The root directory for executing programs should always be /csgo_encounter_prediction/ Otherwise issues with paths and directories may arise.

The datataset has to consist of recorded CS:GO matches in form of .dem files.

Inside the dataset_config.json file you can find multiple variables related to preparing the dataset:

• demo_files_path: Location of downloaded and unpacked .dem files of CS:GO matches
• parsed_files_path: Location of parsed .dem files stored as .csv
• processed_files_path: Location of .feather files that have undergone preprocessing and whose data is ready for training
• training_files_path: Location of randomized chunks of data that are actually used for training
• death_time_window: Time window for classification labels
• chunk_row_size: Row size of randomized chunks.

All paths are relative to the root directory. All paths may point to the same location, but for visibility purposes different paths are recommended. After this setup the dataset preparation can begin.

After setting up the configuration,execute the preparation/prepare_dataset.py with the -mode all argument for complete parsing, preprocessing and randomization of the .dem files.

By using the -config PATH option, the script will use the specified config. Otherwise it will use the default config config/dataset_config.json

Verbosity of logging is specified by the option -v in stages up to -vvvv

If the -override option is left out, no files are replaced. Othweise all files are being replaced during the process.

For separate execution of the steps use the argument option -mode of prepration/prepare_dataset.py:

• parse: Run to parse all .dem files specified in config. Output is .csv files of matches
• preprocess: Add and remove all neccesary features of data for training. Output is .feather files of matches
• randomize: Randomize all matches and generate chunks as .feather files. These can be used for training.

Example:

python preparation/prepare_dataset.py -mode preprocess -config ./config/special.json -vv -override

With py prepare_dataset.py -h a help page is be displayed.

Hyperparameter search is perfomed by running the script training/hyperparameter_search.py. Runs a certain amount of named training runs. Generates configs for each run and saves them to config/<NAME_OF_RUN>. The results are also saved to tensorboard The parameters of the hyperparameter search have to be adjusted inside the script.

As with the dataset preparation train_config.json is the configuration for training the network.

• shared_layers, dense_layers: Arrays that contain sizes of shared and dense layers
• feature_set, label_set: Feature and label sets for training (seen in features_info.json
• lr: Learning rate of network
• batch_size: Amount of samples in one batch
• num_epoch: Amount of epochs
• checkpoint_epoch: Interval at which to save state of model

Network training is done by executing the training/csgotest.py script By deault the config/train_config.json config is used. In addition a name can be specified for the naming run.

For example:

python train.py -trainconf config/train_config.json -name test

Trained models are saved to ./models/<NAME_OF_RUN>

For a description of the different parameters enter:

python train.py -h

All training runs are recorded and can be viewed in real time in tensorboard.

If installed type:

tensorboard --logdir=results/

to start tensorboard.