Osu Neural Network Created Using Pytorch

• DISCLAIMER : I am not responsible for any consequences that stem from the illicit use of the contents of this repository.

Info

• Data extracted from replays produced by a slightly modified version of danser-go

• Aiming works but clicking does not.

• All skins used are located in assets/skins

• Showcase videos Clicking (DOES NOT WORK CURRENTLY) and Aiming.

Quick Start

• The following assume that your monitor/screen size is 1920x1080 (plan to make it dynamic later)
• Clone this repository along with this modified version of danser
• Build danser using the instructions in the repository
• Copy danser-settings.json to "cloned danser repo"/settings/danser-settings.json
• Launch danser using the built binary
• Once danser is done importing in the config dropdown switch it to the settings we copied earlier
• Switch mode to "Watch a Replay" and select the replay you want to train on
• Switch the dropdown on the bottom from "Watch" to "Record"
• Before recording you can configure the settings to use a different skin.
• Once ready click "danse!" and wait for danser to generate the recording.
• Once done it will open a folder with a json file and an mkv file, these are needed to generate the dataset.
• Going back to the osu-ai folder first setup Anaconda
• Run the following in the terminal to create an enviroment

conda create --name osu-ai python=3.9.12
conda activate osu-ai

• Install Poetry
• Run the following in the enviroment we created

poetry install
# For cuda support run "poe force-cuda"
# For win32Mouse support run "poe use-win32"

• now we can run main

python main.py

• You should see this menu

What would you like to do ?
    [0] Train or finetune a model
    [1] Convert a video and json into a dataset
    [2] Test a model
    [3] Quit

• Select "1" for Convert, name it whatever you want. For the video and json type in the path to the respective files we generated earlier i.e. a/b/c.mkv a/b/c.json.
• For the number of threads I usually use 5 and we will leave the offset at 0.
• For Max images to keep in memory when writing I usually leave it at 0 unless the video is really long
• Now wait for the dataset to be generated
• Now we can train. Select "0" to train then "0" for aim. Name it whatever and select the dataset we just made. I usually set max epochs to a very large number since ctrl+c will stop training early.
• And that's it.

Example Autopilot play below on this map. The model was trained using this map.

Example Relax play below on this map. The model was trained using this map.