Shogi Camera is an experimental project which aims to extract shogi board information from pictures of shogi games.

• Python 3.5
• Keras 2
• Tensorflow 1.1.0
• scipy, numpy, scikit-learn
• hdf5, h5py

• Docker Engine

fits for AWS p2 instance

• GPU (NVIDIA Tesla K80)
• CUDA Toolkit / CUDA Driver
• Docker Engine
• nvidia-docker

$ python3 cli.py predict {image_file_path}

or on a Docker container,

$ docker run -it -v $PWD:/app naoys/shogi-camera:nogpu python3 cli.py predict {image_file_path}

$ docker run --runtime=nvidia -it -v $PWD:/app naoys/shogi-camera:latest python3 cli.py predict {image_file_path}

The prediction consists of two phases.

1. detecting a shogi board precisely from the picture
2. predicting single cell information using a pre-trained NN model and repeating it cell by cell

see the notebook or a program

https://github.com/na-o-ys/shogi-camera/blob/master/shogicam/predict/_predict_board.py

The simple model contains three convolutional layers and two fully connected layers.

See the notebook or a program.