Inconsistency in the dataset loading approach

[PCTsai]

Hi @AndrzejCodahead,

I've been using the scripts in fish-identification/helper/classification/CreateDatasetAndTrain.ipynb to train my own dataset. This process involves two Python scripts: auto_train_triplet.py and dataset_creator_by_coco.py.

I noticed an inconsistency in the dataset loading approach. Specifically, the script auto_train_triplet.py utilizes the class FishialDatasetFoOnlineCutting to load the train and validate datasets, whereas train.py from the classification module employs FishialDataset. It appears that both classes are essential yet they are used differently. Could you please provide more detailed insights into the training flow and clarify this discrepancy?

Furthermore, if FishialDatasetFoOnlineCutting is indeed the correct class for my purposes, I understand that I need to create a voxel dataset using the fiftyone Python package. However, the output from dataset_creator_by_coco.py are JSON files, not a voxel dataset. Should I be saving my data in voxel dataset format instead?

Thanks.