups=groups,
weight_attr=ParamAttr(
initializer=KaimingNormal(), name=name + "_weights"),
bias_attr=False)
bn_name = name + "_bn"
self._batch_norm = BatchNorm(
num_channels=out_channels,
act=act,
param_attr=ParamAttr(
name=bn_name + "_scale", regularizer=L2Decay(0.0)),
bias_attr=ParamAttr(
name=bn_name + "_offset", regularizer=L2Decay(0.0)),
moving_mean_name=bn_name + "_mean",
moving_variance_name=bn_name + "_variance")
def forward(self, inputs):
y = self._conv(inputs)
y = self._batch_norm(y)
return y
class SEBlock(nn.Layer):
def __init__(self, num_channels, reduction_ratio=4, name=None):
super(SEBlock, self).__init__()
self.pool2d_gap = AdaptiveAvgPool2D(1)
self._num_channels = num_channels
stdv = 1.0 / math.sqrt(num_channels * 1.0)
med_ch = num_channels // reduction_ratio
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"))
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"))
def forward(self, inputs):
pool = self.pool2d_gap(inputs)
pool = paddle.squeeze(pool, axis=[2, 3])
squeeze = self.squeeze(pool)
squeeze = F.relu(squeeze)
excitation = self.excitation(squeeze)
excitation = paddle.clip(x=excitation, min=0, max=1)
excitation = paddle.unsqueeze(excitation, axis=[2, 3])
out = paddle.multiply(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):
super(GhostModule, self).__init__()
init_channels = int(math.ceil(output_channels / ratio))
new_channels = int(init_channels * (ratio - 1))
self.primary_conv = ConvBNLayer(
in_channels=in_channels,
out_channels=init_channels,
kernel_size=kernel_size,
stride=stride,
groups=1,
act="relu" if relu else None,
name=name + "_primary_conv")
self.cheap_operation = ConvBNLayer(
in_channels=init_channels,
out_channels=new_channels,
kernel_size=dw_size,
stride=1,
groups=init_channels,
act="relu" if relu else None,
name=name + "_cheap_operation")
def forward(self, inputs):
x = self.primary_conv(inputs)
y = self.cheap_operation(x)
out = paddle.concat([x, y], axis=1)
return out
class GhostBottleneck(nn.Layer):
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
self._num_channels = in_channels
self._output_channels = output_channels
self.ghost_module_1 = GhostModule(
in_channels=in_channels,
output_channels=hidden_dim,
kernel_size=1,
stride=1,
relu=True,
name=name + "_ghost_module_1")
if stride == 2:
self.depthwise_conv = ConvBNLayer(
in_channels=hidden_dim,
out_channels=hidden_dim,
kernel_size=kernel_size,
stride=stride,
groups=hidden_dim,
act=None,
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.ghost_module_2 = GhostModule(
in_channels=hidden_dim,
output_channels=output_channels,
kernel_size=1,
relu=False,
name=name + "_ghost_module_2")
if stride != 1 or in_channels != output_channels:
self.shortcut_depthwise = ConvBNLayer(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=kernel_size,
stride=stride,
groups=in_channels,
act=None,
name=name +
"_shortcut_depthwise_depthwise" # looks strange due to an old typo, will be fixed later.
)
self.shortcut_conv = ConvBNLayer(
in_channels=in_channels,
out_channels=output_channels,
kernel_size=1,
stride=1,
groups=1,
act=None,
name=name + "_shortcut_conv")
def forward(self, inputs):
x = self.ghost_module_1(inputs)
if self._stride == 2:
x = self.depthwise_conv(x)
if self._use_se:
x = self.se_block(x)
x = self.ghost_module_2(x)
if self._stride == 1 and self._num_channels == self._output_channels:
shortcut = inputs
else:
shortcut = self.shortcut_depthwise(inputs)
shortcut = self.shortcut_conv(shortcut)
return paddle.add(x=x, y=shortcut)
class GhostNet(nn.Layer):
def __init__(self, scale, pretrained=None):
super(GhostNet, self).__init__()
self.cfgs = [
# k, t, c, SE, s
[3, 16, 16, 0, 1],
[3, 48, 24, 0, 2],
[3, 72, 24, 0, 1], # x4
[5, 72, 40, 1, 2],
[5, 120, 40, 1, 1], # x8
[3, 240, 80, 0, 2],
[3, 200, 80, 0, 1],
[3, 184, 80, 0, 1],
[3, 184, 80, 0, 1],
[3, 480, 112, 1, 1],
[3, 672, 112, 1, 1], # x16
[5, 672, 160, 1, 2],
[5, 960, 160, 0, 1],
[5, 960, 160, 1, 1],
[5, 960, 160, 0, 1],
[5, 960, 160, 1, 1] # x32
]
self.scale = scale
self.pretrained = pretrained
output_channels = int(self._make_divisible(16 * self.scale, 4))
self.conv1 = ConvBNLayer(
in_channels=3,
out_channels=output_channels,
kernel_size=3,
stride=2,
groups=1,
act="relu",
name="conv1")
# build inverted residual blocks
self.out_index = [2, 4, 10, 15]
self.feat_channels = []
self.ghost_bottleneck_list = []
for idx, (k, exp_size, c, use_se, s) in enumerate(self.cfgs):
in_channels = output_channels
output_channels = int(self._make_divisible(c * self.scale, 4))
hidden_dim = int(self._make_divisible(exp_size * self.scale, 4))
ghost_bottleneck = self.add_sublayer(
name="_ghostbottleneck_" + str(idx),
sublayer=GhostBottleneck(
in_channels=in_channels,
hidden_dim=hidden_dim,
output_channels=output_channels,
kernel_size=k,
stride=s,
use_se=use_se,
name="_ghostbottleneck_" + str(idx)))
self.ghost_bottleneck_list.append(ghost_bottleneck)
if idx in self.out_index:
self.feat_channels.append(output_channels)
self.init_weight()
def init_weight(self):
if self.pretrained is not None:
utils.load_entire_model(self, self.pretrained)
def forward(self, inputs):
feat_list = []
x = self.conv1(inputs)
for idx, ghost_bottleneck in enumerate(self.ghost_bottleneck_list):
x = ghost_bottleneck(x)
if idx in self.out_index:
feat_list.append(x)
return feat_list
def _make_divisible(self, v, divisor, min_value=None):
"""
This function is taken from the original tf repo.
It ensures that all layers have a channel number that is divisible by 8
It can be seen here:
https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
"""
if min_value is None:
min_value = divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_v < 0.9 * v:
new_v += divisor
return new_v
@manager.BACKBONES.add_component
def GhostNet_x0_5(**kwargs):
model = GhostNet(scale=0.5, **kwargs)
return model
@manager.BACKBONES.add_component
def GhostNet_x1_0(**kwargs):
model = GhostNet(scale=1.0, **kwargs)
return model
@manager.BACKBONES.add_component
def GhostNet_x1_3(**kwargs):
model = GhostNet(scale=1.3, **kwargs)
return model
|