tensorflow-onnx icon indicating copy to clipboard operation
tensorflow-onnx copied to clipboard

Published 20 hours ago • verified publisher icon onnx

Inconsistency in conv+bn fusion and addition of useless reshapes in tf2onnx==1.16.1

Open riccardoalberghi opened this issue 1 year ago • 3 comments

Describe the bug When I convert a u-net architecture a Conv operator and a BN operator are not fused together and 2 Reshape operators are added which reshape to the same shape [-1, 512, 1, 1].

Urgency none

System information

  • OS Platform and Distribution: macos 14.2.1 arm64
  • TensorFlow Version: 2.13.1
  • Python version: 3.10.13
  • ONNX version: 1.15
  • ONNXRuntime version: 1.16
  • tf2onnx version: 1.16.1 (from github)

To Reproduce https://drive.google.com/file/d/1xzaAXWxUqXUTBVgDSiyrbjs_TSHXuhwW

by running python -m tf2onnx.convert --saved-model bg_model --output bg_model.onnx

Screenshots from the top the first conv+bn is fused correctly, then the following is not. Screenshot 2024-01-31 at 13 10 32 Screenshot 2024-01-31 at 13 12 21

Additional context Every other conv+bn is fused correctly except one.

riccardoalberghi avatar Jan 31 '24 12:01 riccardoalberghi

I conducted a little more testing. Downgrading to onnx==1.14 and tf2onnx==1.15.1/37820d solves the issue. With tf2onnx==1.16.0 the issue reappears, I dont know if this helps for identifying the problem.

riccardoalberghi avatar Jan 31 '24 18:01 riccardoalberghi

@fatcat-z any news?

riccardoalberghi avatar Feb 23 '24 09:02 riccardoalberghi

This is a regression bug brought by PR #2272 . @Rikyf3 , could you please share a code snippet of this module so I can prepare a test to cover it?

fatcat-z avatar Mar 10 '24 12:03 fatcat-z