Unlearning algorithm check

[YuyangXueEd]

Hello, thank you for the great repository.

I am going through the code and start from mu_saliency_unlearn_salun, and I found that in generate_mask.py file, from line 56:

forget_batch = {
                    "edited": images,
                    "edit": {"c_crossattn": prompts}
                }

null_batch = {
                    "edited": images,
                    "edit": {"c_crossattn": prompts}
                }

and I double check the original repo of Salun, in line 44, it was:

forget_batch = {"jpg": images.permute(0, 2, 3, 1), "txt": prompts}

null_batch = {"jpg": images.permute(0, 2, 3, 1), "txt": null_prompts}

I wonder if there is any misunderstanding here.

Looking forward to your reply. Thank you very much.

Best

[NormalUhr]

Hi, thanks for your interest in our work! The difference is made in order to cutomize the code/model used from SalUn to the one in this repo, which was inherited from the InstructPix2Pix. They are just different names and should function similarly.

[NormalUhr]

Actually, if you go through other unlearning methods included in this work, similar modifications can be found as well, just in order to make it possible to load the DM checkpoint fine-tuned on our dataset correctly.