search

NVlabs / I2SB

Other
260 stars 23 forks source link

Training I2SB on cross-domain image-to-image translation tasks such as Day to Night #2

Open prahlad-anand opened 1 year ago

prahlad-anand commented 1 year ago

Hello, Trying to train the model on a task similar to Day to Night (RGB to IR). I do not want to perform any of the available restoration tasks. How exactly do I perform the same?

The flag --cond-x1 is mentioned, however, this seems insufficient and I am looking for some more detail if possible.

I would also like to know how to modify my dataset of RGB and IR images so I can train the model.

Thank you.

cychoi97 commented 1 year ago

Hello, I also tried to train the model on a general img2img with paired dataset. In my case, I gave the option --corrupt 'mixture', then I passed the following code in train.py.

if opt.corrupt == "mixture":
    import corruption.mixture as mix
    train_dataset = mix.MixtureCorruptDatasetTrain(opt, train_dataset)
    val_dataset = mix.MixtureCorruptDatasetVal(opt, val_dataset)

If you already have paired dataset, this process doesn't need. Maybe it works... try it!

Thank you.