fix: train.py post-loop bugs in image and audio classification

[musicalplatypus]

Problem

Both image_classification/train.py and audio_classification/train.py have a "Log best epoch results" section that runs after for epoch in range(args.start_epoch, args.epochs). Two independent bugs there:

1. audio_classification/train.py logged best['f1'] under the "AUC ROC Score" label instead of the best['auc'] value that's actually computed and stored in best on every improving epoch:

   best['accuracy'], best['f1'], best['auc'], best['conf_matrix'], best['epoch'] = avg_accuracy, avg_f1, auc, avg_conf_matrix, epoch
   ...
   logger.info(f"AUC ROC Score {best['f1']:.3f}")  # should read best['auc']

   image_classification/train.py already logs this correctly.

2. Both scripts crash when the loop runs zero iterations — e.g. --resume pointed at a checkpoint that already satisfies --epochs, a real and expected use case for re-running only the post-training export/compile steps without retraining. best = dict(accuracy=0.0, f1=0, conf_matrix=dict(), epoch=None) never gets 'predictions'/'ground_truth'/'auc' populated if no epoch runs, but the post-loop section unconditionally reads them:

   utils.print_file_level_classification_summary(dataset_test, best['predictions'], best['ground_truth'], phase)

   raising KeyError: 'predictions' (or KeyError: 'auc' in image_classification, which hits that key first).

Fix

Both scripts now guard the whole "Log best epoch results" block on best['epoch'] is not None — reusing the epoch=None sentinel already present in the initial dict — and log a clear warning instead of crashing when no epoch actually ran.

Testing

Added tinyml-tinyverse/tests/test_train_best_epoch_bugs_vision_audio.py, which drives each script's real main() through a heavily mocked model/data pipeline (mocking every heavy helper from common/train_base.py and common/models.py) with args.start_epoch == args.epochs, reproducing the "--resume to an already-completed checkpoint" scenario, and asserts main() completes without raising. A separate test for audio_classification runs the loop for one improving epoch with f1 and auc set to distinct values and asserts the "AUC ROC Score" log line reports the auc value, not f1.

• Verified all three tests fail on pre-fix code with the exact errors described above (KeyError: 'predictions', KeyError: 'auc', and the mislabeled log line).
• Verified all three pass with the fix.

Scope note

This PR is one of a pair — the same two bugs also affect timeseries_classification/train.py and timeseries_forecasting/train.py, but those files are already touched by the open #23 (which is itself stacked on #22), so that fix was pushed there instead of opened as a new PR, to keep file-touch history consolidated. image_classification/audio_classification aren't touched by any other open PR, hence this standalone PR.

🤖 Generated with Claude Code