When using onnxruntime v1.14.1, all output values of Sigmoid are Nan.

[PINTO0309]

Describe the issue

When using onnxruntime v1.14.1, all output values of Sigmoid are Nan. However, if the Sigmoid is in the last layer, it works correctly, and if it is before the last layer, it becomes Nan.

It is inferred correctly in onnxruntime v1.13.1.

To reproduce

• onnx file with Sigmoid one before the last layer
  test_cut.onnx.zip
  
  • test code

    import numpy as np
    import onnxruntime as ort
    
    onnx_session = ort.InferenceSession(
        'test_cut.onnx',
        providers=[
            'CPUExecutionProvider'
        ],
    )
    
    model_input = onnx_session.get_inputs()[0]
    model_outputs = [output.name for output in onnx_session.get_outputs()]
    input_name = model_input.name
    input_shape = model_input.shape
    result = onnx_session.run(
        output_names=model_outputs,
        input_feed={input_name: np.ones([1], dtype=np.float32)},
    )
    for idx, (ret, name) in enumerate(zip(result, model_outputs)):
        print(f'{idx}. {name}, value:{ret}' if ret is None else f'{idx}. {name}, shape:{ret.shape}')

  • results

    0. /Unsqueeze_output_0, shape:(1, 1)
    1. /Mul_output_0, shape:(1, 64)
    2. /Sin_output_0, shape:(1, 64)
    3. /Cos_output_0, shape:(1, 64)
    4. /Concat_output_0, shape:(1, 128)
    5. /dense.0/Gemm_output_0, shape:(1, 512)
    6. /Sigmoid_output_0, value:None <-------------------------------- here
    7. /Mul_1_output_0, shape:(1, 512)
    

• onnx file with Sigmoid in the last layer
  test_cut_sigm.onnx.zip
  
  • test code

    import numpy as np
    import onnxruntime as ort
    
    onnx_session = ort.InferenceSession(
        'test_cut_sigm.onnx',
        providers=[
            'CPUExecutionProvider'
        ],
    )
    
    model_input = onnx_session.get_inputs()[0]
    model_outputs = [output.name for output in onnx_session.get_outputs()]
    input_name = model_input.name
    input_shape = model_input.shape
    result = onnx_session.run(
        output_names=model_outputs,
        input_feed={input_name: np.ones([1], dtype=np.float32)},
    )
    for idx, (ret, name) in enumerate(zip(result, model_outputs)):
        print(f'{idx}. {name}, value:{ret}' if ret is None else f'{idx}. {name}, shape:{ret.shape}')
    

  • results

    0. /Unsqueeze_output_0, shape:(1, 1)
    1. /Mul_output_0, shape:(1, 64)
    2. /Sin_output_0, shape:(1, 64)
    3. /Cos_output_0, shape:(1, 64)
    4. /Concat_output_0, shape:(1, 128)
    5. /dense.0/Gemm_output_0, shape:(1, 512)
    6. /Sigmoid_output_0, shape:(1, 512)
    

Urgency

There is a regression in ONNX Runtime functionality.

Platform

Linux

OS Version

Ubuntu 20.04

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

v1.14.1

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[PINTO0309]

This critical problem was resolved in onnxruntime v1.15.0. Thank you very much.