why we should multiply softmax result by 5

[Randylcy]

why we should multiply softmax result by 5. And there aren't code about using "decode" part. I will try it later, while if some one have done this part job, please share.

[HankKung]

I've been contacting the paper author and to the best of my knowledge, Decoder doesn't involve in the arch search. More accurate, they only search for encoder (feature extractor). Then attach ASPP after the encoder. The thing is, according to the author, the four ASPP output segmentation maps should be summed together instead of being concatenated as currently. I'll try to do this part recently.

[Randylcy]

我觉得decode part指的不是 decode 网络部分，这篇paper 所说的decoder， 是为了将cell的内部结构 和 外部 arch 结构进行可视化表示——人为可以看到，搜索出来的网络长什么样（通过print）。这部分代码，在 auto_deeplab.py中，cell decoder 部分我已经检查过了（代码中的 genotype(self)函数部分）；关于arch decoder（代码中的 decode_network (self) 部分），paper中说的是用的维特比算法（classic Viterbi algorithm），暂时还没有看懂。

[Randylcy]

I will try to sum the four ASPP output segmentation maps together, and see the MIoU.

[Randylcy]

I've been contacting the paper author and to the best of my knowledge, Decoder doesn't involve in the arch search. More accurate, they only search for encoder (feature extractor). Then attach ASPP after the encoder. The thing is, according to the author, the four ASPP output segmentation maps should be summed together instead of being concatenated as currently. I'll try to do this part recently.

I have changed the crop size to 321, and I have created a pull request (this work is included in cityscapes.py). BTW， I have changed ASPP cat to AsPP add， but when I set channel num = 40, it still runs into the "out of memory" problem! I wonder why we cann't reproduce the experiment in the paper. Now, I only can set batch size = 1, when channel number =30.