使用pytorch进行迁移学习修改预训练权重模型的分类层/输出类别（附mmaction相关实现代码）

本文链接：https://blog.csdn.net/OTime77/article/details/105938268

文章目录

example
step1 文件建立
step2 获得预训练模型
step3 模型修改
参考文章

example

以mmaction中的tsn+resnet50+kinetics训练的模型作为预训练模型，拿来训练TSN+resnet+ucf101。

step1 文件建立

建立文件tools/transfer_model.py。

step2 获得预训练模型

在mmaction/modelzoo中找到TSN的kinectis模型，并下载。

step3 模型修改

因为kinetics数据集共有400类，所以最后一层分类层的权重W为(400,2048)，bias为(400)。
ucf101数据集共有101类。
不同的预训练文件可能结构不太一样？以TSN的kinectis模型作为例子，用torch.load(model_dir)加载后得到一个dict文件checkpoint，这个文件里面还有两个键，分别为meta和state_dict，meta存的是模型相关配置数据，state_dict存的是模型训练后得到的每一层的权重值。
state_dict中以name为key，param为value，param也是一个dict，param中有data这个key，里面存储的都是torch类型的数据。
所以修改思路就是把模型最后一层分类层的W数据改为(101,2048)，bias改为(101)。下面的代码是数值统一改为1。
最后将模型保存为.pth格式，直接可以用了。
关于mmaction中的--resume_from，适用于训练过程中被突然打断，--load_from，适用于迁移模型，所以这个修改后的模型用的是--load_from。

from __future__ import print_function
from __future__ import division
import torch
import copy

# model_dir = '/media/data2/ggwang/modelzoo/tsn/tsn2d_kinetics400_rgb_r50_seg3_f1s1-b702e12f.pth'
model_dir = '/path/to/your/pretrained/model'
checkpoint = torch.load(model_dir) # 加载模型，checkpoint是一个dict
model = copy.deepcopy(checkpoint) # 复制checkpoint，并修改

for name, param in model['state_dict'].items():
    if name=='cls_head.fc_cls.weight':
        param.data = torch.ones(101, 2048)
        model['state_dict'][name] = param
    if name=='cls_head.fc_cls.bias':
        param.data = torch.ones(101)
        model['state_dict'][name] = param
    print('{},size:{}'.format(name, param.data.size()))
torch.save(model, '/media/data2/ggwang/modelzoo/tsn/resnet_kinetics_ucf101.pth')

