search

knazeri / edge-connect

EdgeConnect: Structure Guided Image Inpainting using Edge Prediction, ICCV 2019 https://arxiv.org/abs/1901.00212
http://openaccess.thecvf.com/content_ICCVW_2019/html/AIM/Nazeri_EdgeConnect_Structure_Guided_Image_Inpainting_using_Edge_Prediction_ICCVW_2019_paper.html
Other
2.5k stars 530 forks source link

learning rate policy used in image inpainting #59

Closed zengyh1900 closed 5 years ago

zengyh1900 commented 5 years ago

Hi @knazeri , I have a question that, have you ever tried any learning rate policy in image inpainting model training? Such as cosine decay, step policy etc. Do you think these policy can help better convergence in such models' training?

Besides, thank you for your contributions! I have benefited a lot from this repo!

knazeri commented 5 years ago

@1900zyh You are very welcome.

We tried exponential learning rate decay in the beginning; however, later we didn't find it much effective and instead stuck with our fixed decay policy as was explained in the paper. Having said that, we didn't try other decay methods and I'm interested to see how other decay policies affect the convergence!

2018hello commented 5 years ago

It is a great work! Thank you so much! @knazeri And I have a question, how to plot the losses values in this project? Because I want to know when will the model become converge. Can you give some suggestions?

knazeri commented 5 years ago

@2018hello You can use tensorboardX which lets you visualize PyTorch code in Tensorboard. You can also write your custom code to plot loss values using the training log file. We save every loss value in a log file. you can change LOG_INTERVAL to set log saving interval!

2018hello commented 5 years ago

Thank you so much! I can plot the loss image!