[Bug]: wrong shape of results produced by OneHot operator

[coffezhou]

OpenVINO Version

2025.3.0-19357-ef1dae8b89d

Operating System

Ubuntu 20.04 (LTS)

Device used for inference

CPU

Framework

None

Model used

No response

Issue description

For the following onnx model,

it only contain an OneHot operator.
When I run this model using onnxruntime, the results are as follows:

ONNXRuntime:
 [array([[[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 1.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 1., 0.],
        [0., 0., 0.],
        [0., 0., 1.],
        [0., 0., 0.],
        [0., 1., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 1., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 1.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]]], dtype=float32)]
shape:
 (10, 5, 3)

However, when I run it using openvino, the results are as follows:

OpenVINO:
 [array([[[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 1.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 1., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 1., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 1., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 1.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]]], dtype=float32)]
shape:
 (10, 5, 3)

In the above results, the third block is different from that in the results produced by onnxruntime. Only two elements equals to 1, while there are 3 elements that equal to 1 in the onnxruntime's results. I also run this model using onnxruntime-gpu and tensorRT and the third block in the results are also containd 3 elements equaled to 1.

Step-by-step reproduction

This bug can be reproduced by the following code with the model in the attachment.

import sys
import os

import numpy as np
import onnx
import onnxruntime

import openvino as ov
import pickle

def main():
    onnx_model = onnx.load("555.onnx")
    
    with open("inputs.pkl", "rb") as fp:
        inputs = pickle.load(fp)
    
    try:
        ort_session = onnxruntime.InferenceSession(
            onnx_model.SerializeToString(), providers=["CPUExecutionProvider"]
        )
        ort_output = ort_session.run([], inputs)
    except Exception as e:
        print(e)
        sys.exit(1)
        
    print("ONNXRuntime:\n", ort_output) 
    print("shape:\n", ort_output[0].shape)  
        
    ov_model = ov.convert_model("555.onnx")
    ir_path = "_temp_OVIR_555.xml"
    
    ov.save_model(ov_model, ir_path)
    core = ov.Core()
    model = core.read_model(ir_path)
    
    compiled_model = core.compile_model(model=model, device_name='CPU')
    
    output_key = compiled_model.outputs

    ov_output = []
    for output in output_key:
        ov_output.append(compiled_model(inputs)[output])
        
    print("OpenVINO:\n", ov_output)
    print("shape:\n", ov_output[0].shape)
    
    np.testing.assert_allclose(ov_output, ort_output, rtol=0.1, atol=0.1)

    
if __name__ == "__main__":
    
    main()

testcase.zip

Relevant log output

Issue submission checklist

• I'm reporting an issue. It's not a question.
• I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
• There is reproducer code and related data files such as images, videos, models, etc.

[rkazants]

Hi @coffezhou,

The problem is on OpenVINO side, namely, OV OneHot operation does not work for negative indices (behavior is undefined).
13-th index in your input is negative -1 for which output element is different between ONNX-RT and OV:

@gkrivor, can you fix ONNX FE for OneHot negative indices by computing mod?
@mitruska, or we can consider updating OneHot to support negative indices similar to ONNX?

@coffezhou, in the meantime, it is critical to your application, you can patch our model to have non-negative indices.

Best regards,
Roman