Machine Learning

PyTorch defaults to using Pickle for tensor storage, which poses security risks as malicious pickle files can execute arbitrary code upon unpickling. In contrast, safetensors specialize in securely storing tensors, guaranteeing data integrity during storage and retrieval.

safetensors also uses zero-copy operations, eliminating the need to copy data into new memory locations, thereby enabling fast and efficient data handling.

Link to safetensors.

After training your machine learning model, deploying it for real-world predictions can be complex.

MLFlow simplifies this by offering a user-friendly interface to deploy models across various platforms without requiring boilerplate code.

With MLflow, the model, code, and configurations are packaged with the deployment container, ensuring consistency between training and deployment environments.

Learn more about MLFlow deployment.

There’s a common belief that fresher data yields better results, but how frequently should you update your models?

To figure this out, train models on different past timeframes and test them on current data.

For instance, train model A on January-May data, model B on April-August, and model C on July-November, then evaluate all on December data.

If model A performs much worse than model C, you should consider updating your model more frequently to maintain high performance.

Reference: Designing Machine Learning Systems by Chip Huyen.

Logging parameters, metrics, and models allow you to reproduce experiments and select the best-performing model. However, writing multiple log statements can be cumbersome.

For automatic logging, add mlflow.autolog() before your training code.

MLflow aliases enable you to assign meaningful names to different versions of machine learning models.

With aliases, you can effortlessly switch between different model versions without modifying the deployment code. This is particularly useful when promoting a new model version to production.

Learn more about MLflow aliases.

To get estimated prediction intervals for predictions made by a scikit-learn model, use MAPIE.

In the code above, we use MapieRegressor to estimate prediction intervals for a scikit-learn regressor.