When my model is conv3d, avgpool3d seems to have some issues

[jungyin]

Issue Type

Feature Request

OS

Linux

onnx2tf version number

1.26.2

onnx version number

1.16.1

onnxruntime version number

1.18.1

onnxsim (onnx_simplifier) version number

0.4.36

tensorflow version number

2.17.0

Download URL for ONNX

https://1drv.ms/u/c/4008ad016f04b6a0/ET1fYPRg0fxNk9NAPr870pkBcKNHr_xKjm7wZF5Ii2hXLg?e=nKf0KE

Parameter Replacement JSON

{
    "inputs": [
        {
            "name": "input_0",
            "shape": [
                1,
                3,
                150,
                128,
                128
            ]
        }
    ],
    "outputs": [
        {
            "name": "face_PPG",
            "shape": [
                1,
                150
            ]
        },
        {
            "name": "finger_PPG",
            "shape": [
                1,
                150
            ]
        }
    ]
}

Description

My model structure is conv3d, and its input type is ncdhw. When I try to convert the onnx model to tflite, there is always an error in the screenshot. After changing it to a fixed value, it can export the model normally. I am not sure if this modification is correct. Can you provide me with some guidance

[PINTO0309]

onnx file

[jungyin]

onnx 文件

i was update https://1drv.ms/u/c/4008ad016f04b6a0/ET1fYPRg0fxNk9NAPr870pkBcKNHr_xKjm7wZF5Ii2hXLg?e=nKf0KE

[PINTO0309]

TensorFlow cannot pool properly when count_include_pad = 1. see: https://github.com/PINTO0309/onnx2tf/issues/124

https://github.com/onnx/onnx/blob/main/docs/Changelog.md#averagepool-11

• AveragePool-11

  

[jungyin]

TensorFlow cannot pool properly when count_include_pad = 1. see: #124

https://github.com/onnx/onnx/blob/main/docs/Changelog.md#averagepool-11

• AveragePool-11

It seems that this approach requires adjustments to the existing model, and such adjustments may require me to retrain the model? Besides, you can take a look. After I made changes to the source code, this error will pause the report. Is this modification correct? Looking forward to your reply, thank you very much!

[PINTO0309]

Check the accuracy of the model by specifying the -cotof option. It is probably incredibly inaccurate.

It seems that this approach requires adjustments to the existing model, and such adjustments may require me to retrain the model?

Unfortunately, that's right.

[jungyin]

Check the accuracy of the model by specifying the option. It is probably incredibly inaccurate.-cotof

It seems that this approach requires adjustments to the existing model, and such adjustments may require me to retrain the model?

Unfortunately, that's right.

i was try add -cotof,it print this，maby it is surcess work?

[PINTO0309]

The conversion was completely successful.

However, AvgPool has a very large number of edge cases. This time it just happened to work, so you need to be careful that it won't always be successful.

[github-actions[bot]]

If there is no activity within the next two days, this issue will be closed automatically.