search

allenai / satlas-super-resolution

Apache License 2.0
190 stars 24 forks source link

Training ESRGAN on PROBA-V #13

Closed yunseok624 closed 5 months ago

yunseok624 commented 5 months ago

Hello!

I'm trying to train the esrgan on proba-v dataset and I came across some questions I coudln't find in the paper.

  1. You took the dataset from the competitions, am I correct?
  2. Did you pre-train a PSNR-oriented model with the L1 loss with proba-v dataset (just like in esrgan and real-esrgan did)?
  3. In the probav_esrgan.yml I don't understand the number of input channels. in RRDBNet shouldn't the input channel be 9 since proba-v (from the competition) has only 1 channel? Why is the output channel 3? I also do not understand why the input channel for discriminator is 30. image

Thank you in advance.

piperwolters commented 5 months ago

Hello, the experiments branch is a work in progress. You can adjust the args in the config to match what you have described. And the discriminator takes in the low-res images + the real/fake image as input.

We did not pretrain a PSNR-oriented model like in real-esrgan. We just trained the whole model with all losses that are specified in the config files.

And yes, we used imagery from the PROBA-V challenge.