The loss value does not converge

[pthai1204]

Hi @AlfredXiangWu ,

Thank you for your great work. I have tried your code training on your MS-Celeb list from scratch. First of all, I used learning rate of 0.001 and I got Nan value after a few iteration . So, I tried to running with smaller learning rate of 0.0001 but the loss value did not decrease and fluctuated after 80 epoches.

Here is my configuration:

for name, value in model.named_parameters(): if 'bias' in name: if 'fc2' in name: params += [{'params':value, 'lr': 10 args.lr, 'weight_decay': 0}] else: params += [{'params':value, 'lr': 2 args.lr, 'weight_decay': 0}] else: if 'fc2' in name: params += [{'params':value, 'lr': 10 args.lr}] else: params += [{'params':value, 'lr': 1 args.lr}]

Thanks, Hai

[AlfredXiangWu]

How do you align the face images for training? The data preparation is important for training.

Besides, do you shuffle the image list when training? The train.py script sets shuffle=False for the train loader because I shuffle the image list manually.