This is a pytorch implementation of models based on BicycleGAN and Soft-Intro VAE on Image to Image Translation. In this project, we explore the tradeoff between fidelity and diversity for image translation tasks.
- We inherit the model architecture from BicycleGAN for the Soft-Intro VAE.
- We explore different hyperparameters for Soft-Intro VAE and obtain the best result of 73 in FID score.
- We explore different loss function for two different architecture configurations.
All of our models are trained on the Edge2Shoes dataset which consists of 50,025 shoe images and their corresponding edges split into train and test subsets.
$ pip install requirements.txt
For training please run the following command to train the model, the output image will save under the auto create directory BCGAN_results/.
$ bash scripts/experiment.sh
For training please run the experiment_vae.sh to train the model, the output image will save under the auto create directory VAE_val_results/.
$ bash scripts/experiment_vae.sh
To run the inference, you first need to set the config_name="eval" in the hydra decorator and then change the pth name you want to run in the variable of best_model_path, then please run the following command:
$ bash scripts/ex_infer.sh
This would also calculate lpips score for each image and create directory for fid score. The results would stored under inference_results/ and name based on the experiment name.
To run the inference, you first need to set the config_name="eval_vae" in the hydra decorator and then change the pth name you want to run in the variable of best_model_path, then please run the following commands.
$ bash scripts/ex_infer_vae.sh
This would also calculate lpips score for each image and create directory for fid score. The results would stored under inference_results/ and name based on the experiment name
Download weights
weights
