do you have code for multivariate classification datasets with categorical values ?

[Sandy4321]

do you have code/data for multivariate classification datasets with categorical values ?

[oguiza]

Hi @Sandy4321, We've just added a multimodal module that you can use for this purpose. It's still experimental, but I've used it in several cases and it works. You can find more details here: https://timeseriesai.github.io/tsai/models.multimodal.html

[norikazu99]

Hello, timeseriesAI is a great repo thanks for everything @oguiza. I am kind of lost on how to implement it. Should I make a ts_model and a tab_model and then use MultiInputNet to join (like in the docs under MultiInputNet)? Could you share a little example which you used? If there's a way but you don't have the code I wouldn't mind writing it and contribute it to tsai. I have a limited dataset and using cat embeddings would allow me to use all samples in a single model.

[norikazu99]

After some time studying the code base and reproducing results, it seems to work just fine. The tests you provided are very well coded and commented. Thanks for everything you did

[Sandy4321]

yes it seems to be something great inside , but not clear at all how to use it may you just share from 0 to end code example 0 read data from file 1 preprocess data 2 build model 3 predict 4 test performance

some clues https://github.com/nkafr/medium-article-tutorials/blob/main/TFT_Energy_Demand_Forecasting_Lightning_2_0.ipynb and https://towardsdatascience.com/temporal-fusion-transformer-time-series-forecasting-with-deep-learning-complete-tutorial-d32c1e51cd91

[norikazu99]

I'd recommend you check https://timeseriesai.github.io/tsai/tutorials.html. You'll find everything you asked for in the notebooks, if not you'll find them in the docs. There are multiple ways to do each one of them, so it depends on your data, domain, etc..

[oguiza]

Thank you for your suggestions, @Sandy4321. But this would require additional contributors that we don't unfortunately have. I'll leave it open just in case anybody volunteers to add additional tutorials.