Question about fine-tuning on custom dataset for instance segmentation

[adityakankariya]

I want to perform instance segmentation by fine-tuning Mask2Former on a custom drone thermographic image COCO dataset. I follow instructions to register the dataset in the DatasetCatalog, and I am able to get the metadata to work with that. Now I don't know how to use train_net.py as DefaultTrainer(cfg) does not work with Mask2Former. For some reason, the train_net config is still looking for the standard COCO 2017 prebuilt dataset. How do I get it to ingest my custom dataset? Is It something to do with the environment variable DETECTRON2_DATASETS? Does it go back to registering the COCO dataset? Clearly, the issue is seen in the last screenshot, where the datasets are already set to coco_2017 for some reason...

[leo96code]

@adityakankariya were you able to sort out?

[Ihafnonaim]

I want to perform instance segmentation by fine-tuning Mask2Former on a custom drone thermographic image COCO dataset. I follow instructions to register the dataset in the DatasetCatalog, and I am able to get the metadata to work with that. Now I don't know how to use train_net.py as DefaultTrainer(cfg) does not work with Mask2Former. For some reason, the train_net config is still looking for the standard COCO 2017 prebuilt dataset. How do I get it to ingest my custom dataset? Is It something to do with the environment variable DETECTRON2_DATASETS? Does it go back to registering the COCO dataset? Clearly, the issue is seen in the last screenshot, where the datasets are already set to coco_2017 for some reason...

Hi, Did you figure it out?