Reproducing results for AR1* with 500 latent replay size

[WeiWeic6222848]

Dear Mr. Lomonaco,

Thank you for making the code of AR1* publicly available and re-implement it in pytorch.
I'm trying to recreate the result of Figure 4 of the paper https://arxiv.org/pdf/1912.01100.
The setting is NICv2-391 with replay buffer size 500.

I have kept all the parameter of params.cfg the same and only changed rm_rz from 1500 to 500, the result is an accuracy of 63.62%, specifically: Accuracy: 0.6362827923521566.

This is far from the reported ~70% as shown below.

Is there some parameter I should change to recreate the results from the paper?

Thank you in advance,

Best regards,
Wei