Pseudo-labels for real domain segmentation

cc-ai / climategan

Code and pre-trained model for the algorithm generating visualisations of 3 climate change related events: floods, wildfires and smog.

https://thisclimatedoesnotexist.com

GNU General Public License v3.0

72 stars 18 forks source link

Pseudo-labels for real domain segmentation #140

Closed alexrey88 closed 3 years ago

alexrey88 commented 3 years ago

Really small modifications so that we directly load the .pt tensors of the seg maps for the real domain (instead of the .png images), to allow using pseudo-labels for segmentation task.

For now, .json file with .pt extension are in:

/network/tmp1/ccai/data/omnigan/base/train_r_full_pl.json
/network/tmp1/ccai/data/omnigan/base/val_r_full_pl.json

When (if) this PR is merged, I will change {train/val}_r_full_pl.json for {train/val}_r_full.json.

Experiment with use_pseudo_labels=true: here

vict0rsch commented 3 years ago

@alexrey88 just changed names so that it makes more sense