Results not reproducible

[bhat-prashant]

Dear authors,

I am trying to reproduce your results on Seq-CIFAR100 with multiple runs with different seeds.
I see some of the issues in your code. The CL model has three heads namely: one for SSL, one for classification and one for rotation prediction. However, the paper does not talk about the rotation prediction. Could you give more intuition how rotation prediction actually helps and how much is the boost coming from such an auxiliary task? Please let me know.

Without the rotation prediction, results drop by quite a lot and are on par with DER++. Please let me know how this can be fixed.

[lijincm]

Please refer to our reference [Ji et al., 2021a] Zhong Ji, Jin Li, QiangWang, et al. Complementary calibration: Boosting general continual learning with collaborative distillation and self-supervision. arXiv preprint arXiv: 2109.02426, 2021. https://arxiv.org/abs/2109.02426
The latest version will be released after acceptance.