I have some parameters in my checkpoint that are unused and not needed anymore in my network. But when I call load_fromcheckpoint() this of course throws in error because it finds parameters in the checkpoint that aren’t in the model. How can I prevent this. Removing the parameters from the state_dict by fist manually loading the checkpoint with torch.load() and then removing the items from the state_dict seems to break the checkpoint, so it can’t be read by load_fromcheckpoint().
torch_checkpoint = torch.load(args.checkpoint) import re r1 = re.compile('model.up_compress_r2p_layers.*') r2 = re.compile('model.ds_compress_r2p_layers.*') for k in list(filter(r1.match, torch_checkpoint['state_dict'].keys())): del torch_checkpoint['state_dict'][k] for k in list(filter(r2.match, torch_checkpoint['state_dict'].keys())): del torch_checkpoint['state_dict'][k] model = FFB6DModule.load_from_checkpoint(torch_checkpoint)
AttributeError: 'dict' object has no attribute 'seek'. You can only torch.load from a file that is seekable. Please pre-load the data into a buffer like io.BytesIO and try to load from it instead.