We provide a few tutorials to get you started with Snorkel MeTaL. We recommend starting with the Basics Tutorial. The others can be completed in any order.
This tutorial walks through the basic training and evaluation process for the two primary classes of a Snorkel MeTaL pipeline: a label model for combining the votes from multiple weak supervision sources, and a discriminative end model for improved generalization and/or transferability to new form factors.
Learn how to use the multi-task versions of our models to utilize supervision sources that (implicitly or explicitly) label multiple tasks at once and capitalize on the benefits of multi-task learning. (NOTE: This is separate from the more recently introduced MMTL (Massive Multi-Task Learning) package, which has its own tutorial below).
See an example of how to easily run the standard machine learning task of classifying the CIFAR 10 image dataset using Snorkel MeTaL; building upon existing datasets like this can be a useful gateway to exploring multi-task learning in a well-known setting.
This tutorial demonstrates estimation of the class balance P(Y) using the ClassBalanceModel class.
This tutorial introduces our MMTL package, whose purpose is to enable flexible prototyping and experimentation in what we call the massive multi-task learning setting, where we have large numbers of tasks and labels of varying types, granularities, and label accuracies. Try the tutorial and checkout the package README for more information and resources.