While the num_classes isn't 1000 , some error in sample.py!

[Y1ZzLu]

Q1:when using single gpu train model , model.module is not available
Q2: sample.py y_null = torch.tensor([1000] * n, device=device) should be y_null = torch.tensor([args.num_classes] * n, device=device)
Q3:after training,i use my trained model to generate data,the sample.png looks like a gauss noise,my train loss is below 0.05! but i cant use this model !

[jizaza]

Q1: you can just use model when using single gpu
Q2: In my opinion, n means the number of classes, 1000 maybe just a constant
Do you have any more ideas? Welcome to discuss!
Q3: you can change the args.num-sampling-steps

[Y1ZzLu]

@jizaza Thanks for your reply, I have solved the questions!

[racoonykc]

thanks!I found the same guassion noise

[Y1ZzLu]

@racoonykc By using a fine-tuning method to transfer learning from imagenet dataset to the target dataset, I use Lora fine-tuning provided by PEFT