Our repository is at link Quickstart notebook is available at the top level of our repo, and is named sim2real_quickstart.ipynb
-
- FID loss
- cycada implementation
- choice of the GAN evaluation metrics huggingface metrics
-
- add code for logging the metrics and loss at each epoch, to the code in our repo
- CycleGAN
- tune (refer to the original paper and the huggan implementation at link to better understand their meaning) :
- learning rate
- decay_epoch (start decaying the learning rate after the epoch number='decay_epoch' )
- (optional ?) lambda_id (identity loss weight)
- (optional ?) lambda_cyc (cycle loss weight)
- (optional ?) n_residual_blocks
- tune (refer to the original paper and the huggan implementation at link to better understand their meaning) :
- Cycada
- learning rate
- ...
-
- dataset creation
- test different GAN architectures for the sim2real translation:
- CycleGAN
- Cycada
- ...
- architecture modifications(list the chosen ones and update when finished):
- FID loss CycleGAN
- ...
- hyperparameter optimization for each model
- CycleGAN
- Cycada
- ...
Chris1/sim2real_gta5_to_cityscapes (unpaired image-translation dataset)
To load the dataset within a python script or notebook simply do the following
from datasets import load_dataset
dataset = load_dataset("Chris1/sim2real_gta5_to_cityscapes")
-
- group results of the tested GANs in a benchmark table, with respect to a common metric, examples are the huggingface metrics
-
- upload models with huggingface specification link
-
- showcase results in SPACE demo Huggingface link
-
- create model card
-
- create SPACE card and demo
- (optional) benchmark the found GANs with the purpose of semantic segmentation in the real domain e.g. synthetic images are translated to real and used to train one or more semantic seg models with supervision on the translated data
The full cityscapes dataset, with train, validation and test splits. Images are paired with the associated semantic segmentation (CARE, the test set does not have ground truths, as it is used to produce predictions for the cityscapes evaluation server)
The full GTA5 dataset, with train, validation and test splits. Images are paired with the associated semantic segmentation masks