search

The-Learning-And-Vision-Atelier-LAVA / DASR

[CVPR 2021] Unsupervised Degradation Representation Learning for Blind Super-Resolution
MIT License
395 stars 50 forks source link

Trainning only need HR images? #8

Open Macro03 opened 3 years ago

Macro03 commented 3 years ago

Hello, thank you very much for your excellent work. Does the trainning data only need HR images? Can I delete the self.dir_lr in df2k.py ?

LongguangWang commented 3 years ago

Hi @Macro03, thanks for your interest in our work. In our implementation, the LR images are generated online during training thus you can safely remove self.dir_lr.

Macro03 commented 3 years ago

I trained the Moco by the two real world images instead of 'degrade' images, the contrast_loss increase in early epoch and then decrease slowly. What do you think of it? Thank you so munch!

LongguangWang commented 3 years ago

Hi @Macro03, we also have this observation in our experiments on synthetic images. In my opinion, at early epochs, the random samples in the queue is gradually replaced with the training data. Thus, it becomes more difficult to distinguish positive samples from negative samples and the loss is increased. After a few epochs, the queue is completely replaced by the training data and the loss begins to decrease.