I trained so poorly？

[goldwater668]

100 pairs of data sets with 1024X1024 resolution are trained according to 512, and the pixel2pixelHD inference is used. Why is the result so poor? Do you know why? The data set pairing adopted is here :https://github.com/Sxela/face2comics
image

[takuyaliu]

It may be because the instance map and feature map are not used.

[goldwater668]

@takuyaliu

How to extract instance map and feature map？

Changing to a different data set does not require map features for training, and the effect is ok

[takuyaliu]

@takuyaliu

How to extract instance map and feature map？

Changing to a different data set does not require map features for training, and the effect is ok

You can find the way to use imap and fmap on Google.

[goldwater668]

@takuyaliu
1.On another set of data sets with better quality, I didn't use map, and the effect is not bad. How to explain it?
2.
For example, to stylize the face, how should the map be set? Divide the data face map and background map of train_A into label_0 and label_1?

[takuyaliu]

@takuyaliu 1.On another set of data sets with better quality, I didn't use map, and the effect is not bad. How to explain it? 2. For example, to stylize the face, how should the map be set? Divide the data face map and background map of train_A into label_0 and label_1?

I'm sorry I have no idea about your question1, just Google it.

As to the instance and feature map, you can see this https://github.com/mcordts/cityscapesScripts.
About q2, you may not only need to partition facemaps and backgroundmaps, but you also need to partition nosemaps, eyesmaps, etc.

The above are just my simple opinions, you can give it a try.