search

apple / coremltools

Core ML tools contain supporting tools for Core ML model conversion, editing, and validation.
https://coremltools.readme.io
BSD 3-Clause "New" or "Revised" License
4.44k stars 643 forks source link

How to convert YOLOX model with proper output structure #2157

Closed rudetrue closed 8 months ago

rudetrue commented 8 months ago

I have a YOLOX Object Detection model I successfully converted to a .mlpackage using coremltools.

My issue is interpreting the output. When I open the .mlpackage in Xcode I can see the model output is a MultiArray (Float32 1 x 13125 x 15). How do I interpret that output, or get easy to use outputs like I do from a model trained with CreateML? I'd like to have a model Preview tab, confidence output, and coordinate output.

This is roughly how I've converted my model:

# This file is from running https://github.com/Megvii-BaseDetection/YOLOX/blob/main/tools/export_torchscript.py
model_path = 'yolox.torchscript.pt'

input_image = ct.ImageType(name='inputs', shape=(1,3,800,800),scale=1/255,
              bias=[0,0,0])

coreml_model = ct.convert(
    model_path,
    source="pytorch",
    inputs=[input_image],
    minimum_deployment_target=ct.target.iOS16,
)

Is there a guide similar to https://apple.github.io/coremltools/docs-guides/source/classifiers.html but for object detection models? It doesn't appear that there is an ObjectDetectionConfig. How do I set the class names for object detection? Any help/guidance would be appreciated!

TobyRoseman commented 8 months ago

The output of your .mlpackage should be the same as your PyTorch model. So I would suggest checking the documentation for your PyTorch model.

Since this isn't really a coremltools question, I'm going to close this issue. You can ask Xcode related questions in our developer forum: https://developer.apple.com/forums/.