About val data

[holylone]

Hi, thanks for your amazing work.
I train the network with my own database, it works well.
However, I don't have so many data for training(just about 700 images).
My question is

1. How many data should I put in the val folder? Or, the val folder is just for testing the network?
2. Besides the paired data, I also have some unpaired data, for example, (Many B but lack of A, training direction: A2B), let me know if you are consider an algorithm for such semi-supervised situation.

Thanks.

[lxj616]

1. The val folder is just for testing the network
2. Try cyclegan https://github.com/junyanz/CycleGAN , BicycleGAN adds bidirectional cycle-consistency losses to "discover multiple modes"

[holylone]

@lxj616 Thanks for replying.
I tried cyclegan, the network can only generate one sample between domain A and B.
However, I want to generate many kinds of B from A, corresponding to diversity of B. The problem is a lot of B, but no A. Any idea of this?
I have another question, I was test BicycleGan on exist paired dataset. The generated B are diverse. But some images are quite strange like some net is added. Are there some way to deal with this?

|z|: 4 (Diversity is lost at |z|=2)

[junyanz]

See these two papers for the multimodal image-to-image translation in the unpaired case:
Augmented CycleGAN
MUNIT

[holylone]

@junyanz Thanks for replying.
The MUNIT didn't work very well.
The dataset A contains only binary images like circuit pattern. Because of the diversity of the dataset B, I think it difficult to train the network only use the unsupervised learning.
Maybe I need to train the unpaired data using the pretrained network trained by paired data.

Now I try to add some data augmentation to increase paired dataset.

[junyanz]

Adding paired data often helps constrain the problem. Good luck with your project.