What is the relationship between sktime and pytorch-forcasting?

[travis-leith]

Hi! I am new to both sktime and pytorch-forecasting. I am about to start a new project to analyse some time series data. I want to be able to fit a LSTM or similar RNN to some time series data, consisting of various individual time series of different lengths.

It is not clear to me if I should be using sktime or if I should be using pytorch-forecasting. Is one of these being absorbed by the other? There seems to be a lot of overlap. How to choose which to use?

[penknife6153]

For most ecosystems, having specialized tools that interoperate well tends to be more sustainable than trying to build one tool that does everything.

sktime does support pytorch-forecasting via wrappers if you want to mix neural networks with classical methods later. But since you're specifically targeting LSTM/RNN approaches, using pytorch-forecasting directly will give you the cleanest experience - optimized for GPU training with minimal dependencies.

[fkiraly]

exactly - pytorch-forecasting specializes in torch based forecasters.

sktime specializes in providing a composition and unified API layer for forecasters and time series models in general.

From an architectural perspective, it is as @penknife6153 says - "trying to do one thing really well".

Where the scope of pytorch-forecasting is wider than LSTM/RNN, it is "anything torch", ultimately including, hopefully, the latest generation of foundation models just as well as previous generation attention based models, and early generation LSTM/RNN based approaches.

To do this, a rework with dsipts is underway, towards version 2: #1736

[fkiraly]

How to choose which to use?

Rule-of-thumb: if you want to use only torch based forecasting models, and do nothing else, pytorch-forecasting is probably your right entry point. Same if you want to are looking for these as components in a framework that you are building yourself.

If you want other models too (e.g., naive models, statistical models, ML based models, etc), pipelines, tuning, meta-models, autoML, benchmark the models against each other, you should probably use sktime, while making use of the interfaces to pytorch-forecasting to include the LSTM type models you want.