print出的为：也就是每一层的名字和参数的shape。

backbone.conv1.weight,size:torch.Size([64, 3, 7, 7])
backbone.bn1.weight,size:torch.Size([64])
backbone.bn1.bias,size:torch.Size([64])
backbone.bn1.running_mean,size:torch.Size([64])
backbone.bn1.running_var,size:torch.Size([64])
backbone.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer1.0.conv1.weight,size:torch.Size([64, 64, 1, 1])
backbone.layer1.0.conv2.weight,size:torch.Size([64, 64, 3, 3])
backbone.layer1.0.bn1.weight,size:torch.Size([64])
backbone.layer1.0.bn1.bias,size:torch.Size([64])
backbone.layer1.0.bn1.running_mean,size:torch.Size([64])
backbone.layer1.0.bn1.running_var,size:torch.Size([64])
backbone.layer1.0.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer1.0.bn2.weight,size:torch.Size([64])
backbone.layer1.0.bn2.bias,size:torch.Size([64])
backbone.layer1.0.bn2.running_mean,size:torch.Size([64])
backbone.layer1.0.bn2.running_var,size:torch.Size([64])
backbone.layer1.0.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer1.0.conv3.weight,size:torch.Size([256, 64, 1, 1])
backbone.layer1.0.bn3.weight,size:torch.Size([256])
backbone.layer1.0.bn3.bias,size:torch.Size([256])
backbone.layer1.0.bn3.running_mean,size:torch.Size([256])
backbone.layer1.0.bn3.running_var,size:torch.Size([256])
backbone.layer1.0.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer1.0.downsample.0.weight,size:torch.Size([256, 64, 1, 1])
backbone.layer1.0.downsample.1.weight,size:torch.Size([256])
backbone.layer1.0.downsample.1.bias,size:torch.Size([256])
backbone.layer1.0.downsample.1.running_mean,size:torch.Size([256])
backbone.layer1.0.downsample.1.running_var,size:torch.Size([256])
backbone.layer1.0.downsample.1.num_batches_tracked,size:torch.Size([])
backbone.layer1.1.conv1.weight,size:torch.Size([64, 256, 1, 1])
backbone.layer1.1.conv2.weight,size:torch.Size([64, 64, 3, 3])
backbone.layer1.1.bn1.weight,size:torch.Size([64])
backbone.layer1.1.bn1.bias,size:torch.Size([64])
backbone.layer1.1.bn1.running_mean,size:torch.Size([64])
backbone.layer1.1.bn1.running_var,size:torch.Size([64])
backbone.layer1.1.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer1.1.bn2.weight,size:torch.Size([64])
backbone.layer1.1.bn2.bias,size:torch.Size([64])
backbone.layer1.1.bn2.running_mean,size:torch.Size([64])
backbone.layer1.1.bn2.running_var,size:torch.Size([64])
backbone.layer1.1.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer1.1.conv3.weight,size:torch.Size([256, 64, 1, 1])
backbone.layer1.1.bn3.weight,size:torch.Size([256])
backbone.layer1.1.bn3.bias,size:torch.Size([256])
backbone.layer1.1.bn3.running_mean,size:torch.Size([256])
backbone.layer1.1.bn3.running_var,size:torch.Size([256])
backbone.layer1.1.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer1.2.conv1.weight,size:torch.Size([64, 256, 1, 1])
backbone.layer1.2.conv2.weight,size:torch.Size([64, 64, 3, 3])
backbone.layer1.2.bn1.weight,size:torch.Size([64])
backbone.layer1.2.bn1.bias,size:torch.Size([64])
backbone.layer1.2.bn1.running_mean,size:torch.Size([64])
backbone.layer1.2.bn1.running_var,size:torch.Size([64])
backbone.layer1.2.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer1.2.bn2.weight,size:torch.Size([64])
backbone.layer1.2.bn2.bias,size:torch.Size([64])
backbone.layer1.2.bn2.running_mean,size:torch.Size([64])
backbone.layer1.2.bn2.running_var,size:torch.Size([64])
backbone.layer1.2.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer1.2.conv3.weight,size:torch.Size([256, 64, 1, 1])
backbone.layer1.2.bn3.weight,size:torch.Size([256])
backbone.layer1.2.bn3.bias,size:torch.Size([256])
backbone.layer1.2.bn3.running_mean,size:torch.Size([256])
backbone.layer1.2.bn3.running_var,size:torch.Size([256])
