That code resulted in the first 3 checkpoints getting saved and no improvement thereafter, no matter which backbone or preset I used. At inference time, predict_proba assigned equal probs to all classes for each row. Note that I had to specify column type else the text predictor would get used since the img column is detected as text instead of filepaths.
Tested that the Quick Start shopee example does work. Initially I thought changing the column names helped but turns out that was just reverting to a text predictor because I hadn't specified the column type.
rxjx
commented
8 months ago
Ubuntu 20.04 autogluon 1.0.0
My original train_df had column labels 'img' and 'lbl'. This code failed:
That code resulted in the first 3 checkpoints getting saved and no improvement thereafter, no matter which backbone or preset I used. At inference time, predict_proba assigned equal probs to all classes for each row. Note that I had to specify column type else the text predictor would get used since the img column is detected as text instead of filepaths.
Tested that the Quick Start shopee example does work. Initially I thought changing the column names helped but turns out that was just reverting to a text predictor because I hadn't specified the column type.