training from scratch, loss decreased very slow

facebookresearch / TimeSformer

The official pytorch implementation of our paper "Is Space-Time Attention All You Need for Video Understanding?"

Other

1.51k stars 209 forks source link

training from scratch, loss decreased very slow #137

Open chrisx599 opened 2 weeks ago

chrisx599 commented 2 weeks ago

have anyone who training from scratch, not use pre-trained weight of ViT-B? Can author or anyone released the training log of train from scratch？ the author said it takes more epochs to training from scratch, but did'nt released a concrete number, can author says this number? when i trained 15 epochs, the accuracy is only 16%, and from the beginning, it's just 4%, i'm so confused is that normal?

chrisx599 commented 2 weeks ago

first figure is loss, seconde figure is accuracy