custom_dataset using train_net.py

[Kevin-Delnoije]

Hey I succesfully added a custom dataset in a jupyter notebook using the steps below. (mostly following the balloon tutorial notebook)

• define custom_function to create list[dict]
• DatasetCatalog.register(name, lambda: custom_function(dirname) )
• MetadataCatalog.get(name)
• add train/test to config
• instantiate DefaultTrainer(cfg)

Now i am trying to move the code out of a jupyter notebook into a similar training script as train_net.py the confusing part for me is where the coco/lvis/pascal are normally registered and what setup magic is hidden behind default_setup(cfg, args)

It is possible to use the same steps as in the notebook, but if possible I would like to modify the train_net.py as little as possible so i can enjoy the nice default settings :)

[ppwwyyxx]

You can register the dataset in train_net.py the same way as in the notebook - I'm not sure what exactly you're trying to ask.

default_setup is documented in the API docs: https://detectron2.readthedocs.io/modules/engine.html#detectron2.engine.defaults.default_setup

[Kevin-Delnoije]

I don't know how the coco dataset is registered in train_net.py and i would like to understand that

[ppwwyyxx]

It's automatically registered by default by detectron2 at https://github.com/facebookresearch/detectron2/blob/125e4ff9d2e628a2d9db77e787570093c1ecf522/detectron2/data/datasets/builtin.py#L217