Why did you choose the last num_outputs is 12 in the network?

[Keying96]

conv10 = slim.conv2d(conv9, 12, [1, 1], rate=1, activation_fn=None, scope='g_conv10')
out = tf.depth_to_space(conv10, 2)

As stated above, in the last part of the network you set conv10's num_outputs is 12 and then do depth_to_space(conv10, 2) to make the channel of out to 3.
Now, I want to know why you didn't set conv10's num_outputs 3 directly.

[manfred-git]

I think the reason is that conv10's output size is H/2*W/2 for each channel, and in order to get an H*W-size 3-channel output RGB image, you need 3*(H*W)/(H/2*W/2) = 12 channels output.

[purva98]

Hey! Is there any alternative to using this :
out = tf.depth_to_space(conv10, 2)
thanks