[Question] ABMIL training tricks

[Rainydu184]

What's your question?

Hi, thanks for your great work on this project!

I have a question regarding the ABMIL implementation in your codebase. I noticed that during training, a fixed number of features (patches) are sampled from each slide, regardless of the total number of available instances.

Since ABMIL relies on attention to identify important instances within a bag, wouldn't using a fixed-size random subset risk missing key patches, especially in slides with a large number of instances? Did you observe that this improves stability or generalization during training?

Would you recommend using dynamic sampling (e.g., variable number of patches) or even using all available instances during validation/testing for better performance evaluation?

Looking forward to your insights. Thanks again for your work!

[ajv012]

Thanks for your question. A couple of thoughts:

Why we sample during training

• Regularization: Sampling a fixed number of random patches per slide each epoch helps prevent the model from overfitting to any single region.

• Computational efficiency: Working with fewer patches reduces memory and compute requirements, enabling larger effective batch sizes and smoother, faster iterations.

• Diversity over time: Because we resample randomly on every epoch, the model still “sees” a wide variety of patches across training, improving generalization.

• Scalability: Fixed-size sampling can allow you to train with batch sizes > 1.

• Better sampling: You can swap in more “intelligent” sampling strategies—e.g. cluster all patches first and then sample evenly from each cluster—to focus on representative regions.

• Sampling will not always work: random sampling should not be used with MIL models that implement position encoding, like GigaPath.

What happens at validation/test time

• No sampling: We use all patches, one slide per batch (i.e. batch size = 1).

• Deterministic evaluation: Because there’s no randomness in validation or testing, results are reproducible and not affected by different random seeds.

Please let me know if you’d like more detail on any of these points!