Hello, I was amazed to see the incredible results you've achieved with the VAR model. I sincerely want to ask you about some training details because I'm trying to train the base model (16, 256*256) for category-specific generation on the official website.
However, after switching datasets to tiny-imagenet(200 classes, 100000 images), my training results were bad. This has left me quite puzzled, so I'm looking forward to any insights you might share, such as which default parameters might need adjustment during training or any other important training tips you think are crucial.
Thank you so much for your help!
Here are my results using all default parameters with 200 epochs. The final converged loss and accuracy are:
Lm: 6.082 Lt: 5.355
Acc m&t: 4.92 7.55
![Screenshot from 2024-07-22 00-46-37](https://private-user-images.githubusercontent.com/55437654/350783514-08b304bf-142f-402c-8081-a6f12bbacf59.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.OKz1uvwtZUjON0C-HkxviqotcaUQ7yyTdMlXZURkHQ4)