BYOL and SimSiam Model Collapse Problem
jqwangai opened this issue · 1 comments
jqwangai commented
In training BYOL or SimSiam, the model tends to fall into collapse. Have you ever encountered this problem and how did you solve it? Looking forward to your reply!
Tian0426 commented
Hi @jqwangai , we also noticed that BYOL and SimSiam are sensitive to the choice of hyper-parameters and easy to collapse. For BYOL, we empirically found using a larger learning rate for the predictor effectively prevent collapse. You might want to consider using the settings in the script we provided in the README file, FYR:
python main.py --framework 'byol' --backbone 'DCL' --dataset 'ucihar' --aug1 't_warp' --aug2 'negate' --n_epoch 60 --batch_size 128 --lr 5e-4 --lr_cls 0.3
python main.py --framework 'simsiam' --backbone 'DCL' --dataset 'ucihar' --aug1 't_warp' --aug2 'negate' --n_epoch 60 --batch_size 128 --lr 5e-4 --lr_cls 0.3
Hope this helps.