search

liruilong940607 / glcic.pytorch

This is a implement of the Siggraph2017 paper: "Globally and Locally Consistent Image Completion"
15 stars 7 forks source link

Black frame generated after server epochs #1

Closed miaojiang1987 closed 5 years ago

miaojiang1987 commented 5 years ago

Dear Ruilong Li,

Thank you for implementing this. It is excellent work to reproduce the training code with python.

However, I found two issues after I trained the model several epochs. I trained with my own data with initial model completionnet_places2.t7. On the first of the epochs, it would do an OK job, however, there is a gray frame covered.

This image is generated from my own data set. And you can see a grey box is covered on top of that. image

After several epochs, the completion part is totally black. Could you explain why? Is it because overfiting?

Here is an artifact generated after 10 epochs. You can see a black box in the image. This is generated from the Places dataset. image

Again, thank you for your work and I really appreciate your hard work. Many thanks to your reply.

Best, Miao Jiang

liruilong940607 commented 5 years ago

Thank you for watching.

This repo is an old version with some bugs. I've just updated this repo to the newest. But I'm sorry to say I don't think I successfully reproduce the author's result, although the code seems fine now.

  1. The result of pretrained model seems ok. The author of this paper used a blend postprocess to fade this edge problem. You can find the codes of this part in blend.py

  2. I also just try this on my own dataset. Here are some good & bad results. I think the good results prove the codes are ok now, but the bad results indicate this repo need improvement.

image image

miaojiang1987 commented 5 years ago

Thanks a lot! Ruilong! It is really nice of you for helping me with the issue. Please ignore my email sent to you. I will test it again.