Depth of CFAT

[PhuTran1005]

Depth of CFAT

[rayabhisek123]

Thank you for pointing out. Six alternative layers is the right one and one 8 is missing in the supplementary material. You can confirm it from the official code also. We will update it in our subsequent arXiv version. Thanks.

[rayabhisek123]

If our work is really able to serve the benefit of the CV community, please be with us in expanding the reach by sharing this work and adding stars to this repo. Thank you!

[PhuTran1005]

Sure, and I hope you can release the pre-trained models and the config file (.yml) soon that help us to retraining or reproduce your work.
Thank you again!!!

[rayabhisek123]

Yes, we will complete the #ToDoList soon. Thank you!

[rayabhisek123]

Yeah, in some paper you can see 500k iters because they are only consider the DIV2K dataset for training with a batch size of 16. But for larger model, we consider both DIV2K and Flikr2K as training dataset with a batch size of 32. For this setting, 250k iters is enough to sufficiently converge.

[PhuTran1005]

Yeah, i understood. Thank you for your quick reply.

[PhuTran1005]

I have one more question about training setting. You used only 250k iters for training CFAT whereas many other papers used 500k iters. Model only needs 250k iters for convergence? Have you tried with 500k iters for training, and can it improve the current performance?