backbone.layer1.2.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer2.0.conv1.weight,size:torch.Size([128, 256, 1, 1])
backbone.layer2.0.conv2.weight,size:torch.Size([128, 128, 3, 3])
backbone.layer2.0.bn1.weight,size:torch.Size([128])
backbone.layer2.0.bn1.bias,size:torch.Size([128])
backbone.layer2.0.bn1.running_mean,size:torch.Size([128])
backbone.layer2.0.bn1.running_var,size:torch.Size([128])
backbone.layer2.0.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer2.0.bn2.weight,size:torch.Size([128])
backbone.layer2.0.bn2.bias,size:torch.Size([128])
backbone.layer2.0.bn2.running_mean,size:torch.Size([128])
backbone.layer2.0.bn2.running_var,size:torch.Size([128])
backbone.layer2.0.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer2.0.conv3.weight,size:torch.Size([512, 128, 1, 1])
backbone.layer2.0.bn3.weight,size:torch.Size([512])
backbone.layer2.0.bn3.bias,size:torch.Size([512])
backbone.layer2.0.bn3.running_mean,size:torch.Size([512])
backbone.layer2.0.bn3.running_var,size:torch.Size([512])
backbone.layer2.0.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer2.0.downsample.0.weight,size:torch.Size([512, 256, 1, 1])
backbone.layer2.0.downsample.1.weight,size:torch.Size([512])
backbone.layer2.0.downsample.1.bias,size:torch.Size([512])
backbone.layer2.0.downsample.1.running_mean,size:torch.Size([512])
backbone.layer2.0.downsample.1.running_var,size:torch.Size([512])
backbone.layer2.0.downsample.1.num_batches_tracked,size:torch.Size([])
backbone.layer2.1.conv1.weight,size:torch.Size([128, 512, 1, 1])
backbone.layer2.1.conv2.weight,size:torch.Size([128, 128, 3, 3])
backbone.layer2.1.bn1.weight,size:torch.Size([128])
backbone.layer2.1.bn1.bias,size:torch.Size([128])
backbone.layer2.1.bn1.running_mean,size:torch.Size([128])
backbone.layer2.1.bn1.running_var,size:torch.Size([128])
backbone.layer2.1.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer2.1.bn2.weight,size:torch.Size([128])
backbone.layer2.1.bn2.bias,size:torch.Size([128])
backbone.layer2.1.bn2.running_mean,size:torch.Size([128])
backbone.layer2.1.bn2.running_var,size:torch.Size([128])
backbone.layer2.1.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer2.1.conv3.weight,size:torch.Size([512, 128, 1, 1])
backbone.layer2.1.bn3.weight,size:torch.Size([512])
backbone.layer2.1.bn3.bias,size:torch.Size([512])
backbone.layer2.1.bn3.running_mean,size:torch.Size([512])
backbone.layer2.1.bn3.running_var,size:torch.Size([512])
backbone.layer2.1.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer2.2.conv1.weight,size:torch.Size([128, 512, 1, 1])
backbone.layer2.2.conv2.weight,size:torch.Size([128, 128, 3, 3])
backbone.layer2.2.bn1.weight,size:torch.Size([128])
backbone.layer2.2.bn1.bias,size:torch.Size([128])
backbone.layer2.2.bn1.running_mean,size:torch.Size([128])
backbone.layer2.2.bn1.running_var,size:torch.Size([128])
backbone.layer2.2.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer2.2.bn2.weight,size:torch.Size([128])
backbone.layer2.2.bn2.bias,size:torch.Size([128])
backbone.layer2.2.bn2.running_mean,size:torch.Size([128])
backbone.layer2.2.bn2.running_var,size:torch.Size([128])
backbone.layer2.2.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer2.2.conv3.weight,size:torch.Size([512, 128, 1, 1])
backbone.layer2.2.bn3.weight,size:torch.Size([512])
backbone.layer2.2.bn3.bias,size:torch.Size([512])
backbone.layer2.2.bn3.running_mean,size:torch.Size([512])
backbone.layer2.2.bn3.running_var,size:torch.Size([512])
backbone.layer2.2.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer2.3.conv1.weight,size:torch.Size([128, 512, 1, 1])
backbone.layer2.3.conv2.weight,size:torch.Size([128, 128, 3, 3])
