pytorch 实现冻结部分参数训练另一部分

for p in self.parameters():
 p.requires_grad = False

class RESNET_MF(nn.Module):
 def __init__(self, model, pretrained):
  super(RESNET_MF, self).__init__()
  self.resnet = model(pretrained)
  for p in self.parameters():
   p.requires_grad = False #预训练模型加载进来后全部设置为不更新参数，然后再后面加层
  self.f = SpectralNorm(nn.Conv2d(2048, 512, 1))
  self.g = SpectralNorm(nn.Conv2d(2048, 512, 1))
  self.h = SpectralNorm(nn.Conv2d(2048, 2048, 1))
  ...

filter(lambda p: p.requires_grad, model.parameters())

optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=0.001, \
 betas=(0.9, 0.999), eps=1e-08, weight_decay=1e-5)

model_dict = torch.load('net.pth.tar').state_dict()
dict_name = list(model_dict)
for i, p in enumerate(dict_name):
 print(i, p)

0 gamma
1 resnet.conv1.weight
2 resnet.bn1.weight
3 resnet.bn1.bias
4 resnet.bn1.running_mean
5 resnet.bn1.running_var
6 resnet.layer1.0.conv1.weight
7 resnet.layer1.0.bn1.weight
8 resnet.layer1.0.bn1.bias
9 resnet.layer1.0.bn1.running_mean
....

for i,p in enumerate(net.parameters()):
 if i < 165:
  p.requires_grad = False

path="你的 .pt文件路径"
model = "你的网络"
checkpoint = torch.load(path, map_location=device)
model.load_state_dict(checkpoint)

# model_dict 是一个字典，保存网络 各层名称和参数，
model_dict = model.state_dict()
print(model_dict.keys()
# 这里打印出 网络 各层名称

checkpoint = torch.load(path,map_location=device)
for k, v in checkpoint.items():
 print("keys:".k)
# 这里打印出 预训练模型网络 各层名称, 是字典 【键】显示的另一种方式。

model.load_state_dict(checkpoint, strict=True)

# 对于 字典而言，in 或 not in 运算符都是基于 key 来判断的
model_dict = model.state_dict()
checkpoint = torch.load(path,map_location=device)
# k 是预训练模型的一个关键字， ss是 网络的有一个关键字
for k, v in checkpoint.items():
 flag = False
 for ss in model_dict.keys():
 if k in ss: # 在每一个元素内部匹配
 s = ss; flag = True; break
 else:
 continue
 if flag:
 checkpoint[k] = model_dict[s]

# 模型保存
state = {
 'epoch': epoch,
 'state_dict': model.state_dict(),
 'optimizer': optimizer.state_dict(),
  ... # 有其他希望保存的内容，也可自定义
 }
 torch.save(state, filepath)
# 加载模型，恢复训练
 model.load_state_dict(state['state_dict'])
 optimizer.load_state_dict(state['optimizer'])
 start_epoch = checkpoint['epoch'] + 1

Init_Epoch = 0
Freeze_Epoch = 50
Unfreeze_Epoch =100
#------------------------------------#
# 冻结一定部分训练
#------------------------------------# 
for param in model.backbone.parameters():
 param.requires_grad = False
for epoch in range(Init_Epoch,Freeze_Epoch): 
 # I`m Freeze-training !!
 pass
#------------------------------------#
# 解冻后训练
#------------------------------------#
for param in model.backbone.parameters():
 param.requires_grad = True
for epoch in range(Freeze_Epoch,Unfreeze_Epoch):
 # I`m unfreeze-training !!
 pass