Apart from the requirements.txt from the helpers provided, I have used tensorflow-gpu 1.12.0 (normal tensorflow should work fine as well), keras 2.2.4, and h5py to save the keras models. Do note that I have written the code to run in Python 3, not 2! The venv I used is contained within the topmost requirements.txt. The fs1_8_offset contains experiments done with fs1, but the main implementation is found in nostate_32_offset_2_generalist.ipynb. In case of some unforeseen troubles I have also added the ipynb-code to a backup.py.