search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
415 stars 77 forks source link

EMA model not working #103

Closed Caspeerrr closed 1 year ago

Caspeerrr commented 1 year ago

I use the default configurations etc on a custom dataset and I didn't change anything in the code. However, during evaluation the map score is very low (10 %)

When I change: model.load_state_dict(checkpoint['state_dict_ema']) to: model.load_state_dict(checkpoint['state_dict'])

the map score is a lot higher (~40%). Why isn't the EMA model working? thank you!

happyharrycn commented 1 year ago

To get the EMA model working, you will need to train the model for many iterations, so that the learned weights can be aggregated into the EMA model. With a short training schedule (# epochs) or a small dataset (# mini batch per epoch), the number of iterations might not be sufficient. If an early stopping is strongly preferred (usually due to a small dataset), EMA should be turned off.

happyharrycn commented 1 year ago

Closed due to inactivity.