search

junyanz / CycleGAN

Software that can generate photos from paintings, turn horses into zebras, perform style transfer, and more.
Other
12.23k stars 1.93k forks source link

Is the performance good or bad? #85

Closed zhuqunxi closed 6 years ago

zhuqunxi commented 6 years ago

After training 140 epochs, 1000 images for each epoch, I got the image from horse to zebra. It seems not bad but also not good compared to the performance in the original paper (at least, the grass in the picture vanished). So, I am wondering in which processing may exist a bug or how can I improve its performance？I will appreciate it if you can help me. Thank you. image image the following picture shows the error from the latest epochs. It seems the optimization is no longer forward. image

junyanz commented 6 years ago

This looks reasonable. You have some checkboard artifacts. The artifacts will go away if you train the model for more iterations with a smaller learning rate. You can also use upsampleConv to replace Deconv as pointed out by this post.

zhuqunxi commented 6 years ago

Great thanks for your help. I'll try it.