Higher learning rate than paper's
dungscout96 opened this issue · comments
We are using lr=3*10^-5 https://github.com/sccn/sound2meg/blob/db17dbf5055a45fb80001c52250ade6919b84494/reimplementation_training.py#LL149C57-L149C62
@makhalid1999 Can you reduce it to 3*10^-4 before running the experiment again?
Have updated that, and submitted to NSG, to see how it goes...