backbone.layer2.3.bn1.weight,size:torch.Size([128])
backbone.layer2.3.bn1.bias,size:torch.Size([128])
backbone.layer2.3.bn1.running_mean,size:torch.Size([128])
backbone.layer2.3.bn1.running_var,size:torch.Size([128])
backbone.layer2.3.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer2.3.bn2.weight,size:torch.Size([128])
backbone.layer2.3.bn2.bias,size:torch.Size([128])
backbone.layer2.3.bn2.running_mean,size:torch.Size([128])
backbone.layer2.3.bn2.running_var,size:torch.Size([128])
backbone.layer2.3.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer2.3.conv3.weight,size:torch.Size([512, 128, 1, 1])
backbone.layer2.3.bn3.weight,size:torch.Size([512])
backbone.layer2.3.bn3.bias,size:torch.Size([512])
backbone.layer2.3.bn3.running_mean,size:torch.Size([512])
backbone.layer2.3.bn3.running_var,size:torch.Size([512])
backbone.layer2.3.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer3.0.conv1.weight,size:torch.Size([256, 512, 1, 1])
backbone.layer3.0.conv2.weight,size:torch.Size([256, 256, 3, 3])
backbone.layer3.0.bn1.weight,size:torch.Size([256])
backbone.layer3.0.bn1.bias,size:torch.Size([256])
backbone.layer3.0.bn1.running_mean,size:torch.Size([256])
backbone.layer3.0.bn1.running_var,size:torch.Size([256])
backbone.layer3.0.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer3.0.bn2.weight,size:torch.Size([256])
backbone.layer3.0.bn2.bias,size:torch.Size([256])
backbone.layer3.0.bn2.running_mean,size:torch.Size([256])
backbone.layer3.0.bn2.running_var,size:torch.Size([256])
backbone.layer3.0.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer3.0.conv3.weight,size:torch.Size([1024, 256, 1, 1])
backbone.layer3.0.bn3.weight,size:torch.Size([1024])
backbone.layer3.0.bn3.bias,size:torch.Size([1024])
backbone.layer3.0.bn3.running_mean,size:torch.Size([1024])
backbone.layer3.0.bn3.running_var,size:torch.Size([1024])
backbone.layer3.0.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer3.0.downsample.0.weight,size:torch.Size([1024, 512, 1, 1])
backbone.layer3.0.downsample.1.weight,size:torch.Size([1024])
backbone.layer3.0.downsample.1.bias,size:torch.Size([1024])
backbone.layer3.0.downsample.1.running_mean,size:torch.Size([1024])
backbone.layer3.0.downsample.1.running_var,size:torch.Size([1024])
backbone.layer3.0.downsample.1.num_batches_tracked,size:torch.Size([])
backbone.layer3.1.conv1.weight,size:torch.Size([256, 1024, 1, 1])
backbone.layer3.1.conv2.weight,size:torch.Size([256, 256, 3, 3])
backbone.layer3.1.bn1.weight,size:torch.Size([256])
backbone.layer3.1.bn1.bias,size:torch.Size([256])
backbone.layer3.1.bn1.running_mean,size:torch.Size([256])
backbone.layer3.1.bn1.running_var,size:torch.Size([256])
backbone.layer3.1.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer3.1.bn2.weight,size:torch.Size([256])
backbone.layer3.1.bn2.bias,size:torch.Size([256])
backbone.layer3.1.bn2.running_mean,size:torch.Size([256])
backbone.layer3.1.bn2.running_var,size:torch.Size([256])
backbone.layer3.1.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer3.1.conv3.weight,size:torch.Size([1024, 256, 1, 1])
backbone.layer3.1.bn3.weight,size:torch.Size([1024])
backbone.layer3.1.bn3.bias,size:torch.Size([1024])
backbone.layer3.1.bn3.running_mean,size:torch.Size([1024])
backbone.layer3.1.bn3.running_var,size:torch.Size([1024])
backbone.layer3.1.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer3.2.conv1.weight,size:torch.Size([256, 1024, 1, 1])
backbone.layer3.2.conv2.weight,size:torch.Size([256, 256, 3, 3])
backbone.layer3.2.bn1.weight,size:torch.Size([256])
backbone.layer3.2.bn1.bias,size:torch.Size([256])
backbone.layer3.2.bn1.running_mean,size:torch.Size([256])
backbone.layer3.2.bn1.running_var,size:torch.Size([256])
