Closed Tinaliu123 closed 2 months ago
I just so happened to implement that yesterday in my fork, for the same reasons (using my fine-tune with it). You can replace (temporarily rename) the original pre-trained CLIP model in .cache/clip and put your fine-tune of the same model in place of it (though you could also modify the code to use torch.load and just load the pickle / .pt from somewhere else instead):
Can I load my custom pretrained weights by providing the path to the weights when I load the mode?