How to use Neural ODE in Time Forecasting?

This is a repo for Neural ODE and CDE forecasters.

Installation and Data Preparation

You should install these necessary packages:

• matplotlib>=3.5.1
• numpy>=1.22.2
• pandas>=1.4.1
• torch>=1.10.2
• torchcde>=0.2.5
• torchdiffeq>=0.2.2
• tqdm>=4.63.0

Then place your own data files in data/. Please change models/utils.py/get_data function to get dicts for train, val (if optional else None) and test. These dicts must be formulated by Pytorch tensors as:

"x_time": torch.Size([all_samples, x_time_points, 1])  # Timestamps
"x_data": torch.Size([all_samples, x_time_points, x_dims])  # input data
"x_mask": torch.Size([all_samples, x_time_points, x_dims])  # input mask. Bool tensor.
"y_time": torch.Size([all_samples, y_time_points, 1])  # Timestamps
"y_data": torch.Size([all_samples, y_time_points, x_dims])  # output data
"y_mask": torch.Size([all_samples, y_time_points, x_dims])  # output mask. Bool tensor.

Run commands to train & test

I've implemented a function models/utils.py/argschange to change the Args dict in run_models.py (A simple alternative to argparse package!), so you can run commands like:

python run_models.py --arch Recurrent --using CDE --no_save

Or simply click and run run_models.py. This way makes sense because the arguments mentioned above have default values in Args.

To use different models, you can set:

--arch: "Recurrent", "Seq2Seq" and "VAE" are available.
--using: "ODE_RNN" or "CDE". "CDE" only for "Recurrent" now.

--no_save argument prevents the model to create experiment results files. Without this, visualization results can be found in fig/, result logs in log/ and checkpoint files in ckpt/.

delete.py is used to clear files above. You can type:

python delete.py --ID 999999

In order to clear files for experiment ID 999999, for instance.