I'm new to Xcode and have some trouble getting my model into an appropriate format. I have quite a similar problem as in this unanswered question but for OBBs:
I trained a YOLOv8n-obb model on a custom dataset containing just one label in order to use it in an iOS app. When I export it as on ultralytics website:
model.export(format='coreml', int8=True, nms=True, imgsz=640)
and add it to Xcode I can't use the preview in Xcode. The bigger problem however, its output becomes a MultiArray.
I get that this multiarray also represents the predictions made by the model without nms. If I understood it correctly, the model normally outputs the confidence and coordinates separately (e.g when downloaded from ultralytics and put in Xcode).
Is there some 'easy' way to convert the model that it outputs the data in the same way with my custom class and its confidence (+ coordinates)?
I assumed that the model is not identified as a classifier by CoreML and I tried converting it as in CoreML docs. I got stuck at
# Convert the model using the Unified Conversion API
model = ct.convert(
keras_model, inputs=[image_input], classifier_config=classifier_config,
)
ValueError: In op 'classify', number of classes must match the size of the tensor corresponding to 'probabilities'.
'Classifier_config' should be a list containing the labels but it works with neither one entry (custom label) nor with 80 (same as coco on which the model was trained on originally).
I'm sorry if I overlooked the obvious but can't imagine that there's no straight forward way to use a self trained OBB model in an iOS app.
