Could you provide a pytorch demo

[lll143653]

Hello Author @skokec ,
I am reading your paper and trying to understand the relevant codes, such as the translation of function in the paper, but the large number of cuda operations and some related codes based on caffe (caffe has not been maintained for a long time) make it difficult to read and connect with the formulas in the paper. Can you provide a simple pytorch demo?

best regards

[skokec]

There is also a pure TensorFlow version without any CUDA code available in https://github.com/skokec/DAU-ConvNet-TF repo. This code is quite a bit slow, but is much easier to understand.

Note that most of the equations in the paper relate to the efficient implementation and the computation of backpropagation, but in pure TensorFlow implementation this is done by automatic differentiation so you just need to provide gaussian kernel as weights to conv2d.

[lll143653]

thanks