I hope you can help me

[userDJX]

Hello, thank you very much for your contribution. I recently simulated the code written in your paper (Pytorch), and the result is quite different from your paper. I hope you can help me find out what the problem is. (data set: cifar100)

Code link: https://github.com/sairin1202/BIC

My email address: [email protected]

[wuyuebupt]

@userDJX I take a quick look at their results and find two potential issues that might affect the results.

1. Results from the first 20 classes are lower than our results, e.g. 0.830 vs 0.852. You might want to improve this part as the first initialization model is important for the distillation in the first incremental training.

2. Although lower in the first 20 classes, the first incremental training (or the second batch training) shows a 0.699 result, which is much lower than our 0.7459. You might want to check this part in detail to see if any parts are wrong.

Once you get close results in the frist two batches of data, you should be able to reproduce the rest results as well.

I am sorry that I am busy with other projects recently (CVPR deadline is soon) and can not take a close look at their code in detail.

[userDJX]

Thank you for your Suggestions and I will modify it according to what you said. If you are free, I hope you can help us to have a look at the code. Wish you all the best in sending the CVPR paper.

[userDJX]

很抱歉，又来打扰您了，我查看了所有可能出现问题的地方，包括损失函数，样例集制作等等，但是都无法达到您论文中的效果，但是毫无疑问，您论文中提出的方法准确有效。 希望您能抽空看看 Https://github.com/sairin1202/BIC 这篇代码到底为何准确率不同 如果可以，希望您能分享一下这篇代码的Pytorch版本，万分感谢

[wuyuebupt]

@userDJX Hey, I make some changes to your code and get one close result. Some implementations for the distillation are still different. But I think your implementation is kind of fine. Check the pr I create for the repo.