I’m trying to train the Mask2former model using the TAO platform on a binary instance segmentation task (1 object + background). Since there are differences in how the pipeline is configured to make this work per mmdetection documentation, I need to know all the parameters I need to setup correctly (num_classes, contiguous_id, etc.) and changes I need to make to make this work within TAO.
My dataset is in COCO format with my images in their respective folders and annotations in the annotation folder. My annotations are in RLE format, but I can also use polygons if necessary. I only have 1 category; hence, my label map consists of only 1 item/thing. My category ID starts from 1 as advised.
What I need:
- My training logs clearly indicating losses going down and accuracy/IoU going up.
- Indication of the model learning the right thing or object and not the background.