This is probably better suited for the PyTorch forums but…
So I tried using the new nn.utils.weight_norm API on the convolutional layers of one of my UNet models, dropping all BNs. It trained like normal, getting to about 90% of the loss/accuracy of the BN model, then dropped to about 50% and got stuck there permanently… no longer training at all.
Have any of you experimented with / had successes with weight_norm in deep convolutional networks?