Reconstructing Trajectories with Transformer Variational AutoEncoder

TimeLine

Dataset process: Geolife and Porto (done)
NVAE model (done)
baseline model:
- EDR (done)
- EDwP (done)
- t2vec (done)
- TrjSR
- E2DTC
- TrajCL
metric:
- cityemd (done)
- NMD (done)
- NMA (done)
- RRNSA (done)

Installation

Python 3.9

conda env create -f environment.yml   # Create environment
conda activate nvib                   # Activate environment

NB This environment also installs the nvib package which is assumed to be in a directory one level above the current directory.

Original datasets can be downloaded from: Geolife and Porto. The preprocessed dataset can be downloaded from here

Quick Start

To train NVAE or baseline models, and encode latent vectors:

python Train.py --dataset porto --model model_name

To retrieve reconstructed trajectories:

python retrieve.py --dataset porto --model model_name

To test reconstructed trajectories with cityemd or Mobility Tableau metrics:

python cityemd.py --dataset porto --model model_name
python evaluate_yao.py --dataset porto --model model_name

Datasets

To use your own datasets, you may need to create your own pre-processing script like ./utils/preprocessing_porto.py. Also, the configuration is required to fill into dataset_config.py. (See ./utils/preprocessing_porto.py and dataset_config.py for more details.)

Baselines

To use your own baselines, you may need to create your own baseline script like ./baseline/AE.py. Also, the configuration is required to fill into model_config.py. (See ./baseline/AE.py and model_config.py for more details.)

Repository Structure

├── baseline
    ├── AE.py
    ├── t2vec.py
    ├── transformer.py
    └── VAE.py
├── data
    └── .gitkeep
├── exp
    └── .gitkeep
├── model
    ├── NVAE.py
    ├── NVAE_trainer.py
├── utils
    ├── dataloader.py
    ├── preprocessing_geolife.py
    ├── preprocessing_porto.py
    ├── seq2multi_row.py
    └── tool_funcs.py
├── .gitignore
├── cityemd.py
├── dataset_config.py
├── environment.yml
├── evaluate_yao.py
├── model_config.py
├── README.md
├── retrieve.py
└── Train.py