Closed XueZ-phd closed 4 years ago
Hi, Glad you find the code useful!
I experimented with a the top-k pooling, however couldn't get it to work probably, so I deleted it (at least I thought so) before making the repo public. Normal max-pooling is just top-1 pooling, so you can either delete k or set k=1. I will update the repo with this change, when I have more time.
Yes, I had all the images in the same folder. Such that you get a list of all you images with line https://github.com/FrederikWarburg/Burst-Image-Deblurring/blob/1359093179ebca9c1d53fb55590e70bc35844f8d/tools.py#L141
I don't have the weights. Sorry about this. But since this repo have gained so much attention, I will definitely upload the weights soon (probably in the start of February, as I'm busy with my thesis and a paper this month).
Good luck with the code. Let me know if you have other quick questions :)
Dear FrederikWarburg:
Thanks for your reply!
But I think there may be another bug in your code:
https://github.com/FrederikWarburg/Burst-Image-Deblurring/blob/1359093179ebca9c1d53fb55590e70bc35844f8d/model.py#L96-L99.
After setting those “in_channels” to (96) 384+384+384, (97) 256+256+256, (98) 192+192+128, (99) 96+96+64, the code works for testing.
Good luck tou you!
发件人: Frederik Warburg 发送时间: 2020年1月13日 16:40 收件人: FrederikWarburg/Burst-Image-Deblurring 抄送: XueZ-phd; Author 主题: Re: [FrederikWarburg/Burst-Image-Deblurring] 3Questions: k,training_set, weights (#3)
Hi, Glad you find the code useful!
Dear FrederikWarburg: Thanks for your great code! But 3 questions make me confused for several days, I do need your generous help:
1. About value k: the value k is not defined in line 78 of 'class UNet(nn.Module)', specifically in model.py, so, how should I set the k value?
2. About training_set: I have downloaded the ImageNet 2012 dataset for training. As we know, there are 1,000 subfolders in folder "ILSVRC2012_img_train", so, should I unzip the 1,000 subfolders to 1 same "train" folder instead of 1,000 folders?
3. About weights: May you please release your weights?
Eagerly waiting for your response, Xue Zhang!