Dataset preprocessing problems.

[alexaex]

Hi, I am working on my thesis in image colorization. I noticed your implementation of dataset processing, where you use OpenCV to read images and skimage for LAB color space conversion. I have implemented a similar approach but only used the L-channel as the input to the network. However, I found that skimage uses uint8 for the L-channel, but the network implicitly converts the L-channel to float. The output of the ab-channels is also in float, so I normalized both the L and ab channels. However, the prediction results in a black image, which might be caused by the uint-to-float conversion.

I noticed that you used the L-channel and padded it with two zero ab channels, then converted it back into RGB as the input to the network. I would like to know if you have any suggestions if we just want to use the L-channel as the input.
@piddnad