spiking-fullsubnet icon indicating copy to clipboard operation
spiking-fullsubnet copied to clipboard

Published 20 hours ago • verified publisher icon haoxiangsnr

训练非常慢

Open shenbuguanni opened this issue 1 year ago • 3 comments

你好，我设置的batch_size=12，音频长度4s,显卡是4090,训练非常慢，请问是不是GSN本身的算力非常大？

shenbuguanni avatar Feb 19 '24 01:02 shenbuguanni

你好，我设置的batch_size=12，音频长度4s,显卡是4090,训练非常慢，请问是不是GSN本身的算力非常大？

你好，请问训练的速度大概是怎样的

xxh0515 avatar Feb 20 '24 07:02 xxh0515

你好，请问训练的速度大概是怎样的

5s一个step吧，正常训练一个降噪模型大概是3steps每s，应该是GSN里面的大矩阵操作太多了？很难训

shenbuguanni avatar Feb 20 '24 07:02 shenbuguanni

Hi, @shenbuguanni @xxh0515 . Thank you for your attention.

For the current track (N-DNS track 1), we only have a naive GSN implementation that doesn't support cuda-level GPU acceleration. Hopefully, in the track 2, we will implement it using the Lava Platform, which supports acceleration by the Intel Loihi 2 chip.

haoxiangsnr avatar Feb 20 '24 13:02 haoxiangsnr