This project is a alexnet forward pass implmentation using Arrayfire, done at the API level.

edit the py_caffe variable in caffe_util.py to point to your caffe/python directory

python main.py

I'm switching the order of tuples in the code from (input, channel, x, y) to (x, y, channel, input) as arrayfire stores arrays colun-major. You can call a function with an image in the form image = (x, y, channel) but internally it gets switched around a lot, since I didn't know what I was doing when I started

The convolution kernel is very slow. It's a very naive implementation, but I didn't really know what was going on at first. Should probably rely on the already written af convolve kernel if possible.