I attempted to use 50 population sizes and 20 iterations to search in both Autoformer-Small and Base spaces according to the source code. However, it's challenging to reproduce the subnet performance as stated in the paper.
For instance, the subnet performance within 30M in the Small space mostly concentrates around 81.7 accuracy (compared to the 82 accuracy in the original paper) , and this phenomenon is more pronounced in the Autoformer-Base space. Could you please provide the random seed and the upper and lower limit parameters for each search to facilitate the reproduction of the search results? Thank you.
I attempted to use 50 population sizes and 20 iterations to search in both Autoformer-Small and Base spaces according to the source code. However, it's challenging to reproduce the subnet performance as stated in the paper. For instance, the subnet performance within 30M in the Small space mostly concentrates around 81.7 accuracy (compared to the 82 accuracy in the original paper) , and this phenomenon is more pronounced in the Autoformer-Base space. Could you please provide the random seed and the upper and lower limit parameters for each search to facilitate the reproduction of the search results? Thank you.