Set num_features of an EfficientNet?

[sk-zk]

I'm trying to change the size of the feature vector in an EfficientNet. What's the best way to do that? I've tried just passing it to create_model, but that didn't work:

>>> timm.create_model("efficientnet_b1", num_classes=0, num_features=512)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Python39\lib\site-packages\timm\models\factory.py", line 74, in create_model
    model = create_fn(pretrained=pretrained, **kwargs)
  File "C:\Python39\lib\site-packages\timm\models\efficientnet.py", line 1355, in efficientnet_b1
    model = _gen_efficientnet(
  File "C:\Python39\lib\site-packages\timm\models\efficientnet.py", line 835, in _gen_efficientnet
    model_kwargs = dict(
TypeError: dict() got multiple values for keyword argument 'num_features'

[rwightman]

@sk-zk num_features is considered part of the backbone architecture def so it's currently fixed for each model def (ie efficientnet_b1), just to be clear you are aware that's the output chs of the penultimate pre-classifier layer? It is not the number of outputs for the model classifier head (num_classes).

It could be changed to support making it easier by explicitly handling kwargs vs defaults, but it'd be some code to write/test as some of the num_features for different sub-families change based on size, rounding modes, etc. Making it work with pretrained weights though would be a bit of a mess.. hmm

[sk-zk]

Oh, I see.

just to be clear you are aware that's the output chs of the penultimate pre-classifier layer? It is not the number of outputs for the model classifier head (num_classes).

Yeah, I specifically want the embeddings without a classifier

[rwightman]

but you're trying to change the size of the embeddings, not use them, what you want to do would invalidate the pretrained weights and require you to train the model yourself... if you want the embeddings you just need to remove the classifier with num_classes=0 or model.reset_classifier(), covered in the docs