Deep Gamer

Requirements

Windows OS
Python 3.5+
See requirements.txt

Installation

Install the python requirements: pip3 install -r requirements.txt.

Development

To create your custom network, just extend the src/networks/abstract.py class and set it to that class inside src/deep_gamer.py (network = MyAwesomeNetwork()). As an example, check out the WASDNetwork (src/networks/wasd.py).

Existing networks

WASD Network

The WASD Network is primary meant for driving games. The network already captures, preprocesses, trains and evaluates the controls for driving games. It has width (640) x height (360) x RGB channels (3) inputs, and runs on the Inception V3 network and has 9 one-hot outputs (forward, backward, left, right, forward+left, forward+right, backward+left, backward+right & none).

Workflow

Note: For now, you can leave --activity="{ACTIVITY}" --mode="{MODE}" --game="{GAME} out, as the default value for --activity is general, for --mode it's default and for --game="{GAME} it's game. The reason for {ACTIVITY} is, that you can capture different activities in your game, such as driving, walking, running, .... The mode on the other hand, is ONLY relevant for preprocessing, training & capturing. For example with activity, you'll always have the same raw data, if you select a mode, then you can preprocess, train & capture different modes, like: try different images sizes (large_sizes_input, small_sizes_input, ...), different color manipulations (black_and_white, lower_brightness, higher_contrast, ...), ... and then later you can compare which mode works the best, without having to retrain your existing model.

Capture

First you will need to capture all the footage on which you want your "AI" to learn on. You'll do that with python src/deep_gamer.py capture. This will capture your (primary) screen and save all the screenshots into data/{ACTIVITY}/raw/{TIMESTAMP}.

Preprocess

After that we need to preprocess all the gathered data. For that, just run python src/deep_gamer.py preprocess

Train

Now that we have all our data prepared, we will need to train the model. We do that with python src/deep_gamer.py train (you can add the --force-new-model, if you don't want to continue to train your existing model). This may take a couple of hours/days. To view the training, start a separate terminal and run python -m tensorflow.tensorboard --logdir=data/{GAME}/{ACTIVITY}/network/{MODE}/logs (ex. python -m tensorflow.tensorboard --logdir=data/game/general/network/default/logs). Now you can go to http://localhost:6006.

Evaluate

Now we are finally ready to evaluate the model. We can do that with python src/deep_gamer.py evaluate.

CS Fixes

We use autopep8 for coding standard fixes. After installing it, just run autopep8 -a -r src.

Author

Borut Balazek bobalazek124@gmail.com (http://bobalazek.com)

License

Deep Gamer is licensed under the MIT license.

bobalazek / deep-gamer