Benchmarks that measure the cost of validating training data with pandera, and the cost of not validating it.
Data validation
Tutorials that validate training data with pandera and measure what it costs. Each one is a benchmark you can run on your own cluster, and each pairs a validated mode with an unvalidated one on the same data, so the difference between them is the thing being measured.
To run one, clone unionai-examples and run its main.py from the
tutorial’s directory under v2/tutorials/. A cluster run connects to the cluster in your
Flyte configuration file; each
tutorial’s “Run it” section also shows a local run.
Per-batch validation overhead
Measure how much wall-clock time pandera adds to a training loop when it validates every TensorDict batch and skips the ones that fail.
Whole-dataset validation
Measure how pandera validation time grows with dataset size when you validate a whole TensorDict dataset once, where it’s produced.
The cost of not validating
Measure what happens to training when corrupt batches reach the model, compared with a loop that validates every batch with pandera.