Deepak Pathak

This is a CPU implementation and does not use GPU. One way is to resize the image to smaller size and later upsample the flow image.

@jremmons I would recommend using it at 100x100 and then resize flow field back to 256x256. I wrote `pyflow` only because Farneback and other default algorithms in OpenCV were too...

@ryanjay0 @jremmons There is also an OpenMP multi-processing pull request here #3 . Depending on how many cores you have that can significantly improve the running time.

Oh, it has been updated -- thanks for spotting it. Are you sure that changing the version of universe doesn't break other things?

Alright, let me answer your "difficult" question since I myself went through these steps 1.5years back. I guess universe-starter-agent has correct implementation of A3C but definitely with quite a few...

You can try tuning the hyper-parameters of [state-predictor version of curiosity](https://github.com/pathak22/noreward-rl/blob/master/src/model.py#L315-L384) from the `noreward-rl` codebase on the Montezuma game. That definitely has higher chances of vanilla `universe-starter-agent` as the latter...

@TianyuanYu There should not be any need for child mode. I am able to run it as mentioned in README. @AdamStelmaszczyk You have to install all dependencies on host machine....

Can you try also passing `--savio` flag to train.py ?

I am sorry -- I realized that later. Can you debug inside your tmux sessions (using pdb or Ipython debugger) as to which version of python is being called inside...