[DOC] Fix rocket transformers which document unequal length capability

[maorgranot1]

Describe the bug

Aeon's minirocket/multirocket wont support data of unequal length although it suppose to

The documentation states "Input data, any number of channels, equal length series of shape ( n_cases, n_channels, n_timepoints) or list of numpy arrays (any number of channels, unequal length series) of shape [n_cases], 2D np.array (n_channels, n_timepoints_i), where n_timepoints_i is length of series i", the class wont fit data of type list length N, containing np arrays of size (2XM[i]).

Steps/Code to reproduce the bug

import numpy as np
from aeon.transformations.collection.convolution_based import MiniRocket, MultiRocket

x = [np.random.rand(2,n) for n in np.floor(np.random.rand(100)*20).astype(int)+100]
y = np.round(np.random.rand(100)).astype(int)
print(f"x is a list of length {len(x)} containing np arrays of size {[data.shape for data in x]}")
print(f"y length is {y.shape}")

minirocket = MiniRocket(n_kernels=1000, random_state=42)
minirocket.fit(x, y) # result in fail because of unequal lengths

multirocket = MultiRocket(n_kernels=1000, random_state=42)
multirocket.fit(x, y) # result in fail because of unequal lengths

Expected results

no error is thrown

Actual results

Versions

python: 3.10.12 (main, Nov 6 2024, 20:22:13) [GCC 11.4.0]
executable: /usr/bin/python3
machine: Linux-6.1.85+-x86_64-with-glibc2.35
Python dependencies:
aeon: 1.0.0
pip: 24.1.2
setuptools: 75.1.0
scikit-learn: 1.5.2
numpy: 1.26.4
numba: 0.60.0
scipy: 1.13.1
pandas: 2.2.2

[baraline]

Thanks for adding the code to reproduce the issue. It seems that unequal length capability was removed in a previous version. So what you are experimenting is "normal" for the current version.

We can converts this issue into a enhancement one to implement this capability back when we retrieve the technical problems that caused this change before.

[maorgranot1]

Thanks for adding the code to reproduce the issue. It seems that unequal length capability was removed in a previous version. So what you are experimenting is "normal" for the current version.

We can converts this issue into a enhancement one to implement this capability back when we retrieve the technical problems that caused this change before.

so should it be labled as "documantion" and "enhancement"?

[baraline]

Better keep this one as is and open another one for the enchancement. If you wish to create it go ahead, otherwise i'll do it in the evening.
Do you want to work on the issue of re-implementing it ?

[maorgranot1]

Better keep this one as is and open another one for the enchancement. If you wish to create it go ahead, otherwise i'll do it in the evening. Do you want to work on the issue of re-implementing it ?

haha I'm afraid I'm not fluid enough in python to understand where does this loss of function accrued. Not to mention fix it..

[baraline]

No worries ! I'll tag you in the issue if you didn't find time to create it before I do so you can keep track then.

[TonyBagnall]

thanks for this, it is a general issue for all classifiers, we should make it clear in the base class fit that it only accepts list of unequal length numpy if "capability:unequal" is true

[TonyBagnall]

hi @maorgranot1 can you point to the documentation you looked at? I suspect it was from an old version. The latest

https://www.aeon-toolkit.org/en/v1.0.0/api_reference/auto_generated/aeon.transformations.collection.convolution_based.Rocket.html#aeon.transformations.collection.convolution_based.Rocket.fit

has the text explaining about the tag values for capabilities.

[maorgranot1]

hi @maorgranot1 can you point to the documentation you looked at? I suspect it was from an old version. The latest

https://www.aeon-toolkit.org/en/v1.0.0/api_reference/auto_generated/aeon.transformations.collection.convolution_based.Rocket.html#aeon.transformations.collection.convolution_based.Rocket.fit

has the text explaining about the tag values for capabilities.

Second row after "Or list of numpy arrays..."

[TonyBagnall]

ah ok, so need to more explicitly link the sentence with the secod paragraph that explains the constraints of list of numpy, thanks.

discussed unequal length with angus today and agreed on best way to handle it, more to follow

[Kaustbh]

Hi, is the issue open for a pull request, or is a decision still pending?

[MatthewMiddlehurst]

Feel free to edit the documentation to make it more accurate

[Kaustbh]

@MatthewMiddlehurst I have made the PR, please review and let me know if any changes are needed.

[Kaustbh]

@aeon-actions-bot assign @Kaustbh

[Kaustbh]

Hi @MatthewMiddlehurst I noticed that when checking the documentation for MiniRocket.fit(), it displays the docstring from the BaseCollectionTransformer.fit() method rather than MiniRocket-specific details. However, MiniRocket does not explicitly define fit()—it only implements _fit().
Since fit() is inherited from BaseCollectionTransformer, it retains the base class's documentation, because of this reason I am adding the documentation for unequal-length in the Notes section of MiniRocket class docstring

[MatthewMiddlehurst]

Ah, sorry I thought this was an issue in the minirocket docstring, not confusion from the generic fit/predict docstrings. Your first PR is more accurate, I will reopen and make a comment.