CoachAI
Tensorflow implementation of badminton serving machine parameter predictions using AI models.
Paper
Training a Group of Badminton Serving Machines to Reproduce a Rally (Work in Progress)
Yu-Fu Wu, Huang-Yi Cheng, Yun-Tang Lin, Jen-Jee Chen, Ting-Hui Chiang, Yu-Chee Tseng
ICPAI 2020 (IEEE pdf)
Requirements
This code was tested with Tensorflow 1.13.1, CUDA 10.0 and Ubuntu 16.04.
Packages installation:
pip install pydot==1.2.3
apt install graphviz
pip install piexif
pip install similaritymeasuresDataset
You can download datasets from here (permission needed).
Training
We provide several models for training. You can choose one of the following models for training.
- cnn_regression_rmse_image.ipynb
- cnn_regression_rmse_trajectory.ipynb
- dense_regression_rmse_image.ipynb
- dense_regression_rmse_trajectory.ipynb
- lstm_regression_rmse.ipynb
You can monitor the learning process using tensorboard and pointing it to your chosen logs.
Testing
We provide a single model (LSTM model). You can try it by runnung lstm_regression_rmse.ipynb and uncomment by following:
# Load model
model_dir = "models/20201216_lstm-rmse-position5-trajectory3-all-diff-euclidean-32-8-batchsize128-relu-epoch1000/"
model_name = "20201216_lstm-rmse-position5-trajectory3-all-diff-euclidean-32-8-batchsize128-relu-epoch1000.h5"
model_path = os.path.join(model_dir, model_name)
model = keras.models.load_model(model_path,
custom_objects={'root_mean_squared_error': root_mean_squared_error})
# eval_his = model.evaluate_generator(test_generator(), steps=len(test_list), verbose=1)