[BUGs] Code that used to train OSCD

[pvmilk]

I believed that backbone from "random" and "pretrained" resnet model still have fully-connected layer in the last layer.
While the one from "pretrained" does not.

seasonal-contrast/main_oscd.py

Lines 91 to 98 in 8285173

	if args.backbone_type == 'random':
	backbone = resnet.resnet18(pretrained=False)
	elif args.backbone_type == 'imagenet':
	backbone = resnet.resnet18(pretrained=True)
	elif args.backbone_type == 'pretrain':
	model = MocoV2.load_from_checkpoint(args.ckpt_path)
	backbone = deepcopy(model.encoder_q)
	else:

seasonal-contrast//models/moco2_module.py

Lines 23 to 25 in 8285173

	# remove fc layer
	self.encoder_q = nn.Sequential(*list(self.encoder_q.children())[:-1], nn.Flatten())
	self.encoder_k = nn.Sequential(*list(self.encoder_k.children())[:-1], nn.Flatten())

I also believed that this could affect the report result in the paper, and the comparison might not be fair.