测试训练>> basketball_trajectory_prediction>> 返回
项目作者： zhaoyu611

项目描述 ：
This repo is an open source of paper : Applying bidirectional LSTM and Mixture Density Network for Basketball Trajectory Prediction.
高级语言： Python
项目地址: git://github.com/zhaoyu611/basketball_trajectory_prediction.git


basketball_trajectory_prediction

This repo is an open source of paper : Applying bidirectional LSTM and Mixture Density Network for Basketball Trajectory Prediction.
I strongly recommend you to review Rajiv and Rob’s repo at first. the URL is https://github.com/RobRomijnders/RNN_basketball.
I think they made cool job and details about basketball prediction. Also you can find their paper and referrences in the repo.
Based on their contribution, I set up a new repo, which proposed Bidirectional LSTM and Mixture Density Network (BLSTM-MDN) for the same prediction problem.
I did 2 jobs in the main, Hit or miss classification and trajecotry generating.
In the first job, users can choose one of models, including CNN, LSTM, BLSTM, LSTM-MDN and BLSTM-MDN. And trajectory genarating only works for LSTM-MDN and BLSTM-MDN.

Setup

  • TesnsorFlow 1.0
  • sklearn
  • hyperopt

The files

  • data: the original data is in ‘seq_all.csv.tar.gz’, and the ‘seq_all.csv’ is the unziped dataset.
  • plot_staff: the scripts and final figures based on the models
  • dataloader.py: data pre-process
  • model.py: build model by TensorFlow
  • util_MDN: utility functions for building model
  • sample.py: functions used for generating trajectory
  • main.py: main steps for classification and generating

Run

Simply run file “main.py” in terminal with default argpases: python main.py
Here is the explanation of each argpase.

  2.   paser.add_argument("--hidden_layers", type=int,
  3.                      default=2, help="number of hidden layer ")
  4.   paser.add_argument("--seq_len", type=int, default=12,
  5.                      help="sequence length")
  6.   paser.add_argument("--dist", type=float, default=5.0,
  7.                      help="distance from point to center")
  8.   paser.add_argument("--hidden_size", type=int, default=64,
  9.                      help="units num in each hidden layer")
  10.   paser.add_argument("--drop_out", type=float, default=0.7,
  11.                      help="drop out probability")
  12.   paser.add_argument('--learning_rate', type=float, default=0.005,
  13.                      help="learning_rate")
  14.   paser.add_argument('--epoch', type=int, default=1,
  15.                      help="epoch")
  16.   paser.add_argument('--batch_size', type=int, default=64,
  17.                      help="batch size")
  18.   paser.add_argument('--model_type', type=str, default='BLSTM_MDN_model',
  19.                      help='the model type should be LSTM_model, \
  20.                        bidir_LSTM_model, CNN_model, Conv_LSTM_model, \
  21.                        LSTM_MDN_model or BLSTM_MDN_model.')

If you want to generate some trajetories, please set “generate_trajectory” as True in code. Because it is False in default.
It should be noted that it only generates traejctory with BLSTM-MDN or LSTM-MDN.

Contact me

Be free the ust the code for studying. But please contact me if you want for commercial applying.

You are welcome to pull requests or issues.

E-mail: zhaoyuafeu@gmail.com

Facebook: zhaoyuafeu