A checklist for open-sourcing your code for reproducibility

• Start with a fresh computer

• Document in detail / document hierarchically

• Use docker or anaconda

Reproducibility in machine learning;

• Use yaml for hyperparameters; .yaml better for comments

• Have separate config file for each experiment

• Save within epochs

• Log all features of training and evaluation

• Experiment management; pytorch lightnight; scared; mlflow

• Set seed and save it in config

pip freeze > requirements.txt

• Test and validate (test the setup on a separate machine to ensure reproducible build;)

• Make sure dependencies are ok and not hardcoded paths exist in the code

• Checklist in the readme:

• - dependencies

• - training script

• - evaluation script

• - pretrained models

• - results

• Good to have

• - contributing guide

• - blog post