May I ask how you obtained the checkpoint files for the diffusion model; ckpt.pth？

[xbyym]

Great work！Thank you @XzwHan @JegZheng for this wonderful contribution. I have read the complete paper and code, but I encountered a problem when training on my own. I used your provided snapshot_self_supervised.ckpt pre-training file, then proceeded with the training of the diffusion model, resulting in two checkpoint files: ckpt_best.pth and ckpt.pth. Then I used these checkpoint files for evaluation testing, and the results were very abnormal, as follows：

INFO - card_classification.py - 2023-12-12 18:39:20,375 - Apply default temperature scaling parameter T with a value of 0.1623 40it [10:36, 15.91s/it] INFO - card_classification.py - 2023-12-12 18:49:56,841 - Mean accuracy of all samples at test instance level is 9.8780%.

INFO - card_classification.py - 2023-12-12 18:49:56,841 - Now we compute metrics related to predicted probability quantiles for all classes... INFO - card_classification.py - 2023-12-12 18:49:56,843 - We predict the label by the class with the 1-th narrowest CI width, and obtain a test accuracy of 10.0900% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,845 - We predict the label by the class with the 2-th narrowest CI width, and obtain a test accuracy of 9.7700% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,846 - We predict the label by the class with the 3-th narrowest CI width, and obtain a test accuracy of 9.6300% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,847 - We predict the label by the class with the 4-th narrowest CI width, and obtain a test accuracy of 9.8100% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,848 - We predict the label by the class with the 5-th narrowest CI width, and obtain a test accuracy of 9.8200% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,850 - We predict the label by the class with the 6-th narrowest CI width, and obtain a test accuracy of 9.5800% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,851 - We predict the label by the class with the 7-th narrowest CI width, and obtain a test accuracy of 10.3100% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,852 - We predict the label by the class with the 8-th narrowest CI width, and obtain a test accuracy of 10.1700% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,852 - We predict the label by the class with the 9-th narrowest CI width, and obtain a test accuracy of 10.3200% through the entire test set. INFO - card_classification.py - 2023-12-12 18:49:56,853 - We predict the label by the class with the 10-th narrowest CI width, and obtain a test accuracy of 10.5000% through the entire test set.

My majority voting class prediction results are only at 10%, and I would like to know where my problem lies.

[xbyym]

I know you've provided your own checkpoint file: ckpt.pth. But how should I go about creating the right checkpoint files for my own pre-trained model, huh?

[xbyym]

Hello @XzwHan and @JegZheng, I previously posted this question and haven't received a response yet. I understand that you might be busy, but your insights on this matter would be really valuable to me. Any guidance or response you could provide would be greatly appreciated. Thanks for your time and consideration!