System Information

latest opencv and lastest onnxruntime

Detailed description

Related PR: #23419 and opencv/opencv_extra#1049.

In the above PR, several fixes are indtroduced for the clip head. Those fixes are fine, but the manually generated onnx model named "clip-vit-base-head.onnx" is invalid and cannot be run by onnxruntime:

1. In the Expand operator, shape input should be of type int64 instead of int32. Please use onnx.checker.check_model(model) (it checks operator definitions and so on) before saving a manually generated onnx model.
2. Having that fixed, onnxruntime still complains invalid shape in the Expand operator. Changing the shape value from [1, 1, -1] to [1, 1, 1] fixes the problem. Please run the manually generated onnx model with onnxruntime to check whether it is all valid.

cc @dkurt

Steps to reproduce

Install onnxruntime, then

import numpy as np
from onnxruntime import InferenceSession


net = InferenceSession("./clip-vit-base-head.onnx")
out = net.run([], {"input": np.random.randn(1, 1, 3).astype(np.float32)})
print(out[0].shape)

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)