backbone.layer3.2.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer3.2.bn2.weight,size:torch.Size([256])
backbone.layer3.2.bn2.bias,size:torch.Size([256])
backbone.layer3.2.bn2.running_mean,size:torch.Size([256])
backbone.layer3.2.bn2.running_var,size:torch.Size([256])
backbone.layer3.2.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer3.2.conv3.weight,size:torch.Size([1024, 256, 1, 1])
backbone.layer3.2.bn3.weight,size:torch.Size([1024])
backbone.layer3.2.bn3.bias,size:torch.Size([1024])
backbone.layer3.2.bn3.running_mean,size:torch.Size([1024])
backbone.layer3.2.bn3.running_var,size:torch.Size([1024])
backbone.layer3.2.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer3.3.conv1.weight,size:torch.Size([256, 1024, 1, 1])
backbone.layer3.3.conv2.weight,size:torch.Size([256, 256, 3, 3])
backbone.layer3.3.bn1.weight,size:torch.Size([256])
backbone.layer3.3.bn1.bias,size:torch.Size([256])
backbone.layer3.3.bn1.running_mean,size:torch.Size([256])
backbone.layer3.3.bn1.running_var,size:torch.Size([256])
backbone.layer3.3.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer3.3.bn2.weight,size:torch.Size([256])
backbone.layer3.3.bn2.bias,size:torch.Size([256])
backbone.layer3.3.bn2.running_mean,size:torch.Size([256])
backbone.layer3.3.bn2.running_var,size:torch.Size([256])
backbone.layer3.3.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer3.3.conv3.weight,size:torch.Size([1024, 256, 1, 1])
backbone.layer3.3.bn3.weight,size:torch.Size([1024])
backbone.layer3.3.bn3.bias,size:torch.Size([1024])
backbone.layer3.3.bn3.running_mean,size:torch.Size([1024])
backbone.layer3.3.bn3.running_var,size:torch.Size([1024])
backbone.layer3.3.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer3.4.conv1.weight,size:torch.Size([256, 1024, 1, 1])
backbone.layer3.4.conv2.weight,size:torch.Size([256, 256, 3, 3])
backbone.layer3.4.bn1.weight,size:torch.Size([256])
backbone.layer3.4.bn1.bias,size:torch.Size([256])
backbone.layer3.4.bn1.running_mean,size:torch.Size([256])
backbone.layer3.4.bn1.running_var,size:torch.Size([256])
backbone.layer3.4.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer3.4.bn2.weight,size:torch.Size([256])
backbone.layer3.4.bn2.bias,size:torch.Size([256])
backbone.layer3.4.bn2.running_mean,size:torch.Size([256])
backbone.layer3.4.bn2.running_var,size:torch.Size([256])
backbone.layer3.4.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer3.4.conv3.weight,size:torch.Size([1024, 256, 1, 1])
backbone.layer3.4.bn3.weight,size:torch.Size([1024])
backbone.layer3.4.bn3.bias,size:torch.Size([1024])
backbone.layer3.4.bn3.running_mean,size:torch.Size([1024])
backbone.layer3.4.bn3.running_var,size:torch.Size([1024])
backbone.layer3.4.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer3.5.conv1.weight,size:torch.Size([256, 1024, 1, 1])
backbone.layer3.5.conv2.weight,size:torch.Size([256, 256, 3, 3])
backbone.layer3.5.bn1.weight,size:torch.Size([256])
backbone.layer3.5.bn1.bias,size:torch.Size([256])
backbone.layer3.5.bn1.running_mean,size:torch.Size([256])
backbone.layer3.5.bn1.running_var,size:torch.Size([256])
backbone.layer3.5.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer3.5.bn2.weight,size:torch.Size([256])
backbone.layer3.5.bn2.bias,size:torch.Size([256])
backbone.layer3.5.bn2.running_mean,size:torch.Size([256])
backbone.layer3.5.bn2.running_var,size:torch.Size([256])
backbone.layer3.5.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer3.5.conv3.weight,size:torch.Size([1024, 256, 1, 1])
backbone.layer3.5.bn3.weight,size:torch.Size([1024])
backbone.layer3.5.bn3.bias,size:torch.Size([1024])
backbone.layer3.5.bn3.running_mean,size:torch.Size([1024])
backbone.layer3.5.bn3.running_var,size:torch.Size([1024])
