augmentations applied to test and val data

[benjaminsshoffman]

https://github.com/earthspecies/representation-learning/blob/d13e0e88c05ae56d18940f2a2c5a7d15f7db7a00/representation_learning/data/dataset.py#L243

I think this applies augmentations (e.g. mixup) to test and val data, which is not typical

[nkundiushuti]

in theory, yes, we have:

    postprocessors: list[Any] = []
    aug_processor: Optional[AugmentationProcessor] = None
    if cfg.augmentations:
        aug_device = "cpu"  # DataLoader workers run on CPU
        aug_processor = AugmentationProcessor(cfg.augmentations, cfg.sr, aug_device)
        postprocessors.append(make_item_postprocessor(aug_processor))

    # Build datasets from the configuration
    if data_config is None:
        data_config = cfg.dataset_config

    ds_train, ds_val, ds_test = _build_datasets(data_config, postprocessors=postprocessors, label_type=cfg.label_type)

and this is not ok. it's somehow inherited from the training code in run_training (maybe some split of training set was used as validation)
in practice all augmentations are disabled in run_evaluate for all the evaluation sets (except birdset training).
there's also this in the dataset.py

    # Determine augmentation strategy based on context
    if is_evaluation_context:
        # Evaluation context (run_evaluate.py):
        # - Default: no augmentations for train/val
        # - BirdSet: augmentations for train/val (enable_eval_augmentations=True)
        train_aug_processor = aug_processor if enable_eval_augmentations else None
        eval_aug_processor = aug_processor if enable_eval_augmentations else None
    else:
        # Training context (run_train.py):
        # - Train: augmentations enabled (standard training behavior)
        # - Val: no augmentations (standard validation behavior)
        train_aug_processor = aug_processor
        eval_aug_processor = None

so in theory these augmentations are applied with the collater
the initial post-processors are not augmentations but any fixed transformations we apply to the raw dataset in esp-data when loading it such as mapping labels to numbers, sampling, filtering