Rob

People criticizing Jupyter notebooks always mention the same issue: out of order cell execution. The cure is simple: always execute all cells in a notebook in order, as you do for any code in any language. Anything else is silly, to be polite. Then why not use plain python files instead? Notebooks are great when you want to have side by side: - documentation - code - code output Don't use notebooks if this is not important to you. For instance, don't use notebooks to run production code, as you don't need cell outputs display. Running notebooks drags unnecessary runtime. I use notebooks in the early stages of a ML project (e.g. a Kaggle competition), for EDA and for early stages of my pipeline. As soon as a part of the pipeline is mature enough (say data loading), I move it to regular python files. The notebook is used to run the overall pipeline, with detailed execution of the part I am working on. At some point everything is moved to python files, and my notebook contains one cell per training experiment experiment. If using Pytorch, each cell may start with: ! torchrun ... I have the output of the run right below the cell. And yes, I also use logging to files or wandb for long term persistence. Last but not least. I noticed that those who never use notebooks also tend to never look at data. Not a good thing if you build ML solutions.