Customized optimizer

[cswpy]

Feature

Support for customized optimizers during training allows for greater flexibility in training models. Some frameworks in WSI tasks consist of multiple networks. Right now, we can only build a model with a single network.

Pitch

There are some models that require customized optimizers, typically models that are consisted of several modules. For example, WENO is consisted of a student network and a teacher network. Right now, I cannot implement WENO using SlideFlow.

Alternatives

I don't think the current version has support for such knowledge distillation network.

Additional context

[jamesdolezal]

Thanks for this suggestion. At present, Slideflow could be used for image processing and preparation of feature bags (in PyTorch format) for external, customized MIL training loops like WENO. I will look into whether we can extend our MIL training framework to add support for custom optimizers. I don't think it would take too much work.