Run python train.py --task cls to train the classification model. Evaluate it by running python eval_cls.py --load_checkpoint=model_epoch_240 (since it saves up to epoch 240 as it is). The outputted gifs should appear in the output/ directory. Currently, it runs with a batch size of 32, and saves one gif per batch. For different gifs per batch, offset the batch by some constant between 0 and 31 when setting i in eval.cls.py.

Run python eval.py --task seg to train the segmentation model. Evaluate it by running python eval_seg.py --load_checkpoint_mode=model_epoch_240 --i={i}, where i is in the set {382, 224, 608, 311, 386} (the ones I tested). The outputted gifs should appear in the output/ directory.

To run the experiments testing model robustness, uncomment the necessary parts of the code in eval_cls.py and eval_seg.py.