You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
分类数目变大,虽然可以将分类层参数拆分到各个GPU上,但是各个GPU上的隐层特征allgather也带来显存消耗。随着分类层数目变多,虽然可以通过增加GPU数量来保证fc层参数分配到各个GPU上的显存是一个常数,但是隐层特征x,还是会随着GPU数增加而增加。单卡显存有限,这样也限制仅仅通过增加GPU数量来应对分类数量线性增长。这个问题在论文“Partial FC: training 10 million identities on a single machine”提出。
The text was updated successfully, but these errors were encountered:
分类数目变大,虽然可以将分类层参数拆分到各个GPU上,但是各个GPU上的隐层特征allgather也带来显存消耗。随着分类层数目变多,虽然可以通过增加GPU数量来保证fc层参数分配到各个GPU上的显存是一个常数,但是隐层特征x,还是会随着GPU数增加而增加。单卡显存有限,这样也限制仅仅通过增加GPU数量来应对分类数量线性增长。这个问题在论文“Partial FC: training 10 million identities on a single machine”提出。
The text was updated successfully, but these errors were encountered: