Costum Dataloader for Detectron2

[Ksieber26]

📚 Documentation

• Link to the relevant documentation/comment/tutorials:

• dataset_mapper Doc: https://detectron2.readthedocs.io/en/latest/_modules/detectron2/data/dataset_mapper.html
• Costum Data Loader Github: Using Custom Data Loader for Augmentation #1107 (Seems like the data mapper and not the Data Loader, am I wrong ?)

Currently, I'm trying to write a costum dataloader in order to train on the HDF5 format of stacked numpy images to save preprocessing time. The problem is I don't find any understandable solution how to adapt a new dataloader instead of the 'register_coco_instances()' Function.

I already saw the documentary about the dataset_mapper. Isn't the Mapper a step further after the dataloader, because the mapper applies the changes to the dataset_dict and I only want to replace the register_coco_instances function to just use already loaded numpy images.

I hope someone can help me.

Thank you in advance!