This is PyTorch implementation of MSTFCN in the following paper:

Multi-mode Spatial-Temporal Data Modeling with Fully Connected Networks. KSEM, 2024.

All datasets have been allocated in ./data/h5data/ folder.

1. Change model settings in ./model/{MODEL}/config.json.

2. Change running settings in ./config.json.

   In data field, nyc-mix, chicago-mix and beijing-mix are available.

   Your can change expid field to name the experiments.

3. Run train_multi_step_mix.py!

4. Results will be saved in ./saves/{DATASET}/{MODEL}/{expid} folder.

If you want to only inference trained models, run test_multi_step_mix.py.

If you want to add more models:

1. Create a new directory in ./model/, place you MYMODEL.py and create a config.json file to record your hyper-parameters.

2. Register your model in ./model/__init__.py and ./util.py.

   • If your model requires auxiliary information such as predefined graph, write it in get_auxiliary() function.
   • Init your model in get_model() function.

3. A little changes on the forward() function of your model:

   • The forward function of your model must be forward(input, **kwargs).
   • The input has shape (B, T, N, C + 2). The value of C is 4, with the first 2 entries refer to features of modality 1, and the last 2 entries refer to features of modality 2.
   • The remaining 2 refers to time semantic information (B, T, N, 2).
   • The output must have shape (B, T, N, C), without time semantic.

4. Run!