Finetune pretrained model with custom number of classes

[ayanasser]

Hi,
Do you have the ability for finetuning slowfast model but with 2 classes only?
in another word (could I change the number of classes of the last layer of "pretrained slowfastNet model"

because I want to take the feature from that pretrained model then classify those features with a simple 2 dense layers,
but I need to end up with one model, not 2 models (one for feature extraction and the other for classifying)
Thanks in advanced

[dreamerlin]

Refer to: #19 (comment)
Or you can modify load_from to the pretrained model checkpoint path or url, then change the num_classes, it will ignore the unexpected keys.

[ayanasser]

I want to freeze all the layers except the classifier layers (the last 2 layers), Does that accessible ?

[dreamerlin]

Try to use with torch.no_grad(): to wrap x = self.extract_feat(imgs), it will freeze the backbone.

[dreamerlin]

And If you feel we have help you, give us a STAR! : )

[ayanasser]

you helped me a lot, sure I will, thank u ^^

[ayanasser]

Hi again :D, u said before that i could freeze the backbone by wrapping the features extracted with torch.no_grad(),
you mean by the backbone "resnet50" or you mean all the layers above the final softmax layer?
because while finetuning the slowfast model i need to freeze all the layers above the final softmax one,

second thing is i found that the backbone recognizer for slowfast is 3D not 2D, Is there a big difference between the 2 recognizers whilem finetunining the model with my custom video dataset ?