- Step1: Download Porto data (i.e., train.csv.zip) from https://www.kaggle.com/c/pkdd-15-predict-taxi-service-trajectory-i/data.
- Step2: Put the data file in ./data/, and unzip it as porto.csv.
- Step3: Set the dataset info in ./preprocess/preprocess.py), run preprocessing by sh preprocess.sh.
The processed data files (i.e., processed_porto_train.csv and processed_porto_val.csv) will be put in ./data.
python run_loop.py --mode=train --cluster_num=5 --num_epochs=5 --gpu_id=0 \
--model_dir=./ckpt --learning_rate=1e-4 --num_epochs=10 --pretrain_dir=./pretrainMore conveniently, we can run pretraining, training and evaluation via pretrain.sh, train.sh and eval.sh, respectively.
| Name | Type | Description |
|---|---|---|
| mode | enum(str) | pretrain, train or evaluate. |
| data_filename | str | data file (e.g., ./data/processed_porto.csv). |
| map_size | (int, int) | size of the grid map. |
| token_dim | int | dimensionality of grid token. |
| rnn_dim | int | dimensionality of rnn hidden state. |
| cluster_num | int | number of Gaussian components. |
| model_dir | str | directory to save/load a model during training or eval. |
| pretrain_dir | str | directory to save/load a model during pretraining. |
| num_negs | int | number of negative samples during training. |
| optimizer | enum(str) | training optimizer (e.g., adam or sgd). |
| learning_rate | float | learning rate for training. |
| num_epochs | int | number of passes over the training data. |
| log_steps | int | number of batches to print the log info. |
Please kindly cite the paper if this repo is helpful :)
@inproceedings{liu2020online,
title={Online anomalous trajectory detection with deep generative sequence modeling},
author={Liu, Yiding and Zhao, Kaiqi and Cong, Gao and Bao, Zhifeng},
booktitle={2020 IEEE 36th International Conference on Data Engineering (ICDE)},
pages={949--960},
year={2020},
organization={IEEE}
}