howardzorn/st-man

ST-MAN: Spatio-Temporal Multimodal Attention Network for Traffic Prediction (KSEM 2023)

ST-MAN: Spatio-Temporal Multimodal Attention Networks for Traffic Prediction

How to Run this Model

Data

PeMS04 and PeMS08 are provided by ASTGCN. Seattle Loop Detector Dataset is provided by Cui et al..

Our dataset files can be downloaded from Google Drive.

The Datasets files should be placed in the data folder.

Requirements

Requirements should be installed before any operation.

pip install -r requirements.txt

Manual

Run train.py to train the model, run test.py to do the test.

$ python train.py PeMS04
$ python test.py PeMS08

You should specify a dataset before trainning and testing.

$ python train.py --help
usage: train.py [-h] {PeMS04,PeMS08,Loop}
train.py: error: the following arguments are required: dataset

You can modify the settings and hyperparameters via commandline arguments, they must be placed after the dataset argument.

$ python train.py PeMS04 --help
usage: train.py [-h] [--time_slot TIME_SLOT] [--P P] [--Q Q] [--N N] [--L L]
                [--K K] [--d D] [--seed SEED] [--train_ratio TRAIN_RATIO]
                [--val_ratio VAL_RATIO] [--test_ratio TEST_RATIO]
                [--batch_size BATCH_SIZE] [--max_epoch MAX_EPOCH]
                [--patience PATIENCE] [--learning_rate LEARNING_RATE]
                [--decay_rate DECAY_RATE] [--traffic_file TRAFFIC_FILE]
                [--SE_file SE_FILE] [--model_file MODEL_FILE]
                [--log_file LOG_FILE] [--gpu_device GPU_DEVICE]
                [--drop_rate DROP_RATE] [--masked_l1 MASKED_L1]
                {PeMS08,PeMS04,Loop}

positional arguments:
  {PeMS08,PeMS04,Loop}  use a dataset

optional arguments:
  -h, --help            show this help message and exit
  --time_slot TIME_SLOT
                        a time step is 5 mins
  --P P                 history steps
  --Q Q                 prediction steps
  --N N                 number of Cross Att Blocks
  --L L                 number of STAtt Blocks
  --K K                 number of attention heads
  --d D                 dims of each head attention outputs
  --seed SEED           seed of random utils
  --train_ratio TRAIN_RATIO
                        training set [default : 0.7]
  --val_ratio VAL_RATIO
                        validation set [default : 0.1]
  --test_ratio TEST_RATIO
                        testing set [default : 0.2]
  --batch_size BATCH_SIZE
                        batch size
  --max_epoch MAX_EPOCH
                        epoch to run
  --patience PATIENCE   patience for early stop
  --learning_rate LEARNING_RATE
                        initial learning rate
  --decay_rate DECAY_RATE
                        decay rate
  --traffic_file TRAFFIC_FILE
                        traffic file
  --SE_file SE_FILE     spatial emebdding file
  --model_file MODEL_FILE
                        save the model to disk
  --log_file LOG_FILE   log file
  --gpu_device GPU_DEVICE
                        train device
  --drop_rate DROP_RATE
                        drop rate
  --masked_l1 MASKED_L1
                        whether use masked l1 loss