0

❓Question

Hi, Following steps were taken

  1. I trained yolo tiny on a custom data set with just one class

  2. Converted .weights(darknet) to .h5 (keras) (verified and keras model is working fine as well)

  3. Now when I convert Keras to core ml model I am not getting coordinates and confidence as outputs

Command used to convert to core ml

coremltools.converters.keras.convert(
'model_data/yolo.h5',
 input_names='image',
 class_labels=output_labels,
 image_input_names='image',
 input_name_shape_dict={'image': [None, 416, 416, 3]}
 )

Though I have checked a third party Yolo model converted to core ml giving coordinates and confidence

refer the screenshot:

3rd party Yolo model converted to core ml

enter image description here

my Yolo model converted to core ml

enter image description here

System Information

  • Keras==2.1.5

  • coremltools==3.3

Ankit Sachan
  • 7,690
  • 16
  • 63
  • 98

2 Answers2

2

Don't add this: class_labels=output_labels -- It will make your Core ML model into a classifier, which are treated special in Core ML. Since your model is an object detector, you don't want this.

Look here for the rest: https://github.com/hollance/YOLO-CoreML-MPSNNGraph

Basically, you need to decode the bounding box coordinates yourself in Swift or Obj-C code. You can add this to the model too, but in my experience that is slower. (Here is a blog post that shows how to do this for SSD, which is similar but not exactly the same as YOLO.)

Matthijs Hollemans
  • 7,706
  • 2
  • 16
  • 23
  • so I tried the steps mentioned in the blog post, frankly speaking, its overwhelming for a beginner 2 questions: 1) Is there a short way 2) How to do it for yolo? – Ankit Sachan May 16 '20 at 14:07
  • The short way is to use the code from the YOLO-CoreML-MPSNNGraph repo I mentioned. It's also the best way. – Matthijs Hollemans May 16 '20 at 16:06
0

I'll keep on updating this as it may be useful for others:

This will be very specific to the scenario where you have custom darknet weight which detects custom objects in a scene.

Typical follow for this will be:

  1. Train Custom weights (I am assuming you'll be using Darknet to do this)
  2. Convert Darknet weights to keras checkpoints or .h5 (whichever suits you)
  3. Use Coreml tools to convert your keras model to coreml model
  4. Make sure you use Imagetype while creating coreml model
  5. test coreml model if they are giving a multi-array of shape 13x13x125
  6. Use this coreml model in iOS apps

Hope this will be helpful. If need more help drop your request in comments

Regards Ankit

Ankit Sachan
  • 7,690
  • 16
  • 63
  • 98