Somshubra Majumdar

Here is another implementation for Tensorflow which follows the same process as the paper. https://github.com/shaohua0116/Group-Normalization-Tensorflow/blob/master/ops.py

Update: Seems you are right. For TF, it needs to transpose to NCHW first, then perform reshape and then finally transpose back. Thank you for catching it. I will update...

I quite forgot to write this. I've kinda moved on from Keras to TF.keras, and that already has Group Normalize in the TF add-ons. If you would like, could you...

Could you specify what error you got? I tested this with TF backend, Keras 2.1.3 and Python 3.5

Open cv is a requirement for this. Try conda install opencv

On Windows, opencv has many problems during installation. It should be import cv2. Even then, there can be some internal import problems from open cv itself.

That's because there are no pretrained models for scale 5. You'll have to train your own, or pass the scale 2 model 3 (8x scale) times and then downscale.

Could you list the entire error report and where it occurs?

Fantastic. I'll give it a look and try to incorporate batchnorm into it.

Yes that is quite possible. I was thinking of adding gradient clipping eventually. The codebase is quite simple, could you try to add gradient masking ?