gluonts VS flow-forecast

https://github.com/awslabs/gluonts is a great candidate for a sample wiki. It is an OSS lib, not great documentation, very hard to RTFM (unlike, say, sklearn which already has a great wiki), doubtful that awslabs would pay to produce.

I was looking at [`gluonts`](https://github.com/awslabs/gluonts/tree/dev/src/gluonts/core) source code and I found a `py.typed` file. That is something I always put in my type-annotated modules: it's literally an empty file which denotes that the module is marked for "internal or external use in type checking" [mypy docs](https://mypy.readthedocs.io/en/stable/installed_packages.html?highlight=py.typed#creating-pep-561-compatible-packages). However, I never saw before the `.typesafe` file. What does it denote? Does it have to be used alongside a `py.typed`?

-GluonTS

-Flow

How does it compare to Flow Forecast? Honestly people rarely mention FF but I've found it much better than pytorch_forecasting and the like.

PyTorch forecasting has a lot of bugs in it. You should try posting on issue on the actual repository though. Also, I've found Flow Forecast to be an all around much better deep learning for time series forecasting/classification framework.

If you want to use deep learning then Flow Forecast is the best. Many of the latest deep learning models and easy hyper-parameter sweeps.

I haven't personally tried it, but here's a Github Repo called LIME for Time. I'm not sure about the state of attention visualization for timeseries but this repo has several models using attention.

Hey everyone, just released a new Flow Forecast a deep learning for time series forecasting framework written in PyTorch. For this new version we added models, fixed several annoying bugs, and created better error messages. See the improved framework and tutorials Link

This is actually a good case for deep learning techniques that create a learned time series embedding id and/or graph convolutions. The advantage of these methods is the can [learn spatial temporal dependencies across several time series](Cross-Dimensional Self-Attention for Multivariate, Geo-tagged Time Series Imputation). There are some implementations that you can find of these models that you can find in the repository that deep learning for time series repo I maintain.

I mean a lot of deep learning models are more interpretable than you would think. There are a lot of methods to explain model predictions. In the deep learning for time series forecasting framework that I help maintain for example we automatically create SHAP plots to show relevant features. You can also visualize the attention mechanism directly.