This is a implementation of Large-Scale-Cruiosity using Pytorch, here is the origin implementation and here is the paper. I don't use MPI or multiprossing because my laptab can't run with them. I tried but failed.
I test it on Atari and it works, and I have not run it on Roboschool or Unity3D.Maze. I really need Unity3D.Maze environment but can't find it in the origin implementation.
Another thing, I don't implemente pixel2pixel feature, don't use it.
Python 3.6+
Pythorch 0.4
Download this repo and run run.py
python3 run.py --feat_learning none # for random feature
python3 run.py --feat_learning idf # for IDF
python3 run.py --feat_learning vaesph # for VAE