I would like to show the RGB features learned in the first layer of a convolutional neural network similarly to this visualization of the same layer's features from AlexNet:
My learned weights are in the range [-1.1,1.1]. When I use
imshow in python or
imagesc in Matlab, the weight values are clipped to [0,1], leaving only positive weights intact, everything else black (obviously).
Negative weight values could be informative, so I don't want to clip them. Rescaling the weights to [0,1] works fine for grayscale features, but not for RGB features as it is unclear how negative values of a channel should be visualized. In the above picture 0 furthermore seems to map to the middle of the range (gray).
How are such RGB features visualized so that they look similarly to above AlexNet visualization?
(Sorry for the beginner's question.)