Test code for the Emotional Mario challenge of Media Eval 2020
Toadstool is a dataset of humans playing Super Mario Bros. The process_toadstool.py script can be used to replay the action from the dataset and generate frames which agents can be trained with. For example the following command will generate a dataset of frames from all runs in Toadstool. (This may take some time to execute)
python process_toadstool.py -i toadstool/toadstool/participants/ -o toadstool/processed/For processing an individual json file of actions, the following format can be followed. Not that the output path must end in participant_x with x being the participat number.
python process_toadstool.py -i toadstool/toadstool/participants/participant_0/participant_0_session.json -o toadstool/processed/participant_0/Behavior cloning involves training an agent on a datset of (observation, action) tuples taken from an expert. The behavior_cloning.py script can be used to train an agent on Toadstool. For example the following command will train the agent from data stores in a given directory for 100 epochs.
python behavior_cloning.py -i toadstool/processed/ -e 100