List all classifier models

[hendriklohse]

Hello!

For a project I'm trying to evaluate different classifier model architectures, and for that it would be nice to be able to list all timm classifier models to loop over (e.g. something like timm.list_models("*efficientnet*", pretrained=True, type=classifier)). After multiple long dives into the code and documentation, I did not find such a feature, so I decided to ask it here. Currently I am listing all models to loop over, and check manually with hasattr(model, "classifier"), but as this is only possible after creating the model, it seems very inefficient.

[rwightman]

@hendriklohse all of the models are current classification models (different tasks like segmentation, obj detection comming). Whether the attr is named .classifier, or using the Classifier(nn.Module) is down to the original weights. I originally preserved 3rd party weigth compatibility by using the same attr name as the original (so unfortunately there is .classifier, .fc, .head, etc) but they are all nn.Linear(features, num_classes) classifiers for the actual layer. Looking back I should have just remamped them all when I had the chance :)

If you are looking to use all models consistenly in a different task, you can remove the classifier consistently with the reset_classifier() (or on creation with global_pool and num_classes args) as per the docs and then wrap the 'body' / 'backbone' in your own model with custom head.

The name of the classifier (the final nn.Linear) is also currently available in the pretrained_cfg (transitioning away from default_cfg) for each model-weight combo. You can query those with snippet below (now api for get_pretrained_cfg) fairly new in 0.6.x. As to whether the instance is part of a Classifier or nn.Sequential, or a standalone nn.Linear that's a bit trickier, if it's name.fc it's often a Classifier module, but I don't see why this would matter until you actually had an instance that you were going to manipulate...

for m in timm.list_models(pretrained=True):
    print(m, timm.get_pretrained_cfg_value(m, 'classifier'))

[hendriklohse]

@rwightman Thank you for the quick and elaborate answer!
I am pretty sure I solved the problem now.
Just a checkup question: is
testOut = torch.randn(2, 3, imgSize, imgSize)
model.reset_classifier(0)
in_features = model(testOut).shape[1]
a consistent way to get the last layer in_features you'd normally get with model.cl.in_features where in this case cl is one of timm.get_pretrained_cfg_value(m, 'classifier'), e.g. fc ?