Image In -> Image Out

[lolz0r]

Hello everyone, Is it possible to setup a network configuration where the input data is an image and the output data is ALSO an image, i.e teach the network to generate an image of equal dimensions as the input image?

The usecase is to utilize a NN to predict pixel depths given a number photographs of a scene.

Thanks, Michael

[antinucleon]

You need to modify the code:

1. if you have fullc in your net, add a reshape layer to reshape it back to [batch, channel, x, y]
2. modify the loss layer, to set gradient correctly.