LeagueMinimapDetectionCNN

Source code to my blog post

Setup

Windows:

python3 -m venv env
env\Scripts\activate
pip install -r requirements.txt

Linux/Mac

python3 -m venv env
source env/Scripts/activate
pip install -r requirements.txt

Reproducing blog material

Running to generate image layers

• Run show_generated_image_layers.ipynb in a python notebook

• This will generate a file layered_minimap.gif and produce timing results for generation and training

Evaluation

• Contains the evaluation code in the blog

• Go to evaluation folder and follow the README.md folder to view results from the blog post/reproduce results.

Model

• Model trained for 8 hours is in trained_models\fastrcnn_model_8hrs.pt

Training

League version

• Assumes patch is 10.7 -- that there were 149 champions in this patch

Updates (optional)

• Update the folders in league_icons to match the current version

• Most likely you need to updated the number of champions (add the missing champion icons under league_icons/champion)

Running

• Run train.ipynb as a python notebook

• Or

• Run with colab

Architecture

• Runs multiple threads to generate data that is forwarded to the model to train on

Production

Run ingame

• This will load the already trained model shown in the blog post. Change the path to your trained model if desired.

• python pytorch_test_ingame.py trained_models\fastrcnn_model_8hrs.pt 149

• Only tested with the game set as borderless and with a 1920x1080 resolution

• This will run with external window displaying the detected champions like this

Run ingame with gif output

• Outputs test.gif that runs for 1000 frames, which is shown above

python pytorch_test_ingame.py trained_models\fastrcnn_model_8hrs.pt 149 test.gif