VTN not converging on UCF101 about slowfast HOT 2 CLOSED

Hi Daniel,

Thank you for your suggestions- I tried using SGD and it surprisingly works i.e. training loss started decreasing. Previously, when I was using ADAM optimizer, none of these settings worked: pretrained Kinetics/ImageNet, scratch with various learning rates from 1e-3 to 1e-4.

Thanks again! Closing the issue :)

-Ishan

from slowfast.

Hey Ishan,

Let's try to break this issue into smaller parts.
Few questions to start with:

1. Are you using the pre-trained kinetics model? (Transformers are very hard to train from scratch)
2. Did you try different lr's with adam? usually, the right lr zone is very narrow.
3. Did you try different optimizers like SGD? In our original paper, we trained with sgd, if you trained with kinetics pre-trained model it usually the sgd coverages

Thanks,
Daniel

from slowfast.