Training tends to yellowish colors

[i-radwan]

Hi, Thanks for this Pytorch implementation. I'm following your exact tutorial to train the model on my Mac machine CPU, the problem is that as I monitor the visdom console, I see the output fake_reg image goes more yellowish as the losses curve goes down. I've had this same exact issue while training the original colorization model (i.e. the one without hints) introduced by the paper authors. I don't know why this actually happens, any ideas?

Thank you!

[richzhang]

I pushed an update, which "pretrains" the net on just 10 of the categories. See issue #3 .