it's becoming more time-consuming and error-prone to manually re-test all the demos following internal refactorings and API adjustments.

now that the API is fleshed out a bit, it's possible to test a large amount of code (non-granularly) without having to simulate all interactions via Puppeteer or similar.

a lot of code can already be regression-tested by simply running all the demos and val

Hi, Thanks for the awesome library!

So I am running a Kmeans on lots of different datasets, which all have roughly four shapes, so I initialize with those shapes and it works well, except for just a few times. There are a few datasets that look different enough that I end up with empty clusters and the algorithm just hangs ("Resumed because of empty cluster" again and again).

I conceptually

similar to scikit model.get_params

In many time series datasets, the sequences are of variable length. Now TSAI dataloader expects fixed length time sequences. The most natural solution is to transcribe each time sequence to the maximum length in our dataset and pad zeros (or -1 or any other token) in the extra positions. In vanilla PyTorch implementation of Transformer, there is an src_key_padding argument given to forward