This is an extension of caffe that enables the calling of Halide functions as layers.

To test it please.

1. Install Halide
2. build caffe using cmake with the BUILD_halide option
3. run the short python script below

TODOs:

• [build] fix make build

• [build] add Halide to Travis + run layer tests
  The process of building layers is still quite elaborate and a lot of cmake code went into masking this complexity. That said I'm not a cmake expert, so I would greatly appreciate if others had a look at his too. The build server will have to updated to install Halide to tests to pass.

• Make layer zoo?

Once useful Halide layers start cropping up it would be nice to have a designated place to put them, together with some information on usage.

Further work for this is on hold till after the ICCV deadline.

Some additional things worth considering:

1. Halide Python Bindings
   Halide has python bindings which are a nicer user interface. Whats the cleanest way to provide access to these?

2. Automatic Backward pass
   Automatic backward functions for Halide would make it much easier to use. There has been some chatter about implementing this, but I don't know the current state of this.

This is an extension of @longjon prior work.

Simple Test script

import time
import numpy as np
import caffe
from caffe import layers as L

INSTALL_DIR = "."
HEIGHT, WIDTH = 200, 200

def create_neural_net(batch_size=1):
    library = INSTALL_DIR + "/lib/halide/libdevfunc_wrapper.so"
    netspec = caffe.NetSpec()
    netspec.data = L.DummyData(shape=[dict(dim=[batch_size, 1, WIDTH, HEIGHT])], ntop=1)
    netspec.halide  = L.DevFuncLayer(netspec.data)

    with open('net_halide.proto','w') as f:
        f.write(str(netspec.to_proto()))
        f.write('external_layers : "{0}"'.format(library))
        f.flush()
        net = caffe.Net(f.name, caffe.TEST)
    return net

if __name__=='__main__':
    caffe.set_mode_gpu()

    m = 3
    tmp = np.zeros( (1,1,HEIGHT,WIDTH))
    h_c = int(HEIGHT/2)
    w_c = int(WIDTH/2)
    tmp[0,0,h_c,w_c] = 1

    net = create_neural_net()

    net.blobs['data'].data[...]  = tmp
    startTime = int(round(time.time() * 1000))
    net.forward()
    endTime = int(round(time.time() * 1000))
    print(endTime - startTime,'ms')
    print(net.blobs["halide"].data[0,0,h_c-m:h_c+m,w_c-m:w_c+m])