compiler error: Espresso exception: "Invalid blob shape": Cannot broadcast blobs

[franva]

🐞Describe the bug

• A clear and brief description of what the bug is. I successfully converted my ONNX model to mlmodel without any issue. Then I dragged my mlmodel into my Swift project and the project failed at compiling the model:

• Is this a converter issue? If yes, please specify which converter (sci-kit, keras, xgboost etc.) Not sure. I use ONNX converter.

Trace

If applicable, please paste the error trace.

To Reproduce

• If a python script can reproduce the error, please paste the code snippet

from pathlib import Path
import coremltools as ct

root = Path(__file__).parent

model = ct.converters.onnx.convert(
    model = "BiSeNetV2_op12_sim_output_fixed.onnx",
    minimum_ios_deployment_target = '13',
    image_input_names=["x"],
    preprocessing_args={
        "image_scale": 1./(255.0),
        "red_bias": - 0.5/(0.5),
        "green_bias":- 0.5/(0.5),
        "blue_bias": - 0.5/(0.5)
    }
)

saved_model_path = root.parent / "output/bisenetv2_op12_ios13"
model.save("./output/bisenetv2_op12_ios13_outout_fixed_sim.mlmodel")

• If applicable, please attach the source model

And here is the ONNX model(on Microsoft OneDrive): https://1drv.ms/u/s!Asflam6BEzhjgbM4b82jgr-_iALyLA?e=im3LKz

2. create a Swift UI project
3. drag this model into the XCode : https://1drv.ms/u/s!Asflam6BEzhjgbM3PGRG2h_jCY-4tA?e=EEPXno
4. Build and run the app

System environment (please complete the following information):

• coremltools version (e.g., 3.0b5): 5.2.0
• OS (e.g., MacOS, Linux): MacOS
• macOS version (if applicable): 10.15.7
• XCode version (if applicable): 12.4
• How you install python (anaconda, virtualenv, system): anaconda
• python version (e.g. 3.7): 3.8
• any other relevant information:
  • ONNX Operator version12.

[franva]

I wrote an Python script to test the ml model and found more details of the same issue:

Here is the code, helpfully it helps to figure out the issue:

import coremltools as ct
import numpy as np
import cv2
from paddle.vision.transforms import Compose, Resize, Normalize, Transpose

loaded_model = ct.models.MLModel("./output/bisenetv2_op12_ios13_outout_fixed_sim.mlmodel")

def load_image_cv(path):
    img = cv2.imread(path)
    return img

img = load_image_cv("boxhill_079.jpeg")
scaled_img = img.astype("float32") / 255.0

trfms = Compose([
    Resize(size=(1024, 1024)),
    Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), data_format="HWC"),
    Transpose()
])

transformed_img = trfms(scaled_img)

transformed_img = transformed_img[np.newaxis, :]

out_dict = loaded_model.predict({"x": transformed_img})

print(out_dict)

The error happens on the line:

out_dict = loaded_model.predict({"x": transformed_img})

[franva]

@TobyRoseman just wondering any update for this one?

[TobyRoseman]

@TobyRoseman just wondering any update for this one?

No update. Please note: our onnx converter is deprecated. We're planning on removing it in our next major release. I think it's unlikely this will be fixed.

[TobyRoseman]

ONNX support has been removed in coremltools 6.0.