[Python] How to Convert TensorFlow Arbitrary Image Stylization Model to Core ML for Swift?

I'm trying to implement arbitrary image stylization in Swift using TensorFlow's pre-trained model. After downloading the model, I got a saved_model.pb file. However, iOS does not support this format directly, so I tried converting it to Core ML using coremltools.

Python Code for Conversion:

import coremltools as ct

# Load the TensorFlow model
model = ct.convert(
"/Users/kirit/Downloads/saved_model", # e.g., a SavedModel directory
inputs=[
ct.ImageType(shape=(1, 256, 256, 3), name="contentImage", scale=1/255),
ct.ImageType(shape=(1, 256, 256, 3), name="styleImage", scale=1/255)
],
outputs=[ct.TensorType(name="stylizedImage")],
source="tensorflow"
)

# Save the Core ML model
model.save("DualInputStylization.mlmodel")



Issue:

• The conversion process always fails with various errors.
• Common errors include:
• Failed to load _MLModelProxy: No module named 'coremltools.libcoremlpython'
• Converter was called with source="tensorflow", but missing tensorflow package
• TensorFlow import errors due to missing imp module in Python 3.13

System Information: Python Version: 3.13.1 TensorFlow Version: (Latest) coremltools Version: (Latest) OS: macOS (15.1.1 (24B91))

Questions:

• How can I successfully convert a TensorFlow saved_model.pb for arbitrary image stylization to a Core ML mlmodel format?
• Is there an alternative approach for using this model in Swift?
• Is there a specific TensorFlow version that works best for Core ML conversion? Any guidance would be appreciated! Thanks.

Continue reading...