ramos
ramos
在声纹识别的分类训练任务上，我用的是我自己实现的circle loss, gamma=64, m=0.3, 之前的训练收敛速度不是很理想，但是使用cos annealing lr schedule以及balanced data sampler似乎可以加快收敛。目前结果还未出来，完成实验我来贴一下结果
This problem is caused by initializing OOM, since the model opt-66B should occupy around 130G gpu memory which way surpasses the A100 memory. I have updated a shardinit option in...
> Hi @nemoramo Thanks for your contribution. But it fails in CI, could you please fix it? Thanks. It seems no coverage report is included. Any ideas to fix this?
> Hi, @nemoramo , we are currently working on that. Are you interested to be part of the development? sure
Same problem. The links are not available right now. @ankurdhuriya