Setup Steps:
1) Clone: https://github.com/nicklashansen/dmcontrol-generalization-benchmark.git and follow setup instructions
This includes DMC-GB environment as well as previous methods
2) Replace train.py, utils.py and arguments.py in src/ with versions in zip
3) Replace factory.py in src/algorithms with one in zip
4) Add non_naive_rad.py and augcl.py to algorithms/

Training AugCL:
1) python src/train.py --id shift --algorithm non_naive_rad --data_aug shift --train_steps 200k --save_buffer True
This represents the weak augmentation phase
2) python src/train.pyu --algorithm augcl --data_aug splice(or CS_splice if task is cartpole swingup) --prev_id shift --prev_algorithm non_naive_rad --curriculum_step 200000 --eval_mode color_hard
This represents the strong augmenation phase
At the end of train environment and color hard performance over 30 episodes will be printed to console and log. (The last log or print)

Instructions to download COCO from:
https://cocodataset.org/#home