HantingChen

Results 107 comments of HantingChen

How if you use the transposed conv layer?

I think this may caused by some optimization problems. Maybe you can try to adjust the hyper-parameters such as learning rate.

You can directly modify the Pytorch official imagenet training example: https://github.com/pytorch/examples/blob/master/imagenet/main.py, by replacing the conv-resnet with adder-resnet and using cosine learning rate decay with 150 epochs training.

p is reduced for each epoch, but I think it also works if you reduce it for each step.

可以,反卷积也是卷积的一种,只需要同样把反卷积中的乘法替换为加法和绝对值操作即可。

你好,如果是CNN会出现这种现象吗?

谢谢关注,暂时没有

1) 标准正态分布的绝对值方差为(1-2/π) 2) 见batch normalization原论文 3) 这种学习率使得每层的梯度的量级相同,所以更新步长也基本可以看作相同的。

> > > 您好,还想问下公式(9)中求出的是输出特征的方差Var(Y),公式(11)中用到的δB是特征的方差开平方,即标准差,可是文中公式(11)的下一行却说取Var(Y)=δB,这里是为什么?为什么要给Var(Y)取mini-batch的标准差呢?谢谢 你好,这里是写错了,应该是δB的平方,两个都代表方差。