model.load_state_dict(state_dict)报错问题

load_state_dict strict=False 使用说明

load_state_dict strict=False 时不会成功加载不对应位置的参数，这种方式无法解决模型不对应，只是在模型有改动时，只需要要加载骨干网络的参数时，用这种方法，因为骨干网络的那部分key大概率对应（还是打印检测一下保险），所以可以用strict=False，使得骨干网络那部分的参数加载上来。

print(model_path)
# model.load_state_dict(torch.load(model_path, map_location=device), strict=False)
state_dict = torch.load(model_path, map_location=device)
# print(state_dict.keys())
# print(model)
model.load_state_dict(state_dict, strict=False)

# 'module.backbone.layer1.0.weight
bl10w = state_dict["module.backbone.layer1.0.weight"]
print(bl10w.shape)
model_state_dict = model.state_dict()
mbl10w = model_state_dict["backbone.layer1.0.weight"]
print(mbl10w.shape)
print(bl10w[0][0][0][0])
print(mbl10w[0][0][0][0])

D:/python/python_data/SiamRPNPP_Change/save/SiamRPNPP_Transformer_Alexnet_epoch_39.pt
torch.Size([96, 3, 11, 11])
torch.Size([96, 3, 11, 11])
tensor(0.0473, device='cuda:0')
tensor(0.0057)

正确处理模型key不对应的方法：

参考链接：https://blog.csdn.net/xuru_0927/article/details/119274321

# load model
print(model_path)

state_dict = torch.load(model_path, map_location=device)
state_dict = {k.replace('module.', ''): v for k, v in state_dict.items()}
model.load_state_dict(state_dict)

bl10w = state_dict["backbone.layer1.0.weight"]
print(bl10w.shape)

model_state_dict = model.state_dict()
mbl10w = model_state_dict["backbone.layer1.0.weight"]
print(mbl10w.shape)

print(bl10w[0][0][0][0])
print(mbl10w[0][0][0][0])
print(bl10w[0][0][0][1])
print(mbl10w[0][0][0][1])

结果：

D:/python/python_data/SiamRPNPP_Change/save/SiamRPNPP_Transformer_Alexnet_epoch_39.pt
torch.Size([96, 3, 11, 11])
torch.Size([96, 3, 11, 11])
tensor(0.0473, device='cuda:0')
tensor(0.0473)
tensor(-0.0176, device='cuda:0')
tensor(-0.0176)