hearbenchmark / hear-eval-kit

Evaluation kit for the HEAR Benchmark
https://hearbenchmark.com
Apache License 2.0
56 stars 17 forks source link

Store test output #330

Open turian opened 3 years ago

turian commented 3 years ago

task_predictions_test should include the test_fold_str.

_score_epoch_end() should include an optional prediction_pkl_name: Optional[str] = None parameter. If it is not None (because we passed in test_fold_str), the final prediction_logit and target should be dumped to a pickle file.

For event models, we should also dump predicted_events and target_events and maybe prediction.

The idea is that if we do a long expensive prediction on some model, we want the final test output saved. That way we could do a more detailed analysis, e.g. seeing what the scores are for each class.