Recently I’ve found batch renormalization paper after investigating instability in training ( fluctuating validation loss) due to small batch size because of memory constraint - high res segmentation task. One option is to freeze or drop the batchnorm layer but the paper states that it substantially improves the training-inference performance. Does anyone have experience with it or a readily available implementation in pytorch ?
This allows to create larger batches on higher resolution images to compensate batch norm instability with small batches. Very useful in medical imaging overall.
Group norm (https://arxiv.org/abs/1803.08494) is also a great theoretical alternative but I never found pretrained weights on imagenet using group norm.
Did you get good results using inplace abn model in that github repo ? My best model is still single resnet18 Dynamic Unet LB: 0.731 with TTA but unfortunately couldn’t make deeplab v3 to give good results.