Check if optional parameters are supported by a specific model

[materight]

Is there a way to check if a certain parameter to pass to create_model is supported by a certain model?
For instance, we are using timm on our training pipeline and we want to run hyperparam search over multiple architectures. We want to try drop_rate, but that's not supported by every model, so we want to pass it to create_model only if supported.

Is there a simple way to check if a particular param is supported by an architecture?

[rwightman]

Best way is to try it and catch the error if it fails... I feel it would be a significant undertaking and a lot of added complexity to figure out a way to make this easy to check/report (any great ideas here?). I've seen others swallow unused args silently,, but I'd rather it fail so that it's known and one can decide to retry or not...

[materight]

Ok thanks, I was wondering if there was a supported version to do that.

We ended up using the type annotations from the registry's entrypoint functions to get the set of allowed params, like this:

import inspect
from timm.models import registry
model_name = 'resnet18'
model_class = inspect.signature(registry.model_entrypoint(model_name)).return_annotation
model_params = inspect.signature(model_class).parameters