backbone.layer3.5.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer4.0.conv1.weight,size:torch.Size([512, 1024, 1, 1])
backbone.layer4.0.conv2.weight,size:torch.Size([512, 512, 3, 3])
backbone.layer4.0.bn1.weight,size:torch.Size([512])
backbone.layer4.0.bn1.bias,size:torch.Size([512])
backbone.layer4.0.bn1.running_mean,size:torch.Size([512])
backbone.layer4.0.bn1.running_var,size:torch.Size([512])
backbone.layer4.0.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer4.0.bn2.weight,size:torch.Size([512])
backbone.layer4.0.bn2.bias,size:torch.Size([512])
backbone.layer4.0.bn2.running_mean,size:torch.Size([512])
backbone.layer4.0.bn2.running_var,size:torch.Size([512])
backbone.layer4.0.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer4.0.conv3.weight,size:torch.Size([2048, 512, 1, 1])
backbone.layer4.0.bn3.weight,size:torch.Size([2048])
backbone.layer4.0.bn3.bias,size:torch.Size([2048])
backbone.layer4.0.bn3.running_mean,size:torch.Size([2048])
backbone.layer4.0.bn3.running_var,size:torch.Size([2048])
backbone.layer4.0.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer4.0.downsample.0.weight,size:torch.Size([2048, 1024, 1, 1])
backbone.layer4.0.downsample.1.weight,size:torch.Size([2048])
backbone.layer4.0.downsample.1.bias,size:torch.Size([2048])
backbone.layer4.0.downsample.1.running_mean,size:torch.Size([2048])
backbone.layer4.0.downsample.1.running_var,size:torch.Size([2048])
backbone.layer4.0.downsample.1.num_batches_tracked,size:torch.Size([])
backbone.layer4.1.conv1.weight,size:torch.Size([512, 2048, 1, 1])
backbone.layer4.1.conv2.weight,size:torch.Size([512, 512, 3, 3])
backbone.layer4.1.bn1.weight,size:torch.Size([512])
backbone.layer4.1.bn1.bias,size:torch.Size([512])
backbone.layer4.1.bn1.running_mean,size:torch.Size([512])
backbone.layer4.1.bn1.running_var,size:torch.Size([512])
backbone.layer4.1.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer4.1.bn2.weight,size:torch.Size([512])
backbone.layer4.1.bn2.bias,size:torch.Size([512])
backbone.layer4.1.bn2.running_mean,size:torch.Size([512])
backbone.layer4.1.bn2.running_var,size:torch.Size([512])
backbone.layer4.1.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer4.1.conv3.weight,size:torch.Size([2048, 512, 1, 1])
backbone.layer4.1.bn3.weight,size:torch.Size([2048])
backbone.layer4.1.bn3.bias,size:torch.Size([2048])
backbone.layer4.1.bn3.running_mean,size:torch.Size([2048])
backbone.layer4.1.bn3.running_var,size:torch.Size([2048])
backbone.layer4.1.bn3.num_batches_tracked,size:torch.Size([])
backbone.layer4.2.conv1.weight,size:torch.Size([512, 2048, 1, 1])
backbone.layer4.2.conv2.weight,size:torch.Size([512, 512, 3, 3])
backbone.layer4.2.bn1.weight,size:torch.Size([512])
backbone.layer4.2.bn1.bias,size:torch.Size([512])
backbone.layer4.2.bn1.running_mean,size:torch.Size([512])
backbone.layer4.2.bn1.running_var,size:torch.Size([512])
backbone.layer4.2.bn1.num_batches_tracked,size:torch.Size([])
backbone.layer4.2.bn2.weight,size:torch.Size([512])
backbone.layer4.2.bn2.bias,size:torch.Size([512])
backbone.layer4.2.bn2.running_mean,size:torch.Size([512])
backbone.layer4.2.bn2.running_var,size:torch.Size([512])
backbone.layer4.2.bn2.num_batches_tracked,size:torch.Size([])
backbone.layer4.2.conv3.weight,size:torch.Size([2048, 512, 1, 1])
backbone.layer4.2.bn3.weight,size:torch.Size([2048])
backbone.layer4.2.bn3.bias,size:torch.Size([2048])
backbone.layer4.2.bn3.running_mean,size:torch.Size([2048])
backbone.layer4.2.bn3.running_var,size:torch.Size([2048])
backbone.layer4.2.bn3.num_batches_tracked,size:torch.Size([])
cls_head.fc_cls.weight,size:torch.Size([101, 2048])
cls_head.fc_cls.bias,size:torch.Size([101])