vipermdl / arl Goto Github PK

Hey there,

I'm not completely sure Softmax2d does exactly what you what (apply the softmax spatially).
I've implemented this paper by doing using that function at the start, but later I found out that that was a mistake.

I believe that we should be doing this:

1. Add a softmaxLayer to every bottleneck such that every bottleneck has a property

self.softmaxLayer = nn.Softmax(2) #it will apply softmax along the third dimension (we use 2 because we start indexing from 0)

and later override the forward function by doing this:

 out = out + (identity 
                     * self.softmaxLayer(out.view(*out.size()[:2], -1)).view_as(out) 
                     * self.alpha) + (identity)

Tell me what you think

因为中间那个数据集链接进不去 所以我是自己下载的数据集 但可能自己比较小白 真的找不出问题在哪
如果有时间的话 谢谢解答了

Hello,
I felt confued about your BasicBlock module:

class BasicBlock(nn.Module):
    expansion = 1
    def __init__(self, inplanes, planes, stride=1, downsample=None, alpha =0.001):
        ...
    def forward(self, x):
        ...
        out = out + residual + self.alpha * attention * residual

if in_channels != out_channels, the last line can running?
residual: inplanes, out: planes

Looking forward to your reply. Thanks.