User Report

When uploading a trained AutoGluon TabularPredictor artifact to Kaggle, the __version__ file in the metadata is not able to be uploaded, because Kaggle for some unknown reason does not allow this specific file name to be uploaded. This results in a confusing warning message when a user tries to then load the predictor via TabularPredictor.load:

WARNING: Could not find version file at "/kaggle/input/test-ag/ag-20230301_033436/__version__".
This means that the predictor was fit in a version `<=0.3.1`.

To resolve, we can rename the file from __version__ to version.txt, or some equivalent.

• Ensure backwards compatibility, check for __version__ if version.txt doesn't exist.
• Check if the same problem exists for TimeSeries and Mulitmodal: Yes, fixed in [tabular, timeseries] Rename cached version file to version.txt from __version__ #4203 for TimeSeries and Tabular. Multimodal does not check version on load.