This package provides tools for training supervised machine learning models for computational pathology tasks using tile-level embeddings.
This repository accompanies the publication of the Paige Virchow foundation model suite. If you use this repository for academic research, please cite the following paper: Vorontsov, Eugene et al. “Virchow: A Million-Slide Digital Pathology Foundation Model.” ArXiv abs/2309.07778 (2023).
This repository is intended strictly for non-commercial academic research use. For commercial collaborations, please contact Paige AI under the appropriate terms. For detailed terms, please refer to our LICENSE.
cd paige-ml-sdk; pip install -e .
The SDK is equipped with a cli which can be used to train models and run inference. Run python -m paige.ml_sdk --help to get started, or refer to the examples directory for a basic tutorial illustrating how to use the SDK and how to organize your data.
The SDK and its CLI are powered by pytorch lightning which has many customizeable features.
By default, the sdk relies on Lightning's CSV logger which writes outputs to the lightning_logs folder. Use --trainer.logger and choose from any of Lightning's built-in loggers such as WandB or TensorBoard.
Use --optimizer and choose from any of the torch.optim optimizers. Similarly, use --lr_scheduler to adjust the learning rate.
Per pytorch lightning's documentation:
Callbacks allow you to add arbitrary self-contained programs to your training. At specific points during the flow of execution (hooks), the Callback interface allows you to design programs that encapsulate a full set of functionality. It de-couples functionality that does not need to be in the lightning module and can be shared across projects.
Lightning has a callback system to execute them when needed. Callbacks should capture NON-ESSENTIAL logic that is NOT required for your lightning module to run.
Callbacks can be used for many things, like saving model checkpoints, profiling exectution, or early stopping. By default, the CLI only applies a single callback which saves model checkpoints. Users wishing to configure this callback differently, or add additional callbacks may consider the options offered here: https://lightning.ai/docs/pytorch/stable/cli/lightning_cli_advanced_3.html#trainer-callbacks-and-arguments-with-class-type
Lightning's Trainer class automates most of the complexity surrounding model training. It handles the training strategy (e.g., DDP, FSDP), the choice of hardware (cpu, gpu), training precision, the number of epochs to train for, and much more. Most common practice tricks in AI engineering can be configured via the Trainer class. All of the available trainer flags are documented in the cli's help text:
python -m paige.ml_sdk fit --helpfor example, to train a model in 16-bit precision, set --trainer.precision 16.