A colorization framework that disentangles the color multimodality and the structural consistency via adaptively located anchors, so that both aspects can be achieved effectively. [SIGGRAPH Asia 2022]
Hi,
I am trying to finetune DISCO model on other color image datasets. The training settings are kept as default as on ./scripts/anchorcolorprob_hint2class-enhanced-h8.sh. After several loops, "nan" value occurs and the training losses become nan too. I checked the intermediate variables. The "nan" starts after the execution of "totalLoss_idx.backward()". Would you please give me some suggestion? This error usually caused by gradient explosion or gradient vanishing. I tried gradient clipping using torch.nn.utils.clip_grad_norm(model.parameters(), args.clip), and didn't work. Looking forward to your reply.
Hi,
Wonderful work!
As stated on L.35 in ./utils/dataset_lab.py, the ab channels are scaled by 110. instead of 128. Would you please explane the reason? In OPENCV, LAB space should be defined as L -> 0100, a -> -128+128, b-> -128~+128. Thanks again for your wonderful work.
thank you for your awesome work!
I felt confused about superpixel modle “SPixelNet-s16” and "SPixelNet-s8".
how to use it if i want to test colorize and superpixel performance?