Toast

Python implementation of Toast: Task-oriented Augmentation for Spatio-Temporal Data.

Dependencies

pip install -r requirements.txt

Another dependency traj-dist is required, and it can be install from this link .

Datasets

The dataset T-drive can be downloaded from this repo .
The dataset Porto can be downloaded from this repo .
The dataset Geolife can be downloaded from this link .
The dataset TaxiBJ can be downloaded from this link and its data format can be found from this repo .

Run code

The scripts are shown in scripts/.
E.g., to run experiments on trajectory recovery, the script is

usage: run_recovery_task.py [--dataset DATASET] [--model_name MODEL] [--num_epochs EPOCHS] [--batch_size SIZE] [--phase PHASE] [--gpu GPU_ID] [--seed SEED] [--num_virtual_tokens TOKEN_NUM] [--num_augment_epochs AUG_EPOCHS]

optional arguements:
--dataset               dataset name
--model_name            downstream model name
--num_epochs            number of epochs to train downstream model (following the original paper)
--batch_size            batch size for model training
--phase                 training / test / augment phase (for init. training, evaluation, and augmentation)
--gpu                   gpu id
--seed                  seed
--num_virtual_tokens    number of virtual tokens
--num_augment_epochs    number of epochs for data augmentation
--mixup                 store_true type (whether to use mixup augmentation) 

Acknowledgement

The code used for downstream tasks is adapted from their original github repos.