https://www.youtube.com/watch?v=pOnKsGiTVuk&ab_channel=FourgeesGaming

Python / FastAI CNN - Playing any racing game

This code was used to gather and process data while playing any racing games or side scroller. Currently working with Diddy Kong Racing. The network was then trained using the FastAI Libary. The original scope of this project was to fork the code from windows to linux.

Setup

• I used OBS virtual cam (using loopback driver for linux) and with a game window capture playing at a 800 X 600 resolution fed into open cv2 as a webcam input.
• You can start data collection by running CreateData.py. This code was tested on Ubuntu 20.04.
• Pressing esc will start the screen / key grab. These will be stored in lists until the episode is done. Press h to end and save.
• Once the episode ( Round ) ends pressing h will stop the screen / key grab process and all data will be moved to a numpy array.
• Then I used a script in util folder called CreateImages.py to put then onto a disk drive in folders corresponding to their actions.

Train

• Use the file called training.py or use the google colab that is provided here. (It's faster and easier)
• Point it at your image directory

Run Agents

• Fully random agent is RandomAgent.py
• Trained Agent is TrainedAgent.py
• You will have to load in the pkl created from training.

Inputs (Observations)

• Uses inputs to the nural network (Observations) of pixels in the game.
• 224 X 224
• Line detection (might change it to just b&w, testing this currently)

Contact!

• Plz don't