About how to change the trained model

[Lin-plax]

I have trained the vggface_48w in this project with the resnet18. And it is OK, the accruacy can be 90%.
BUT when I try to change the model, such as the resnet18 in PytorchVision. The accruacy is been to 0 ALLWAYS. EVEN 0.0125 it NOT.

Now, I have try to find the problem. I noticed something:

1. I remove the metrics.py in the train.py, then the accruacy is not 0 all the time(even it is 0.2~0.3).
2. If i only changed the num_classes from 13938 into 1333 in the cofig.py, the network will be OK, and the accruacy is appear.

I really not konw where is the problem. And I really want to know the answer.
Thanks for your reading.