Why do teams version their models and datasets, not only their code?
Options
- Because Python requires it
- To make training faster
- To reproduce a result, compare versions and roll back, by knowing exactly which data and code produced each model (Correct option)
- To make files smaller
Answer
To reproduce a result, compare versions and roll back, by knowing exactly which data and code produced each model
Why
A model is the product of code, data, parameters and environment. If you only version the code, you cannot rebuild the model, because the data may have changed.
With versions you can answer "what is running in production?" and "what changed since last month?", and go back if the new one is worse.
Tools range from Git plus data versioning tools to a model registry.
Remember
Version data and models to reproduce and roll back.