Louis Kirsch

In the code here we have just z-normalization for the inputs, no virtual batch norm. Also no hyperparameters for Atari. OpenAI, please be more reproducible! :-)

This is currently not supported. Because the training times are rather short, I am doubtful that this is a very important feature. But you may explore how this could be...

Have you trained your model using mfccs?

Hi Zhuliwen, can you have a look at the ray logs in `/tmp/ray`? Also have a look at #2, perhaps you are using the wrong versions of the libraries.

Seems like an OpenBLAS issue to me ``` OpenBLAS blas_thread_init: pthread_create failed for thread 41 of 56: Resource temporarily unavailable OpenBLAS blas_thread_init: RLIMIT_NPROC 515042 current, 515042 max ``` Perhaps this...

Hi Jiachen, thank you for your kind words. Apologies for the kind of outdated software stack (tensorflow 1.x aged quickly). Can you try running these commands in order? ``` git...

Hi Lucifer, have you compiled the plasma op first? Please also see my response in #2 for more help.

Perhaps the best solution would be to cache jitted executables to disk, eg using a python decorator that loads the executable if the signature and function name matches.

I have hundreds of thousands of files from both runs and artifacts to download and the API is by far too slow. I tried to parallelize it with threading but...

You are using a too new version of gym I believe. Try these: ``` absl-py==0.9.0 astor==0.8.1 atari-py==0.2.6 attrs==19.3.0 box2d-py==2.3.8 cachetools==4.0.0 certifi==2019.11.28 cffi==1.13.2 chardet==3.0.4 Click==7.0 cloudpickle==1.2.2 colorama==0.4.3 Cython==0.29.14 fasteners==0.15 filelock==3.0.12 funcsigs==1.0.2...