openvino: [Bug] Incorrect result for ArgMin ONNX operator

System information (version)

• OpenVINO => 2022.2
• Operating System / Platform => Ubuntu 20.04
• Compiler => from pip
• Problem classification => ❔

Detailed description

inference produce an incorrect result on ArgMin ONNX operator when using larger input size.

Steps to reproduce

run the code below, I have compared the result to pytorch and onnxruntime as well

import torch
import torch.nn as nn
import numpy as np

from openvino.inference_engine import IECore
import onnxruntime as ort

class DummyModel(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x: torch.Tensor):
        _, idx = x.min(dim=0, keepdim=True)
        return idx

def get_input(size, zero):
    x = torch.ones((size,), dtype=torch.int) * -1
    x[:zero] = 0
    return x

if __name__ == "__main__":
    onnx_model_path = "issue.onnx"
    model = DummyModel()
    zero_len = 4
    array_len = 16
    x = get_input(array_len, zero_len)
    y = model(x)

    input_name = "input"
    output_name = "output"
    torch.onnx.export(model, (x,), onnx_model_path, opset_version=11, input_names=[input_name], output_names=[output_name])

    ## openvino
    ie = IECore()
    net = ie.read_network(model=onnx_model_path)
    ovino_model = ie.load_network(network=net, device_name="CPU")
    pred = ovino_model.infer(inputs={input_name: x.numpy()})[output_name]

    ## onnxruntime
    session = ort.InferenceSession(onnx_model_path, providers=["CPUExecutionProvider"])
    ort_pred = session.run(None, {input_name: x.numpy().astype(np.int32)})[0]

    print(f"openvino: {pred.item()}\npytorch: {y.item()}\nonnxruntime: {ort_pred.item()}")
    assert(y.item() == zero_len) # pytorch
    assert(ort_pred.item() == zero_len) # onnxruntime
    assert(pred.item() == zero_len) # openvino

With the code above it will produce different result for openvino. But if the value of array_len is set to 15, it will produce the correct result. In general if array_len is set to >15 it will produce an incorrect result.

Edit: I have tested on openvino 2022.1.0 and did not find this problem

Issue submission checklist

• I report the issue, it’s not a question
• I checked the problem with documentation, FAQ, open issues, Stack Overflow, etc and have not found solution
• There is reproducer code and related data files: images, videos, models, etc.