Dynamic Seq2Seq for Trajectory Understanding

Using Seq2Seq to assign actions to trajectories.

Getting Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. See deployment for notes on how to deploy the project on a live system.

Prerequisites

For this repo you need:

• tensorflow -v 1.4.0
• Numpy -v 1.14.2
• Scikit-learn -v 0.19.1
• argparse -v 1.1

pip install tensorflow-gpu=1.4.0
pip install numpy, scikit-learn, argparse

Installing

A step by step series of examples that tell you how to get a development env running.

Download the repository.

git clone git@github.com:yadrimz/autoencoding-trajectories.git

You can train and infer. Run help for full documentation.

python main.py -h

To train:

python main.py --mode 'train'

Sample output:

Batch: 1
  minibatch_loss: 1.0833992958068848
  accuracy: 0.2222222222222222
   sample: 1
      sequence real id         :> 8
      target start             :> [4 3 5]
      predicted start          :> [3 3 1]
      target end region        :> [3 5]
      predicted end region     :> [3 1]
   sample: 2
      sequence real id         :> 7
      target start             :> [3 2 5]
      predicted start          :> [6 5 5]
      target end region        :> [2 5]
      predicted end region     :> [5 5]
   sample: 3
      sequence real id         :> 14
      target start             :> [4 1 5]
      predicted start          :> [5 3 4]
      target end region        :> [1 5]
      predicted end region     :> [3 4]

To infer after having trained a model:

python main.py --mode 'infer'

Project Structure

Project File Structure:
-data
---timeseries_25_May_2018_18_51_49-walk3.npy
-models
---model.py
-utils
---utils.py
-main.py
-README.md

Authors

• Todor Davchev

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgments

• Tensorflow's NMT Tutorial
• Ematviev's GitHub repo