前言
基于faster_rcnn算法的骨干网络,resnet50_fpn_model。
目录
1.import 相关的包
import os
import torch
import torch.nn as nn
from torchvision.ops.misc import FrozenBatchNorm2d # 该FrozenBatchNorm2d 不会更行bn参数
from .feature_pyramid_network import BackboneWithFPN, LastLevelMaxPool #从FPNnet中应用函数
2、Bottleneck RESNet的三层结构
class Bottleneck(nn.Module):
expansion = 4
def __init__(self, in_channel, out_channel, stride=1, downsample=None, norm_layer=None):
super().__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel,
kernel_size=1, stride=1, bias=False) # squeeze channels
self.bn1 = norm_layer(out_channel)
# -----------------------------------------
self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel,
kernel_size=3, stride=stride, bias=False, padding=1)
self.bn2 = norm_layer(out_channel)
# -----------------------------------------
self.conv3 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel * self.expansion,
kernel_size=1, stride=1, bias=False) # unsqueeze channels
self.bn3 = norm_layer(out_channel * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
def forward(self, x):
identity = x
if self.downsample is not None:
identity = self.downsample(x)
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
out += identity
out = self.relu(out)
return out
3、ResNet
class ResNet(nn.Module):
def __init__(self, block, blocks_num, num_classes=1000, include_top=True, norm_layer=None):
super().__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.include_top = include_top
self.in_channel = 64
self.conv1 = nn.Conv2d(3, self.in_channel, kernel_size=7, stride=2,
padding=3, bias=False)
self.bn1 = norm_layer(self.in_channel)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, blocks_num[0])
self.layer2 = self._make_layer(block, 128, blocks_num[1], stride=2)
self.layer3 = self._make_layer(block, 256, blocks_num[2], stride=2)
self.layer4 = self._make_layer(block, 512, blocks_num[3], stride=2)
if self.include_top:
self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) # output size = (1, 1)
self.fc = nn.Linear(512 * block.expansion, num_classes)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
def _make_layer(self, block, channel, block_num, stride=1):
norm_layer = self._norm_layer
downsample = None
if stride != 1 or self.in_channel != channel * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.in_channel, channel * block.expansion, kernel_size=1, stride=stride, bias=False),
norm_layer(channel * block.expansion))
layers = []
layers.append(block(self.in_channel, channel, downsample=downsample,
stride=stride, norm_layer=norm_layer))
self.in_channel = channel * block.expansion
for _ in range(1, block_num):
layers.append(block(self.in_channel, channel, norm_layer=norm_layer))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
if self.include_top:
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.fc(x)
return x
4、overwrite_eps
def overwrite_eps(model, eps):
"""
This method overwrites the default eps values of all the
FrozenBatchNorm2d layers of the model with the provided value.
This is necessary to address the BC-breaking change introduced
by the bug-fix at pytorch/vision#2933. The overwrite is applied
only when the pretrained weights are loaded to maintain compatibility
with previous versions.
Args:
model (nn.Module): The model on which we perform the overwrite.
eps (float): The new value of eps.
"""
for module in model.modules():
if isinstance(module, FrozenBatchNorm2d):
module.eps = eps
5、resnet50_fpn_backbone
def resnet50_fpn_backbone(pretrain_path="",
norm_layer=FrozenBatchNorm2d, # FrozenBatchNorm2d的功能与BatchNorm2d类似,但参数无法更新
trainable_layers=3,
returned_layers=None,
extra_blocks=None):
"""
搭建resnet50_fpn——backbone
Args:
pretrain_path: resnet50的预训练权重,如果不使用就默认为空
norm_layer: 官方默认的是FrozenBatchNorm2d,即不会更新参数的bn层(因为如果batch_size设置的很小会导致效果更差,还不如不用bn层)
如果自己的GPU显存很大可以设置很大的batch_size,那么自己可以传入正常的BatchNorm2d层
(https://github.com/facebookresearch/maskrcnn-benchmark/issues/267)
trainable_layers: 指定训练哪些层结构
returned_layers: 指定哪些层的输出需要返回
extra_blocks: 在输出的特征层基础上额外添加的层结构
Returns:
"""
resnet_backbone = ResNet(Bottleneck, [3, 4, 6, 3],
include_top=False,
norm_layer=norm_layer)
if isinstance(norm_layer, FrozenBatchNorm2d):
overwrite_eps(resnet_backbone, 0.0)
if pretrain_path != "":
assert os.path.exists(pretrain_path), "{} is not exist.".format(pretrain_path)
# 载入预训练权重
print(resnet_backbone.load_state_dict(torch.load(pretrain_path), strict=False))
# select layers that wont be frozen
assert 0 <= trainable_layers <= 5
layers_to_train = ['layer4', 'layer3', 'layer2', 'layer1', 'conv1'][:trainable_layers]
# 如果要训练所有层结构的话,不要忘了conv1后还有一个bn1
if trainable_layers == 5:
layers_to_train.append("bn1")
# freeze layers
for name, parameter in resnet_backbone.named_parameters():
# 只训练不在layers_to_train列表中的层结构
if all([not name.startswith(layer) for layer in layers_to_train]):
parameter.requires_grad_(False)
if extra_blocks is None:
extra_blocks = LastLevelMaxPool()
if returned_layers is None:
returned_layers = [1, 2, 3, 4]
# 返回的特征层个数肯定大于0小于5
assert min(returned_layers) > 0 and max(returned_layers) < 5
# return_layers = {'layer1': '0', 'layer2': '1', 'layer3': '2', 'layer4': '3'}
return_layers = {f'layer{k}': str(v) for v, k in enumerate(returned_layers)}
# in_channel 为layer4的输出特征矩阵channel = 2048
in_channels_stage2 = resnet_backbone.in_channel // 8 # 256
# 记录resnet50提供给fpn的每个特征层channel
in_channels_list = [in_channels_stage2 * 2 ** (i - 1) for i in returned_layers]
# 通过fpn后得到的每个特征层的channel
out_channels = 256
return BackboneWithFPN(resnet_backbone, return_layers, in_channels_list, out_channels, extra_blocks=extra_blocks)