GlobalAveragePool with ORT_ENABLE_ALL genenerates the incorrect output.

[SuhwanSong]

Describe the issue

Onnxruntime generates the incorrect output of the GlobalAveragePool layer when utilizing the "ORT_ENABLE_ALL" . The inference results diverge from the expected outcome achieved with "ORT_DISABLE_ALL".

To reproduce

poc.zip

1. Download and use the provided ONNX model (poc.onnx) and the following Python script.
2. Execute the Python script with the ONNX model using different optimization levels (ORT_DISABLE_ALL and ORT_ENABLE_ALL).
3. Observe the outputs and compare results.

import onnxruntime
import numpy as np

if __name__ == "__main__" :

    onnx_model_path = 'poc.onnx'

    rt_option_list = [onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL,
                      onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
                      ]

    rt_sess_option_list = []

    for level in rt_option_list:
        sess_options = onnxruntime.SessionOptions()
        sess_options.graph_optimization_level = level
        rt_sess_option_list.append(sess_options)


    ort_sessions = []
    for sess_options in rt_sess_option_list:
        try:
            ort_session = onnxruntime.InferenceSession(onnx_model_path, sess_options)
            ort_sessions.append(ort_session)
        except Exception as e:
            print (e)


    num_of_error = 0
    for _ in range(100):
        input_np = np.random.randn(1, 64, 13, 13).astype('f')
        input_ = {'fire9/squeeze1x1_1': input_np}

        outputs = []
        for ort_session in ort_sessions:
            try:
                ort_output  = ort_session.run(None, input_)
                outputs.append(np.array(ort_output[0]))
            except Exception as e:
                print (e)

        if outputs[0].argmax() != outputs[1].argmax():
            print (f'output {outputs[0].argmax()} != {outputs[1].argmax()}')
            num_of_error += 1

    print (f'result: {num_of_error} / 100')

Urgency

No response

Platform

Linux

OS Version

6.2.0-35-generic

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.17.3

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response