search

OPTML-Group / UnlearnCanvas

UnlearnCanvas: A Stylized Image Dataaset to Benchmark Machine Unlearning for Diffusion Models by Yihua Zhang, Chongyu Fan, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Gaoyuan Zhang, Gaowen Liu, Ramana Kompella, Xiaoming Liu, Sijia Liu
https://unlearn-canvas.netlify.app
48 stars 2 forks source link

Object Erasure Code #5

Open sophistz opened 3 months ago

sophistz commented 3 months ago

Hi, is there any script for object erasure training and evaluation code? Thank you.

NormalUhr commented 3 months ago

Hi, for each machine unlearning method, you can simply replace the artistic style name with the object name to enable object unlearning. The evaluation code is also provided and the instruction is in the machine_unlearning folder. We are also happy to help if you have questions related to a specific unlearning method.

sophistz commented 2 months ago

Thank you for the reply. When I tried to evalutate object erasure, I noticed that for some method (e.g. SalUn), we cannot simply replace theme name with object name. Because in train-erase.py, the dataloader loads pairs of stylized images and seed images (with "photo style" prompts). If we replace themes by objects, the length of forget_dl and remain_dl did not match, and the image pairs also become dismatched. In other word, as I understand it, seed images are clean images for stylized ones, which can provide guidance in style unlearning. I don't know if that true, so I still wonder how can I evaluate object erasure performance for methods like SalUn.