Open MartinJensen37 opened 3 years ago
@MartinJensen37 as can be seen in the printed configs, you still train on 81 classes. You only need to change the NUM_CLASSES (in a class that inherits from ´config´) to 3+1 (background). Don't change anything else
@MartinJensen37 as can be seen in the printed configs, you still train on 81 classes. You only need to change the NUM_CLASSES (in a class that inherits from ´config´) to 3+1 (background). Don't change anything else
@Shakesbeer333 I changed the number of classes in the config file, it now shows 3 classes instead but it still gives the exact same error message as before "could not broadcast shape xx into shape xx"
@MartinJensen37 also make sure that you annotation file does not contain background
as an category as it is added separately in Mask RCNN. So you annotation file starts with category id=1.
Hello! I'm trying to retrain the Mask-RCNN using only 3 classes from the MS COCO dataset. I've been following the step in #247 in order to do this. Firstly, I changed
NUM_CLASSES=3
and then I changed the ´class_ids = [1, 67]and finally I added
exclude=['mrcnn_class_logits', 'mrcnn_bbox_fc', 'mrcnn_bbox','mrcnn_mask']
to the load_weights function on line 469 as in #247. For reference I'm using the 2014 training dataset from MS COCOHowever I get an error message saying that something with the image dimensions are off(the error message is much longer than this but it keeps repeating being unable to fit one shape into another):