Demand forecasting with Temporal Fusion Transformers

---

This repository contains a custom implementation of Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting using Lightning and pytorch_forecasting for demand forecasting on the Stallion dataset.

Technical details

Enhancements compared to the original implementation in the Google Research repo:

• capabilities added through pytorch_forecasting base model e.g. monotone constraints
• static variables can be continuous
• multiple categorical variables can be summarized with an EmbeddingBag
• variable encoder and decoder length by sample
• categorical embeddings are not transformed by variable selection network (because it is a redundant operation)
• variable dimension in variable selection network are scaled up via linear interpolation to reduce number of parameters
• non-linear variable processing in variable selection network can be shared among decoder and encoder (not shared by default)

Run locally

The dependency management system is poetry. Install poetry and run:

poetry install

to set up the environment.

Training

You can get the baseline performance in the dataset by training pytorch_forecasting's Baseline model:

cd src
poetry run python baseline.py

To train the full model and get the performance:

poetry run python train.py
poetry run python evaluate.py

Prediction & Inference

To run prediction on test data in the model, run:

poetry run python predict.py

To inference the model through an API, the api.py file sets up a simple Flask API with a '/predictions' GET endpoint.

poetry run python api.py

then you can access localhost:8501/docs to see a nice UI set up by FastAPI that allows easy inference, or use curl or any other tools to call the API directly.

References

• Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting
• Stallion dataset