Github-pytorch-ssd
vision/ssd
mobilenet_v2
image.png
其中：t表示“扩张”倍数，c表示输出通道数，n表示重复次数，s表示步长stride。
先说两点有误之处吧：
第五行，也就是第7~10个bottleneck，stride=2，分辨率应该从28降低到14；如果不是分辨率出错，那就应该是stride=1；
文中提到共计采用19个bottleneck，但是这里只有17个。
import torch.nn as nn
import math
# Modified from https://github.com/tonylins/pytorch-mobilenet-v2/blob/master/MobileNetV2.py.
# In this version, Relu6 is replaced with Relu to make it ONNX compatible.
# BatchNorm Layer is optional to make it easy do batch norm confusion.
深度可分离卷积用pytorch实现比较简单，conv2d中有个参数叫groups，将它设置为通道数即可
# 普通3*3卷积
def conv_bn(inp, oup, stride, use_batch_norm=True, onnx_compatible=False):
# onnx_compatible决定是使用普通relu还是relu6
'''
普通relu， y=max(0, x), 相当于无限多个bernoulli分布，即无限多个骰子
relu6, y= min(max(0,x), 6), 相当于有六个bernoulli分布，即6个硬币，同时抛出正面，这样鼓励网络学习到稀疏特征。
网络里面每一个输出n，相当于n个bernoulli分布的叠加。
通过实验发现，用6，效果比较好。所以选用了6
'''
ReLU = nn.ReLU if onnx_compatible else nn.ReLU6
if use_batch_norm:
return nn.Sequential(
nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
nn.BatchNorm2d(oup),
ReLU(inplace=True)
)
else:
return nn.Sequential(
nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
ReLU(inplace=True)
)
上面的代码中频繁出现了inplace=True，在此作出说明。
inplace=True：不创建新的对象，直接对原始对象进行修改；
inplace=False：对数据进行修改，创建并返回新的对象承载其修改结果。
这里relu指定inplace=True，则对于上层网络传递下来的tensor直接进行修改，可以少存储变量，节省运算内存。
# 点卷积
def conv_1x1_bn(inp, oup, use_batch_norm=True, onnx_compatible=False):
ReLU = nn.ReLU if onnx_compatible else nn.ReLU6
if use_batch_norm:
return nn.Sequential(
nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
nn.BatchNorm2d(oup),
ReLU(inplace=True)
)
else:
return nn.Sequential(
nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
ReLU(inplace=True)
)
# mobilenetv2的创新点，InvertedResidual
class InvertedResidual(nn.Module):
def __init__(self, inp, oup, stride, expand_ratio, use_batch_norm=True, onnx_compatible=False):
super(InvertedResidual, self).__init__()
ReLU = nn.ReLU if onnx_compatible else nn.ReLU6
self.stride = stride
assert stride in [1, 2]
hidden_dim = round(inp * expand_ratio)
self.use_res_connect = self.stride == 1 and inp == oup
if expand_ratio == 1:
# 当我不扩张时，hidden_dim就等于inp, InvertedResidual结构是: 3*3DWconv, (BN,) Relu, 1*1PWconv, (BN)
if use_batch_norm:
self.conv = nn.Sequential(
# dw(深度可分离卷积)
# torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias=True)
nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
nn.BatchNorm2d(hidden_dim),
ReLU(inplace=True),
# pw-linear 点卷积，实现线性变换
# 点卷积不不需要padding
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
nn.BatchNorm2d(oup),
)
else:
self.conv = nn.Sequential(
# dw
nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
ReLU(inplace=True),
# pw-linear
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
)
else:
# 当我需要扩张时，InvertedResidual结构是: 1*1PWconv, (BN,) Relu, 3*3DWconv, (BN,) Relu, 1*1PWconv, (BN)
if use_batch_norm:
self.conv = nn.Sequential(
# pw(点卷积扩张)
nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
nn.BatchNorm2d(hidden_dim),
ReLU(inplace=True),
# dw(深度可分离卷积)
nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
nn.BatchNorm2d(hidden_dim),
ReLU(inplace=True),
# pw-linear(点卷积压缩)
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
nn.BatchNorm2d(oup),
)
else:
self.conv = nn.Sequential(
# pw
nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
ReLU(inplace=True),
# dw
nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
ReLU(inplace=True),
# pw-linear
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
)
def forward(self, x):
# 残差网络是执行+操作
if self.use_res_connect:
return x + self.conv(x)
else:
return self.conv(x)
这里来梳理几个知识点：
concat与add的区别
shortcut residual是基于layer的concat操作
router是add操作
分支操作是基于通道的concat操作
class MobileNetV2(nn.Module):
def __init__(self, n_class=1000, input_size=224, width_mult=1., dropout_ratio=0.2,
use_batch_norm=True, onnx_compatible=False):
super(MobileNetV2, self).__init__()
block = InvertedResidual
input_channel = 32
last_channel = 1280
# t表示“扩张”倍数，c表示输出通道数，n表示重复次数，s表示步长stride。
interverted_residual_setting = [
# t, c, n, s
[1, 16, 1, 1],
[6, 24, 2, 2],
[6, 32, 3, 2],
[6, 64, 4, 2],
[6, 96, 3, 1],
[6, 160, 3, 2],
[6, 320, 1, 1],
]
# building first layer
assert input_size % 32 == 0
input_channel = int(input_channel * width_mult)
self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last_channel
self.features = [conv_bn(3, input_channel, 2, onnx_compatible=onnx_compatible)]
# building inverted residual blocks
for t, c, n, s in interverted_residual_setting:
output_channel = int(c * width_mult)
for i in range(n):
if i == 0:
self.features.append(block(input_channel, output_channel, s,
expand_ratio=t, use_batch_norm=use_batch_norm,
onnx_compatible=onnx_compatible))
else:
self.features.append(block(input_channel, output_channel, 1,
expand_ratio=t, use_batch_norm=use_batch_norm,
onnx_compatible=onnx_compatible))
input_channel = output_channel
# building last several layers
self.features.append(conv_1x1_bn(input_channel, self.last_channel,
use_batch_norm=use_batch_norm, onnx_compatible=onnx_compatible))
# make it nn.Sequential
self.features = nn.Sequential(*self.features)
# building classifier
self.classifier = nn.Sequential(
nn.Dropout(dropout_ratio),
nn.Linear(self.last_channel, n_class),
)
self._initialize_weights()
def forward(self, x):
x = self.features(x)
x = x.mean(3).mean(2)
x = self.classifier(x)
return x
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
if m.bias is not None:
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
n = m.weight.size(1)
m.weight.data.normal_(0, 0.01)
m.bias.data.zero_()
