Add the current directory to the search path via

export PYTHONPATH=$(pwd)

install requirements via

pip install -r requirements.txt

Automatic code formatting via

autopep8 -r --in-place .

Extract the table region via

python tools/add_table_frames_to_h5py.py -i filename

Add ball positions

python tools/neural_net_detection.py -f tools/frozen_inference_graph.pb -i train/games.h5

Filter out the good games. This scripts weeds out games, where here are few distinct ball positions detected.

python tools/filter_good_games.py -i $INPUT_FILE -o OUTPUT_FILE

Then one can proceed to score the game with

python tools/score_training_data.py -i $INPUT_FILE

This process produces then a file train/training_data.h5 which contains the necessary information to train a neural net.

Training can be corun for 1000 steps via the

python tools/train_1000.py

This updates model.h5 with 1000 training iterations.

Running

python tools/analyze_model.py model.h5

generates graphs displaying the distribution of the weights. These should approach smooth curvers during training.

With

python tools/inspect_data.py

one can visualize what the neural net sees. One sees the three subsequent frames from the collected data set, and the actions that the neural net would perform in the form

[gr, gl, dr, dl, cr, cl, ar, al]

where g = goal, d = defense, c = center, a = attack and r = rotation, l = lateral. 0 corresponds to no movement, 1 to forward, -1 to backward.

The tool

python tools/view_game_file.py

shows a specific game. Useful for inspecting the progress of neural net training.

After training via python tools/train_1000.py, one can use tensorboard to visualize the training via

tensorboard --logdir=tensorboard_logdir