Torch call model inference result is wrong analysis

The modified code available here: https://github.com/BIGBALLON/CIFAR-ZOO/blob/master/eval.py

1. Is the model imported successfully?

1.1 problem

Missing key(s) in state_dict: "features.0.weight"
Unexpected key(s) in state_dict: "module.features.0.weight"

For example, if the following command is used at this time to import the model:

net.load_state_dict(checkpoint["state_dict"], strict=False)

Here the strict parameter is set to False (unmatched network layer parameters will be ignored). How to use the multi-GPU parallel training model name when training, there will be more "module.", the model may not be imported successfully.

reference:

https://blog.csdn.net/shiwanghualuo/article/details/101756876

https://blog.csdn.net/qq_32998593/article/details/89343507

1.2 Solve

    net.load_state_dict({k.replace('module.',''):v for k,v in checkpoint["state_dict"].items()}, strict=True)

2. Input image preprocessing

Torch often uses PIL to read pictures, some preprocessing in transforms (resize, ToTensor, etc.). It has been verified before that PIL and opencv only need to be converted to the same color space, and it is basically no problem.