: Contains functions to send mouse clicks, set mouse position and key presses. Uses win32gui, win32ui, win32con, OpenCV. : Gets screenshot of the CSGO window, does some cropping and downsampling. Also several functions used across many scripts. : Contains key global hyperparameters and settings.
Script dm_run_agent.py runs the model in the CSGO environment. Use tools_create_stateful.py to create a 'stateful' version of the model. Run dm_train_model.py to train a model. (Optional.) Run tools_extract_metaadata.py to pull out metadata from the. hdf5 files containing screenshots and onehot targets. if the metadata from GSI and RAM disagrees for some variables, this is a sign that the data might be unreliable, if there are long periods of immobility ( vel_static), this can mean you were tracking a motionless player. The visulations and stats produced by this script can be used to clean the data - e.g. npy files from step 1 to infer actions from the metadata ('inverse dynamics model'). Use dm_record_data.py to scrape data as a spectator, or dm_record_data_me_wasd.py to record when actively playing. We briefly describe the workflow: dataset capture -> data processing -> training -> testing. 
Also presented at Offline RL workshop, NeurIPS 2021 Counter-Strike Deathmatch with Large-Scale Behavioural Cloning
0 kommentar(er)
