##### ###### ##### ### # # ### # # ###### ## ## ## ## ## ## ## # # # # # ## ##### #### ##### # # # # # # # #### ## # ## ## ## ## # # # # # ## ## # ###### ## ### # ### # ######
##### ###### ##### ### # # ### # # ###### ## ## ## ## ## ## ## # # # # # ## ##### #### ##### # # # # # # # #### ## # ## ## ## ## # # # # # ## ## # ###### ## ### # ### # ######
Make your training reproducible:
Set random seeds for Python, NumPy, and PyTorch
Version your data and code
Log all hyperparameters
Use deterministic algorithms when possible
Perfect reproducibility is hard with multi-GPU training. But close is better than nothing. When something works, you should be able to do it again.
