search

jiamingzhang94 / Unlearnable-Clusters

CVPR2023: Unlearnable Clusters: Towards Label-agnostic Unlearnable Examples
21 stars 1 forks source link

Implement detail question about train_gnet in main.py and Algorithm 1 in original paper #3

Closed pppppkun closed 1 year ago

pppppkun commented 1 year ago

In original paper, the line 9 of Algorithm 1 iter the images all in C_i to optim G. However, in Line 60 in main.py (function train_gnet), it is seem like iter all images in dataset. Is there something wrong with my understanding of Algorithm 1, or am I missing some detail in the code?

jiamingzhang94 commented 1 year ago

Hi @pppppkun , you are very observant. This is actually mentioned in section 4.1: "For each \delta_i, we repeated p times to train the generator G for entire datasets". In short, it is inspired by [1] that using more training data when training the generator of UAP can help to perform a better fit.

[1] Poursaeed, O., Katsman, I., Gao, B., & Belongie, S. (2018). Generative adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4422-4431).

pppppkun commented 1 year ago

Thank you very much for your answer!