PSPnet and auxiliary classifier

I’ve read PSPNet paper so many times. I have searched all git repos and still I’m pretty much confused on how to train it with fastai.

PSPNet50 has
Resnet50:
layer 1
layer 2
layer 3 → auxiliary classifier(Single convolution with interpolation to return to initial size)
layer 4
pyramidpool:
priors
head

At start do I freeze the Resnet50 layers (+ layer4 batchNorm) and the pyramid_pool, and I just train the auxiliary classifier alone?

And if so, then do I unfreeze the whole network (except the auxiliary layer) and train the resnet50 and the pyramid_pool with different learning rates or do I go with sequential unfreezing?

1 Like