tding1 / neural-collapse Goto Github PK

Hello, thanks for the nice work. I'm running the validate_NC.py and I wonder how should I interpret collapse_metric values < 0? Thanks.

Hi, when running validate_NC with gpu_id = 1, there is an issue where certain neural collapse metrics are computed using .cuda() instead of .to(device). The device is correctly specified in args.device as used in compute_info, compute_Sigma_W, and compute_Sigma_B, but in compute_ETF, compute_W_H_relation, and compute_Wh_b_relation, .cuda() is used instead of device. On my machine, .cuda() defaults to GPU 0, which leads to an error whenever I specify gpu_id=1 in which case there are tensors being loaded on both GPU 0 and 1. Maybe this is an issue on my end with how my defaults are set for cuda, but it seems like it could be resolved by using .to(device) instead of .cuda() in these instances. Thanks