search

interpreting-rl-behavior / interpreting-rl-behavior.github.io

Code for the site https://interpreting-rl-behavior.github.io/
Creative Commons Attribution 4.0 International
0 stars 0 forks source link

Fix dataloader for gen model #33

Closed leesharkey closed 2 years ago

leesharkey commented 2 years ago

I think there might be a bug in the dataloader.

Sometimes true observation sequences go black even though the episode is nowhere near done. This will mess up the training of the gen model if it's in the dataloader; if it's downstream (e.g. in visualisation and saving) then it's fine.

I don't know the exact circumstances under which the bug occurs, but simply run train_gen_model.py and taking a look at some of the visualised samples. You may want to make sample-saving more frequent for debugging purposes.

danbraunai commented 2 years ago

I tried and failed to replicate this after looking at 70 samples . On the postaisc branch, I made some recordings, then ran train_gen_model, both with and without a pre-existing model_file.

danbraunai commented 2 years ago

Any ideas @leesharkey?

leesharkey commented 2 years ago

It happens probably more rarely that 1/70. But that's a guess.

Here is one example. The last frame inexplicably goes blank. https://user-images.githubusercontent.com/7349481/136906492-cbc16beb-9cc4-49be-8d97-9015399c4318.mp4

I don't know why this might happen. But I'm also no longer sure this is 'Important-Urgent' because we're retraining the gen model anyway now. I'll close the issue. (and I'll update the priorities for the other tasks)