search

facebookresearch / maskrcnn-benchmark

Fast, modular reference implementation of Instance Segmentation and Object Detection algorithms in PyTorch.
MIT License
9.29k stars 2.5k forks source link

Save best model #326

Open wesleylp opened 5 years ago

wesleylp commented 5 years ago

❓ Questions and Help

I don't see the best model on validation being saved. Did I miss it or it wasn't really implemented? If not implemented, any reason why?

fmassa commented 5 years ago

Hi,

We currently save all the checkpoints, and only perform validation in the end model.

We could perform validation after every checkpoint, but that's not implemented ATM (and follows what was implemented in Detectron). I agree though that it would be nice to have evaluation after every checkpoint. The way it was implemented was to have a watcher job which would wait for checkpoints to be available and it would kick evaluation right away, on a separate machine.

sno6 commented 4 years ago

@fmassa Hi there. Is this still on the cards? this would be a great feature in my opinion.