search

NoSavedDATA / PyTorch-BBF-Bigger-Better-Faster-Atari-100k

12 stars 1 forks source link

Reproduce BBF result #1

Closed hlsafin closed 5 months ago

hlsafin commented 5 months ago

Were you able to reproduce all the BBF results in the paper? I see a few of them posted on the github page, but not all of them.

NoSavedDATA commented 5 months ago

You can refer to results.csv at the main page. The results at the read.me are from an older version. I've executed the bbf.ipynb file.

I have not replicated the result of all the games yet. Unfortunately, using a single 4090 still takes a lot of time, as you have to train the model from scratch for every game and every seed, which takes around 6 hours for this GPU. And currently I'm not at the same city where the 4090 is at.

I'm doing my bachelor's thesis about this network though, but it is not going to be in english.

hlsafin commented 5 months ago

Is it possible to train enviroments with multiple GPU in order to make the training process faster? I assume the bottleneck is the large network?

NoSavedDATA commented 5 months ago

It's a pretty small network of 35 M parameters. There is a certain bottleneck on the evaluation environments, which I did not know how to implement parallely.

Also I do not have access to multi GPU servers.

hlsafin commented 5 months ago

okay thank you