About contribute densnet to Mxnet

[qingzhouzhen]

Hi xiong, I used your densnet core code and fit it to the form as incubator-mxnet/example/image-classification/symbol, then I write the train code for distribute train(because memory of my GPUs is not large enough), the results as below:

Sorry for this picture, I cannot upload file beyound 10K,
my code refer to https://github.com/qingzhouzhen/incubator-mxnet/blob/densenet-symbol/example/image-classification/symbols/densenet.py
Would you mind I push it to mxnet? I would describle my training indetail

[bruinxiong]

Hi qing, sorry, that picture is too blur that I can not see the detailed information. Open-sourcing is a good thing, I'm very happy to see my code can help you to understand densenet architecture. If you want push your code to mxnet, I only have one request, please refer and mention my code or my link when you push to mxnet. Thanks!

[shiyuanyin]

@bruinxiong
作者你好，我想改动一个新的结构，是在SE的地方改动的，有点困惑，mxnet 的symbol，不能直接得到bchw的值，需要根据当前bn3的输出得到bchw,然后reshape,操作，可是symbol只能在喂数据后才能得到每一层的形状
pytorch 的SGE，一个实现架构语句， 对应你提供的模型SE代码位置修改的话，symbol每一层bn3 后边的bchw,我直接得不到，我要mxnet,实现这句话，b, c, h, w = x.size()， x = x.reshape(b * self.groups, -1, h, w) 我对mxnet 不是那么熟悉，不知道作者你有没有好的方式实现这句reshape

下面是对应pytorch 实现

class SpatialGroupEnhance(nn.Module): # 3 2 1 hw is half, 311 is same size
def init(self, groups = 64):
super(SpatialGroupEnhance, self).init()
self.groups = groups
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.weight = Parameter(torch.zeros(1, groups, 1, 1))
self.bias = Parameter(torch.ones(1, groups, 1, 1))
self.sig = nn.Sigmoid()

def forward(self, x): # (b, c, h, w)
b, c, h, w = x.size()
x = x.view(b * self.groups, -1, h, w) ##reshape
xn = x * self.avg_pool(x) # x * global pooling(h,w change 1)
xn = xn.sum(dim=1, keepdim=True) #(b,1,h,w)
t = xn.view(b * self.groups, -1)
t = t - t.mean(dim=1, keepdim=True)
std = t.std(dim=1, keepdim=True) + 1e-5
t = t / std # normalize -mean/std
t = t.view(b, self.groups, h, w)
t = t * self.weight + self.bias
t = t.view(b * self.groups, 1, h, w)
x = x * self.sig(t) #in order to sigmod facter,this is group factor (0-1)
x = x.view(b, c, h, w) #get to varying degrees of importance,Restoration dimension
return x

[liuhangfan]

@shiyuanyin 你好，请问你用mxnet实现了吗？我现在也遇到了这个问题

[sky186]

@liuhangfan
https://github.com/sky186/sge-srm_mxnet

[bruinxiong]

@shiyuanyin @liuhangfan @sky186 你们好，很久没有关注这个帖子了，抱歉才看到。已经转战TF平台了，我目前没有资源做进一步尝试。如果后面有的话，我会试试。不过在这之前，可以参考 @sky186 这位同学的实现。感谢 @sky186 同学的回复，我抽空看看你的代码哈。非常感谢大家！