search

automl / Auto-PyTorch

Automatic architecture search and hyperparameter optimization for PyTorch
Apache License 2.0
2.33k stars 282 forks source link

[Feature request] time series classification #491

Open Aceticia opened 1 year ago

Aceticia commented 1 year ago

Theoretically this should be compatible with the encoder - decoder style framework we have here?

Aceticia commented 1 year ago

Alternatively, adding a method that returns the extracted features after a forecasting model is trained.

dengdifan commented 1 year ago

Thanks for the suggestion, Unfortunately, we are busy with some other work these days and may add this feature in the near future.

However, here are some hints: you could write a new pipeline for time series classification, which should be something similar to the tabular classification pipline and reimplement the [classification network] (https://github.com/automl/Auto-PyTorch/blob/master/autoPyTorch/pipeline/components/setup/network/base_network.py), adding them to the pipeline. Additionally, the network backbone might also need to be modified accordingly: Auto-regressive needs to be removed.

Then you need to create a new class for time series classification task that is inherent from this class. Then everything should work as expected