Merge convolution and batchnorm layers in both Caffe and PyTorch using Python
Basic idea can be found in my CSDN Blog:
Caffe merge bn with equations and explanations
I wrote the Caffe merge script a long time ago while Pytorch merge script recently, thus may have many difference between them.
Both of them have been tested yet.
Before using either of them, you should have a no-bn version model first, it means you should have a caffe network prototxt without bn or a pytorch model with all bn layers commented.
After prepare your model and weights, you should modify the path and imports in python scripts
Finally, just run and merge~
BTW, i offered a no-bn version shufflenetv2 model for example