Issue

The latest napari-deeplabcut version retains the likelihood column when refining machine annotations (they are added to the CollectedData h5 if present).
The matlab file creation function always assumed only x and y are present, and did not filter out likelihood, leading to dataset creation failure.
Note: this raises in terminal but could cause the GUI to hang instead, as reported in #3319.

Related

• Should napari-deeplabcut always discard likelihood when saving to CollectedData ? -> Yes
  • Fixed plugin-side in Prevent likelihood columns from being saved in CollectedData napari-deeplabcut#204
• Do we want to systematically filter at the training set h5/csv creation step ? -> not needed due to above point
• Aims to close Shuffle creation failure #3319

Fix

Improves the robustness of the training data formatting by ensuring that any "likelihood" columns present in the input DataFrame are dropped before .mat formatting, and adds a corresponding test to verify this behavior.

Data formatting fixes:

• Updated format_training_data in trainingsetmanipulation.py to automatically detect and remove "likelihood" columns from the DataFrame before processing, ensuring only "x" and "y" coordinates are used. Added validation to require both "x" and "y" columns and to check that the number of coordinate values per row is even.
• Improved error handling for malformed data, raising clear exceptions when required coordinate columns are missing or when the data shape is unexpected.

Testing:

• Added a new test test_format_training_data_ignores_likelihood_columns in test_trainingsetmanipulation.py to verify that the presence of "likelihood" columns does not affect the output of format_training_data.