Add groups parameter to nengo.Convolution

[Michaeljurado42]

nengo.Convolution supports many of the same arguments and options as its keras counterpart. However there is no groups parameter in nengo as in the keras docs for Conv2D.

A groups paramter would be useful to allow for more robust conversions between keras networks and nengo networks via the dl converter. This would also allow people to implement depthwise convolutions with ease also in nengo.

[Michaeljurado42]

It seems to fix this method you have to add a groups parameter to the conv2d function found here

This is my fix:

def conv2d(x, w, pad='SAME', stride=(1, 1), groups = 1):
    ksize = w.shape[:2]
    x = extract_sliding_windows(x, ksize, pad, stride)
    ws = w.shape
    w = w.reshape([ws[0] * ws[1] * ws[2], ws[3]])
    xs = x.shape

    if groups == 1:
        x = x.reshape([xs[0] * xs[1] * xs[2], -1])
    else:
         x = x.reshape([xs[0] * xs[1] * xs[2], groups, -1])  # seperate input into groups

    y = x.dot(w)  # convolve all
    if groups > 1:
        y = np.sum(y, -2)  # sum up along groups
    y = y.reshape([xs[0], xs[1], xs[2], -1])
    return y

Then you have to manually edit or extend ConvInc found here to use the groups parameter

You have to modify the convolution transform found here and register it with the Builder.

I may edit and elaborate on this process in the coming weeks

[Michaeljurado42]

Sorry I did not mean to close this issue