[DOC] input formats for transformations

[michaelkalz]

Describe the issue linked to the documentation

I am a medical student with some background in CS and I am trying to build a new API to utilise the transformers for our machine learning pipeline tool (https://photon-ai.com/). However, I am having quite a rough time trying to get the transformers running on some of our data, that come in the shape: (rat instance, sensors, measurements) and need to be np.arrays.

There are not that many transformers with examples/example data and (at least to me) it is not clear when to use pd.Df or np.array as input and in which format (input and output) without digging deep(er) into the linked papers (if linked) or github repo.

Suggest a potential alternative/fix

Could you please provide some more information about input/output type and format and/or more examples for the transformation section?

[fkiraly]

Have you seen the transformations tutorial?

Tutorial 3, e.g., here on binder:
https://mybinder.org/v2/gh/sktime/sktime/main?filepath=examples

There's also an extended tutorial on transformations and pipelines here:
https://github.com/sktime/sktime-tutorial-pydata-global-2022

[fkiraly]

that come in the shape: (rat instance, sensors, measurements) and need to be np.arrays.

This sounds like it is already in the numpy3D format (mtype) accepted by all sktime transformers and should work out of the box with all transformers.

it is not clear when to use pd.Df or np.array as input and in which format (input and output) without digging deep(er) into the linked papers (if linked) or github repo.

Both input formats - pd-multiindex and numpy3D should work with all transformers!

The papers are outdated, the online documentation and the tutorial notebooks on the main repo are "living" and should be evergreen.

[jernsting]

But if this is numpy3D Format, how would I then expect the target? My assumption would be, that there is one label for the whole instance, but this seems to be different from what sktime expects. [instance given our input (instances, sensors, measurements)]

[jernsting]

In other words: The numpy3D input is not usable for multivariate input (ref https://sktime.readthedocs.io/en/latest/examples/AA_datatypes_and_datasets.html#Section-1.1.3:-Time-series---the-%22np.ndarray%22-mtype, capabilities: cannot represent multivariate series).

However, it seems that Panels are actually a generalisation of multivariate input? So using the Panel input type with data from just one instance should be equal to using multivariate input?

[fkiraly]

Just noticed that we haven't yet moved the pydata 2022 material to the sktime main repo and it might be a bit patchy... opened an issue here
#4373

In the meanwhile, I hope the link to the pydata 2022 material helps.

[fkiraly]

moving this to the discussion forum

[jernsting]

Documentation and repository are out of sync. E.g. https://github.com/sktime/sktime/blob/v0.16.1/sktime/transformations/series/fourier.py clearly does not contain FourierTransform, but latest documentation (https://www.sktime.net/en/latest/api_reference/auto_generated/sktime.transformations.series.fourier.FourierTransform.html) states that it should exist.

Also version specific links in the documentation do not work. For example the link to the documentation for v0.16.1 does not work: https://sktime.readthedocs.io/en/v0.16.1/

[fkiraly]

Thanks for bringing this up!

Discussion moved to two issues:

• [DOC] Version specific docs on readthedocs are not working - "not found" #4407
• [DOC] Documentation and repository are out of sync #4408