Dear Markus,

I've just came across your old article on fusing batch normalization and convolution.

I play with the code you provided, and spot that the last line (of computing the new bias) has a small typo.
In particular, according to your article, one should also left multiply b_\text{conv} with W_\text{bn}.
Please see the gist for this correction: https://gist.github.com/heytitle/34d2c3f32708d512132534965f0c220d#file-fused-batchnorm-py-L26.

In the example you used, the original code doesn't introduce any error because there is no bias in the first convolution of Resnet18.

Lastly, thank you for writing this nice article!

Cheers,
Pat