Deep learning for time series forecasting framework updates (.95)

Version 0.95 of Flow Forecast includes complete support for probabilistic models, multitask forecasting, bug fixes and much more

Photo by author

I’m excited to announce that today we are releasing Flow Forecast version 0.95 (if you are unfamiliar with Flow Forecast please see the introductory article). Although we are still somewhat far away from version 1.0 there are a lot of cool new features now available. To get the latest version please run pip install flood_forecast --upgrade. You can also go to our

  • Complete support for multitask learning: Flow forecast now allows users to easily forecast multiple targets at once. For an easy to use tutorial see this notebook.
  • Log and sort on multiple metrics: Flow Forecast now allows you to easily log multiple metrics at once and compute test loss on each metric. So for instance, you could evaluate a model with both DilateLoss and MSELoss seamlessly. This should be particularly helpful to researchers who want to benchmark on new metrics on multiple datasets.
  • DilateLoss bug fixed: Previously there was a bug where the DilateLoss function was given an improperly shaped tensor.
  • Variable length test loss: You can now tell Flow Forecast in your config file the number of time steps you want computed with your test_loss .
  • Speed improvements: We have significantly sped up model execution time on the GPU and the CPU. We recommended using a higher number of workers, large batch size, and pinning memory to fully take advantage of our framework’s new speed improvements.
  • TorchScript Support: Our package now includes TorchScript functions. You can easily export trained models to TorchScript (note this will make).
  • Test coverage: Our repository is now at 70% code coverage. All of our main loss functions, training loops, and data-loaders are covered by tests. We are also going the extra step of adding reproducibility tests to make sure things like random seeds don’t mess up attempts to reproduce your research.

We have a number of new features in the works including:

  • Adding support for models in PyTorch Temporal Geometric
  • Adding the Informer model
  • Easy conversion of models to ONNX
  • Distributed hyper-parameter sweeps
  • Improvements to documentation
  • More tutorials and benchmarks

We could use your help! If you have free time to contribute please reach out. Don’t be intimidated we have lots beginner friendly issues. Plus our maintainers are more than welcome to help.

Deep learning researcher. Creator/Maintainer of @FlowTemporal