fuse_bn_into_conv optimizer in pypi linux wheels is wrong

[daquexian]

Reference: daquexian/onnx-simplifier#50

Here are some onnx models, where MBV3_SSD_onnx.onnx is the original model, mbv3_ssd_onnx_sim_macos.onnx is the model after optimizing and simplifying on macOS, MBV3_SSD_onnx_sim_onnx_16_pip.onnx is the model after optimizing and simplifying on ubuntu.

The "1523" tensors of the last two models differ a lot, which is caused by the onnx fuse_bn_into_conv optimizer. The tensor in ubuntu model (no matter produced by onnx 1.5 or 1.6) is totally wrong while the macOS version is right.

However, if I compiled and install onnx from source on ubuntu, the optimized model is correct and the same as that on macOS.

[askhade]

@daquexian : Is this still a problem in the latest onnx version?

[daquexian]

@askhade yes, I just verified it still exits on onnx 1.7.0 installed by pip3 install onnx==1.7.0 on linux. what's more, if I installed onnx by pip3 install onnx==1.7.0 --no-binary onnx (which builds onnx from source instead of installing the prebuilt binary), the problem disappears

[daquexian]

I believe it has been fixed by #3021