search

Hzzone / ProPos

Self-Supervised Learning for Deep Clustering (TPAMI 2022)
104 stars 15 forks source link

About cifar10 performance on SimSiam #13

Closed kasteric closed 1 year ago

kasteric commented 1 year ago

Hi, I reran the experiments for SimSiam, but I can only achieve around 60% ACC, I am not sure where I was wrong, could u please share the config file for training SimSiam? Your response is really appreciated!

kasteric commented 1 year ago

I have tried many times, including using the .json included in the repo. image

kasteric commented 1 year ago

Hi, do you have time for sharing the config file for cifar10_simsiam？I have run the experiments multiple times, but the performance is not satisfying. Seems like the representation gets collapsed... Specifically I have replace the projection_num_layers from 3 to 2, and the feat_dim is set to 2048. Is the feat_dim 2048 for k-means too? Thanks so much! image

Hzzone commented 1 year ago

I have no time to rerun simsiam. There are some important tips on CIFAR if I remember correctly: 1. total bs 512 on single GPU; 2. projection_num_layers=2 instead 3, feat_dim 2048 (for k-means too); 4. weight decay 5e-4

kasteric commented 1 year ago

I have no time to rerun simsiam. There are some important tips on CIFAR if I remember correctly: 1. total bs 512 on single GPU; 2. projection_num_layers=2 instead 3, feat_dim 2048 (for k-means too); 4. weight decay 5e-4

Thanks for your fast and informative response! i will try on single card

kasteric commented 1 year ago

Thanks to your guidance! The problem has been solved!