search

ZK-Zhou / spikformer

ICLR 2023, Spikformer: When Spiking Neural Network Meets Transformer
MIT License
277 stars 43 forks source link

数据处理 #9

Closed whz-snn closed 1 year ago

whz-snn commented 1 year ago

从download->extract 的aedat格式是正常，本应该extract/xx/xx.aedat->events_up/xx/xx.npz，结果现在events_up/xx是空的文件夹，没有npz，尝试换台机器跑也是这个问题。

ZK-Zhou commented 1 year ago

You can delete the processed data folder and try again.

whz-snn commented 1 year ago

感谢答复，这个应该是spikingjelly的原因，我在它的issue中发现了同类问题。

还有一些看论文时的遇到的问题。

1.Table one 你对比了A_I 、A_RELU 、A_leakyRelu 、A_softmax 、A_SSA 其中前4个的QK是float V是spike，最后SSA的QKV都是spike，是这样的吧，这样让QK是float的意义是什么；

2.如果是这样的话，QK如何保持float，是有一个从输入直接到attention的分支吗，按照我的理解，经过了SPS模块的话，都会变成spike；

3.为什么要设置和relu对比的，softmax对比的实验可以理解，因为VSA是softmax，所以我理解的可能设置的实验是VSA的float QKV和SSA的spike QKV对比就可以说明SSA的显著优势？

-----原始邮件----- 发件人:"Zhaokun Zhou" @.> 发送时间:2023-06-02 22:23:27 (星期五) 收件人: ZK-Zhou/spikformer @.> 抄送: whz-snn @.>, Author @.> 主题: Re: [ZK-Zhou/spikformer] 数据处理 (Issue #9)

You can delete the processed data folder and try again.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

ZK-Zhou commented 1 year ago

Hi，

  1. Yep, Q, K, and V are all spike-form. In the first four variant, we test whether the float-point Q,K can improve the performance.
  2. Just cancel the spiking neuron in SSA we can get the float-point Q,K
  3. The more, The better :). No reviewer will criticize you for doing too many experiments