Time series event detection models

[samueljackson92]

Adds four new time series event detection models to toktagger:

• DTW motifs - use user generated annotations as motifs and use Dynamic Time Warping to detect similar events within the time series trace.
• Stumpy - use user generated annotations as motifs for the matrix profile to find similar motifs in the data.
• Minirocket - use minirocket and a linear probe to classify time series windows.
• Shapelets - use a shapelet transform to classify time series windows (using sktime's default RotationForest estimator, an ensemble of decision trees).

Each model supports either single or multi-channel inputs.

To Test:

• open a project with a time series view.
• click on the train models button.
• train one of the models
• once trained, close the training window
• open the model predictions interface
• run the model
• annotations should be generated for this shot and plotted.

Additionally, check the new tests & CI are passing

[abdullah-ukaea]

Known issue: multiple signal inputs

As described in #330, the parameter description says that users can provide multiple signals, but the current schema form only renders one input and provides no way to add another. The Pydantic schema supports multiple values, the problem is in the UI’s list handling.

@wk9874 is looking into this. I have also prototyped a possible fix in https://github.com/ukaea/toktagger/blob/slj/ts-event-detection-models-schemaFix/toktagger/ui/src/app/components/ui/schemaForm.tsx

[abdullah-ukaea]

The new models trained and predicted successfully in my functional testing, but the training forms contain too much text and require excessive scrolling. Overall it is is not a good UI experience.

Could we show only a short model summary and brief parameter descriptions in the form? The detailed explanations could perhaps live in the Zensical documentation, with a link from the UI if needed.

The user currently sees:

I think the model summaries could be closer to:

• DTW: “Detects events that have a similar shape to labelled examples using Dynamic Time Warping.”
• STUMPY: “Detects events that are similar to labelled examples using STUMPY.”
• MiniRocket: “Learns to detect labelled events using MiniRocket time-series features.”
• Shapelet: “Learns to detect events using characteristic shapes found in the training signals.”

Parameter descriptions should be similarly concise.

The form text is generated from the model docstrings in these files, so make the changes here:

• toktagger/api/models/shapelet.py
• toktagger/api/models/stumpy_motif.py
• toktagger/api/models/dtw_motif.py
• toktagger/api/models/minirocket.py

[wk9874]

• tests have broken since merging into dev, please fix :)

[abdullah-ukaea]

We have finished the first review of this branch, feel free to address the changes when you are ready @samueljackson92

[samueljackson92]

@abdullah-ukaea this should be ready for re-review now.

[abdullah-ukaea]

As discussed, could we make these two usability improvements before I start the next phase of code review? @samueljackson92

1. In the training form, make Signal Names a dropdown containing the signals available for the current sample, rather than requiring users to type them manually.

2. In the prediction tool, let users select a specific trained model by name. Currently, selecting a model type such as MiniRocket automatically uses its latest version. If users have trained models named test1 and test2, there is no way to choose between them.