How to use it as a feature extractor

Hi,

Is it possible to use this efficientnet implementation as a feature extractor? If so how should it be done?

Thanks in advance

[rwightman]

@jgmontoya … oh, yeah, if you just want the features prior to the linear, yeah, something like that will work, I assumed you were thinking about features at different resolutions for segmentation, etc.

You can do geffnet.efficientnet_b0(pretrained=True, as_sequential=True)[:-2] to have the flattened output before linear/droput, [:-1] to include dropout, [:-4] if you want to retain the last spatial feature map (before global pooling and flatten).

[rwightman]

@gaceladri I didn’t add the argument to the factory method in timm, effdet was being used more by fastai projects so that’s why it was added there, wasn’t clear anyone wanted that feature in timm. However, the EfficientNet models in both locations have .as_sequential() method that you can call after you create the model.

So, you can m = timm.create_model('efficientnet_b2a', pretrained=use_pretrained).as_sequential()

Another feature in timm, for all models you can just do model.forward_features(input) and you’ll get an unpooled feature output. In the future it’ll be possible to create any model with num_classes=0 or call .reset(0) on the model after creating and get pooled featured outputs by default from the normal forward() method.