search

xinntao / EDVR

Winning Solution in NTIRE19 Challenges on Video Restoration and Enhancement (CVPR19 Workshops) - Video Restoration with Enhanced Deformable Convolutional Networks. EDVR has been merged into BasicSR and this repo is a mirror of BasicSR.
https://github.com/xinntao/BasicSR
1.48k stars 320 forks source link

Do I need to change any model settings for different provided models? #66

Open Cugtyt opened 5 years ago

Cugtyt commented 5 years ago

If I test different models in model zoo, do I need to change some code or settings?

xinntao commented 5 years ago

Just change the data_mode = 'sharp_bicubic' in test_Vid4_REDS4_with_GT.py is OK. The testing codes will do the modifications.

Cugtyt commented 5 years ago

During training, how to use the deblur module? I notice that there is a deblur option, is that just turn it on is ok? (I am using my dataset)

xinntao commented 5 years ago

The only difference between SR and deblur is that they have different input size. So you only need to change HR_in from False to True.

For deblurring, it is better to have a predeblur model before alignment and fusion. So You can also turn the predeblur into True. But in our released model (used in the competition), we first pre-train this predublur module and then train the whole model. If you only set predeblur: True and do not have a pretraining, it may reduce performance.

Cugtyt commented 5 years ago

Could you please provide more information about the predeblur mudule training? I tried to add conv layers after predeblur to reconstruct clean image and train with the whole model, but didn't see any improvement. Could you give some advice? Thanks.

xinntao commented 5 years ago

We first train the preblur part separately (using another model only for reconstructing the clean image). Then in the large EDVR model, we first initialize the predeblur weights with the pretrained one and set X2 smaller learning rate of the predeblur module. And then train the whole model.

I think it is not the optimal strategy. Having another supervision like reconstructing a clean image is reasonable. I think the weight of this deblur loss term should anneal as the training goes. Did you try it?

gulzainali98 commented 2 years ago

hey, @xinntao can you please help me out here with how this would work with basicsr framework?