For the ienet2.py (hr_new in config) we take inspiration from superresolution works: we remove all batch/group normalization and initialize the residual paths such they have a lower initial contribution. Further, we add 2 stages to the HRNet. This further improves final quality as well as training stability.
The training config provided in train.yaml seems to only work for hr (ienet.py) – simply changing hr to hr_new didn't seem to work for me (generated images were really bad). Would you mind submitting the config used that gave the improved final quality and training stability? Thanks!
The README says the following:
The training config provided in train.yaml seems to only work for hr (ienet.py) – simply changing hr to hr_new didn't seem to work for me (generated images were really bad). Would you mind submitting the config used that gave the improved final quality and training stability? Thanks!