zoucheng1991
Results
2
comments of
zoucheng1991
I have the same question. The paper says only a lora is needed for domain-adaptation, but in this project the whole unet is updated during the image_finetune stage
loss = -torch.mean(F.logsigmoid(labels * logits)) / n according to paper, change "mean" to "sum" > Hi, > > Thanks for your great work. I was trying to apply siglip loss...