onnx2tf: [FastestDet] The absolute error of `AveragePool`, `GlobalAveragePool` is quite large
Issue Type
Others
onnx2tf version number
1.5.12
onnx version number
1.13.0
tensorflow version number
2.10.0
Download URL for ONNX
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/fastestdet.onnx https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/ppmattingv2_stdc1_human_480x640.onnx
Parameter Replacement JSON
N/A
Description
- Personal
- The absolute error of
AveragePoolis quite large. In addition, there is almost no error in other OPs.onnx2tf -i fastestdet.onnx -cotof -cotoa 1e-4
- At this time I do not know the cause of the larger error.
- I want to generate a tfite model with small accuracy degradation.
About this issue
- Original URL
- State: closed
- Created a year ago
- Comments: 15 (10 by maintainers)
The images above are outputs of average pooling 2d from tflite, onnx, and tflite output divided by onnx output. As shown in image, output is same except the padded edges. I think we can get same output by padding slightly more and slice useless part. By the way, what was the bug you mentioned in #121?
https://github.com/PINTO0309/onnx2tf/blob/27d635f5a962b14f9655aa48e5c5660083217cc2/onnx2tf/ops/AveragePool.py#L129-L134 It looks line 133 is not appropriate for pooling operation. When stride is larger than 1, this line is always True. After that line, padding mode is always ‘valid’.
Also, according to https://stackoverflow.com/questions/50543862/tensorflow-tf-nn-avg-pool-with-same-padding-does-not-average-over-padded-pi, tensorflow default action looks like
count_include_pad=False, same like onnx. But there is no option for user to change it. I guess, In this case, explicit padding and pad mode ‘valid’ caused output difference between onnx and tflite.Keep trying to fix, but this problem is quite complicated than I thought.