Beginner Classifier

What is this?

This is a program that uses a deep neural net in order to classify players as either beginner or advanced in TrackMania based on statistical data gathered by the server controll EvoSC (https://github.com/EvoTM/EvoSC).

Installation & Setup

• pip install -r requirements.txt
• [Optional] If you want logging to graylog, you must install graypy: pip install graypy
• Rename conf.cfg.example to conf.cfg and modify the settings to your requirements
• python main.py -h for usage information

Usage

• Start the server python main.py server
  • If you get a bunch of errors from tensorflow about the gpu, you can ignore them.
  • If you don't use --detach and the server just closes immediately, an error occured. Check the log.
• Test the classifier using the server: python .\main.py classify --server --logins snixtho
• For json output, use --json: python .\main.py classify --server --logins snixtho --json
  • This will return a list of predictions in the array predictions. Each prediction contains a property sucess which is true on success, and false if an error occured. If an error occured, the property error contains details about what happened. Each prediction also contains the login requested as well as the predictions experienced and beginner. Their sum should be exactly 1, so the predicted class is the one with a higher value. The number itself is a indication about how sure the classifier is about it's prediction.
• To request multiple logins at the same time, just separate them by space: python .\main.py classify --server --logins snixtho brakerb tmexperte

Training

You can train your own classifier model using the python main.py model and python main.py dataset commands. Use the -h for usage of these commands.

The general procedure for creating a classifier model is:

1. Build the dataset with python main.py dataset or create a CSV file containing the data points you would like to use.
2. Run the python main.py model to train a model using the dataset. You can specify training and model options including layers. The program uses Keras to build the model.

The default training and model options are:

• --batch-size (training batch size): 100
• --epochs (num training epochs): 32
• --out-layer-activation: softmax
• --inner-layers: 50:sigmoid
• --dt-values (features): finishes,locals,wins,score,rank
• --in-layer-size (input layer): 32
• --in-layer-activation: relu
• --optimizer: rmsprop
• --loss (loss function): binary_crossentropy
• --metrics: accuracy
• --out: model.h5

(The models already made does not use these options, they also use more layers)

Notes

• tensorflow 2.1.0 which is required, uses the AVX instruction set by default. If you get the error "Illegal Instruction", you can either configure your VM to use SandyBridge (if run in a VM) or build tensorflow from source.
• Coded and tested with python 3.6.