About training setting of real world BurstSR dataset about bipnet

[ZcsrenlongZ]

Hi! How long time you train this model? Thank you!

[ZcsrenlongZ]

Thank you for your wonderful work!

I'm trying to train with your code, and on the synthetic dataset, BipNet shows excellent results. But on the real burst dataset, It does not give as good results as described in your paper. The loss continues to increase except at the beginning. your code set 1e-5 ~ 1e-6 with CosineAnnealingLR as a default setting. If this learning rate is different from the setting of the experiment in the paper, could you share the training settings for the real burst dataset?

I'm really looking forward to your answer.

Have you any advice to tacle this problem?

[akshaydudhane16]

One thing you could try to stabilize the training is to consider a smaller patch size (like 24x24) for training. The aligned L1 loss is computed between the network output (aligned using PWCNet) and ground truth. As you consider a smaller patch size, the minor alignment error between the network output and the ground truth decreases and helps PWCNet in reducing alignment errors. Thus, there is a chance to compute aligned L1 loss with less error.