No difference between resnet50 pretrained and not

vision
1

I’m training a classification model using resnet50 with pretrained and not pretrained one. However i don’t see much differences between them. Can anyone please explain to me?
Thank you so much.
My code:

dict_model = {
        'r18': models.resnet18(pretrained=args.pretrained),
        'r34': models.resnet34(pretrained=args.pretrained),
        'r50': models.resnet50(pretrained=args.pretrained),
        'r101': models.resnet101(pretrained=args.pretrained),
        'r152': models.resnet152(pretrained=args.pretrained)        
    }
    model_ft = dict_model[args.model]

Here is my result:

Screenshot 2022-03-22 215728
Screenshot 2022-03-22 215728889×505 48 KB

2

I’m not sure I understand the issue as the accuracy plot seems to show a better accuracy for the pretrained models? E.g. it seems you are reaching ~0.5 at step 1-2 using the pretrained models while training from scratch needs 5 steps or am I mixing up the plots?

1 Like