langfengQ / MLF-DSResNet

This is the implementation of MLF & spiking DS-ResNet
MIT License
11 stars 1 forks source link

超参数的设置 #3

Closed 2ephyrus closed 8 months ago

2ephyrus commented 8 months ago

作者您好, 请问膜时间常数和阈值这些超参数是如何设置初始值的? 另外,您能否提供针对本文来讲,这些超参数的限制区间?以及,这些超参数是否具有继续训练的价值?(您预设的初始值已经逼近最优解了吗?)

2ephyrus commented 8 months ago

image

langfengQ commented 8 months ago

没有调节过超参数,直接使用的tdBN(Going Deeper With Directly-Trained Larger Spiking Neural Networks)的超参数。除了Vth。

2ephyrus commented 8 months ago

请问Vth的取值有什么范围吗?是否有些取值会导致训练梯度消失或爆炸?

langfengQ commented 8 months ago

我们在工作中没有过多探索过Vth的影响,我们在工作中设置过0.5-0.7,感觉性能对此并不敏感,因此设置合理范围应该都可以。

可以参考Direct training for spiking neural networks: Faster, larger, better

langfengQ commented 8 months ago

image

2ephyrus commented 8 months ago

好,谢谢作者