How to perform 2D convolution with user defined kernel just like sobel filter?

[chowkamlee81]

How to perform 2D convolution with user defined kernel just like sobel filter? Kindly provide example

[ZiyueHuang]

You can fix the weight of convolution.

[thomelane]

With Gluon API you can set_data on a parameter. So for a convolutional layer you can select the weight (which will be the kernel) and call set_data on this with the kernel of your choosing. You must make sure you have the correct shape, with the dimensions being (filter, channel, height, width). So if we wanted a single sobel kernel applied to a single channel input we could for the following:

import mxnet as mx

# kernel of your choosing (e.g. a sobel filter)
kernel = mx.nd.array(((1,0,-1),
                      (2,0,-2),
                      (1,0,-1)))
# add dimensions for filters, and input channel
# both of which will have shape 1, since single filter and single input channel
weight = kernel.expand_dims(0).expand_dims(0)

conv = mx.gluon.nn.Conv2D(channels=1, kernel_size=(3,3), padding=(1,1))
conv._in_channels = 1
conv.initialize()
conv.weight.set_data(weight)

[sojiadeshina]

@marcoabreu can you close out this issue since the question has been answered with example by @thomelane