Error trying to convert model to CoreML

[raulsf6]

Hi, first of all, thanks for you amazing.

I'm facing an error trying to convert tiny-yolo to CoreML suing the procedures explained in the repo. I manage to get the .h5 representation but when I try to export it to CoreML I get this error:

Traceback (most recent call last):
  File "coreml.py", line 17, in <module>
    coreml_model.input_description['image'] = 'Input image'
  File "/home/rsanchez/.local/share/virtualenvs/Convert-r3iBw4lK/lib/python2.7/site-packages/coremltools/models/model.py", line 79, in __setitem__
    raise AttributeError("No feature with name %s." % key)
AttributeError: No feature with name image.

Could it be due to package versions? I have a virtual Python 3 environment for the Darknet -> Keras transformation and another Python 2.7 environment for the Keras -> CoreML transformation. The only pinned version is Keras but I suppose tensorflow has to be 1.x version. Actually, my packages versions are:

• Tensorflow 1.15.0
• Keras 1.2.2
• coremltools 4.0
• Python 2.7

[hollance]

It's probably the coremltools version (I think I used version 2.0 for this or perhaps even 1.0), but the error is that your model does not have an input named 'image'.

[raulsf6]

@hollance thanks for your answer! I'll try with those coremltools versions. Could you also tell me which h5py version you used? I think the problem could be when exporting from Darknet to Keras. The h5 model looks like this in Netron:

So maybe the first transformation is messing up the input/output names?

[hollance]

Just rename the input in the code from image to input_1. It should work then. The name of the input isn't very important.

[raulsf6]

@hollance, I just managed to execute CoreML.py succesfuly with this packages versions without modifying any code:

keras==1.2.2
tensorflow==1.15.0
coremltools==2.0
h5py

From Darknet to Keras everything works as long as yo use Keras 1.2.2 and Tensorflow 1.x.

Maybe it would be a good idea to update scripts to new versions, specially coreml.py because Python2.7 has no longer support.

Amazing work with this repo!