pytorch-SegInpaint

pytorch-SegInpaint copied to clipboard

Deeplabv3+ or FCN? How to train SP-Net and SG-Net 200 Epochs first, and then train 100 Epochs together

1

你好，为什么架构图中和代码里是FCN网络，但是论文中是deeplabv3+网络。并且，如何先分别训练200个Epochs的SP-Net和SG-Net，再联合训练100次。 代码中貌似是直接联合训练100次

isgangzi

Hi, thanks for your sharing code. I have noticed that the label map input of SP-Net is one hot, have you try to replace it with the softmax output of...

super233

用你提供的代码 我在CelebA上进行了训练，我发现测试结果非常糟糕

2

   

wenshinlee

nunu-chong

In paper, you use two dataset, cityscapes and helen face. however, in your code, there are not the define about helen dataset.

ah-ke

Semantic FCN Seems Lost

2

I think your code indicates that the module consists of SPN and SGN. However, the very beginning module of FCN does not appear in the model's structure. Therefore, the condition...

YIYANGCAI

Worship!!

1

HOLY SHIT!!! THIS IS SO FUCKING BRILLIANT!!!!!!

hubert0527

老哥，这篇论文真的可行吗？

1

我毕设就是采用它的思想，发现根本不可行啊，对破损的图像先进行语义分割，破损根本就和RGB原图不对应啊，怪不得发表不了

DequanZhu

Do you have test code?

1

As the title

jialiang66