mnist

[songhaohaohao]

Hello, I'm sorry to bother you, but I have some question, the mnist can be replaced with a pair of image?and can it be extracted to characteristics ?

[Zardinality]

I don't quite get it what you mean by characteristics, unless you are talking about feature extracted by previous layers, and also do you mean whether the MNIST can be replaced by another dataset? If so, the answer is yes, as long as you define it appropriately.

[songhaohaohao]

Thank you for your reply, I mean it can be replaced by an image ?I am trying to do an image as a data set, but i don't know the result is good without labels, and i want to extract the characteristic of the image by G(Z)

[songhaohaohao]

哈哈，我智障了，才发现可以用中文交流，你好，我的意思是我想要用一个256*256图像去代替MNIST，但是没有标签我不知道这个可以提取这个图片的特征么，因为我看中间是需要标签去生成的，所以不知道我的这个想法对不对

[Zardinality]

不需要标签。你指的标签，应该是真假分类的标签，数据集本身是不需要标签的。如果不明白这个问题，可以去看一下 GAN 的原理。 用更高分辨率的图像代替 MNIST 当然可行，只是你要保证数据集的接口一致，并且模型结构要进行相应的改变。 顺便一提，特征的英文在这里应该是 feature ，而不是 characteristic 。 在 GitHub 是提倡使用英文的，因为 issue 理论上并不是只有中文使用者可以看到，用英文能够帮助到更多有同样问题的人。

Labels of dataset are not required, you might mean the real/fake label used in building discriminator loss, which is generated automatically. Replacing MNIST with other dataset is feasible. As long as you keep the api of the dataset you use consistent with calls in main function, and add more layers in both D and G.