Question of DER: freeze the old module is left out in the file der.py

[fanyan0411]

Why "freezing old network" of after_task is left out in file der.py, while other methods, such as podnet and WA execute this step as follows:
self._old_network = self._network.copy().freeze()?

[G-U-N]

Hi, thanks for your interest. WA and PodNet are distillation-based methods, meaning they have to save the model in the last incremental session, while DER does not.

[fanyan0411]

Thanks for your timely responce!

In original paper, DER noticed that "we freeze the previously learned representation" for several times. If the previous convs in self.convnets are not freezed, which is also not executed in the code, does it means that those previous "resnet" modules will be affected by the back propagation? How to understand "freeze the previously learned representation"?

[G-U-N]

I understand your confusion now. Refer to Line51 and Line84.

[fanyan0411]

That the key point I am wondering about. Thanks a lot!

I hope your work keeps getting better！