So I've got the path to the images hardcoded in solver/train_val.prototxt, that'll need to be changed. The .txt files in data folder have the names of the images. One thing to note is that for some reason, there's one batch of images missing - check doit.py. If you're gonna mess with the ITS, it doesn't matter as all that will have to be modified anyway.

Basically using AOD framework. Just to try to get something functional. What was provided on their git repo was just for deployment, not training. My modifications are to try to train. I figure we should be training against un-hazed images so you can see that I have two extra input layers for those. They're then used in the loss layer at the end. This results in enormous loss values, so something is definitely not right. My suspicion right now is that the images need to be transposed before input as you'll notice that's a common theme for this stuff. Not sure how to do that when feeding images using the .txt file stuff. Then, I could just be totally off. Anyway...