I'm trying to create a custom regression output for my CNN. The basic premise is I take an image with various point-spread functions, and output an image with "spikes" at the center of those PSFs, and zero everywhere else.
I input and output images, so my loss function is an L2 loss function based off the different pixel values. To calculate this, I convolve the images with a 2D-Gaussian filter
L = ||T * g - Y * g||^2 + L1; where * denotes the convolution operator, T is the desired output, and Y is the output from the neural net.
My issue is that Y and T are dlarray objects, and imgaussfilt is incompatible with them. Is there a way around this?