wgcban
commented
1 year ago @money6651626 Hi, Thanks for your question. I agree with your observation - a few others also raised this issue in the past, and I would like to bring your attention that, all the models provided here did not utilize the model pre-trained on ADE dataset. Instead, we first trained the model on LEVIR dataset starting from random initialization, and then used this model as the pre-trained model when fine-tuning on other datasets such as DSIFN to speed up the training. If you are training the model on separate dataset, you can use any of the pre-trained models provided here as the starting point which will help the faster convergence. Thanks.
money6651626
Hi, I found a question that your pretrain model parameters（on ADE datasets） seem to no useful because most of the parameters do not match(the weight shape). Althought they have the same Key and load_state_dict(torch.load(self.args.pretrain), strict=False) allow it. I caculate the number of pretrain parameters which can be loaded correctly about 6 items.