reds-lab/lava

This is an official repository for "LAVA: Data Valuation without Pre-Specified Learning Algorithms" (ICLR2023).

lava's Issues

Experiment dataloaders missing

At the moment you have only released the code for the noisy-label experiments: https://github.com/ruoxi-jia-group/LAVA/blob/main/lava.py#L60. Can you release the data loaders for the noisy features, backdoor, poison attack, noisy feature and irrelavant data detection please?

Thanks in advance!

Logical bug in the code

Hi folks, thanks for your excellent work!
I have used the repo and I found a very key bug in the following position.
Actually, in your model architecture, there is no layer called ".fc", therefore, directly replace to identity layer will not transfer the network architecture to embedder (actually not change the network architecture at all, just add another identity layer in the network class not to use). It is better to fix the bug in the repo.
https://github.com/ruoxi-jia-group/LAVA/blob/84e8bd55e62239625d8728240ec041d2fca5b2f4/lava.py#L88