conda create -n 2S-ODIS python=3.10
conda activate 2S-ODIS
pip install -r requirements.txtDownload other required files (VQGAN):
sh setup_vqgan.shCheck if vqvae_models.pth is generated.
Place the SUN360 outdoor data one level up (in the same directory as the 2S-ODIS folder).
sh generate_dataset.shLeave it for about a day, and the dataset will be created automatically.
python train_first_stage.py
python train_second_stage.py --bf16Train each for two days to save the trained models on mlflow.
Please enable bf16 for the second stage.
Ensure to note down the path where weights are saved on mlflow for inference.
To view experiment logs in mlflow, navigate to the 2S-ODIS folder and run:
mlflow uipython image_generation.py --lowres_base_path="path_to_lowres" --highres_base_path="path_to_highres"Example:
python image_generation.py --lowres_base_path=mlruns/1/0910882ab5ad4faeadbbd2288b176980/artifacts/ --highres_base_path=mlruns/2/7e71d11321ce4f0da773f8723861c077/artifacts/Generating 5000 images takes approximately 2 hours in RTX3090. Each image is created in about 1.5s.
To evaluate, navigate to the eval folder and run:
python evaluation.pyEvaluation results will be computed in eval/results.
- NVIDIA RTX 3090
- CUDA 11.8
Apache License 2.0