kumapowerliu / csa-inpainting Goto Github PK

I couldn't find the torch.eval() function for evaluating test set.

That's a great work !
I would like to know whether the mask used in this paper comes from the training set of partial convolution?
Because the mask in this paper will not be updated during training to gradually narrow the mask area.
I find that the mask occlusion area in the training set of partial convolution is too large. Is this conducive to training?

load error

I find that your code is mainly based on deprecated code in pytorch 0.3 branch https://github.com/Zhaoyi-Yan/Shift-Net_pytorch/tree/pytorch0.3.1

However, this branch is slow, now the code in the master branch is must faster and is fully-parallel https://github.com/Zhaoyi-Yan/Shift-Net_pytorch/blob/master/models/shift_net/InnerShiftTripleFunction.py#L36-L49

You may upgrade your code to make it run much faster.

In CSA layer, if D_maxi or D_adi is a negative value, m_i can be an inappropriate value.
Because of this, ng_loss_value diverged to a very large value (about 1e+10).

Do you have any solution to this problem? I would appreciate it if you could share some idea with me.

我有两块GPU 0和1,除了在class Opion() 类中指定 ### 参数 self.gpu_ids=[1] 外，代码中还需要在哪块有指定，我只改了这个它显示了以下错误：（### train.ipy运行到 model.optimize_parameters()出错）。
调用CSA.py 中的224行出现报错
223 def optimize_parameters(self):
--> 224 self.forward()

##报错的提示：
RuntimeError: Expected tensor for argument #1 'input' to have the same device as tensor for argument #2 'weight'; but device 0 does not equal 1 (while checking arguments for cudnn_convolution)

期待您的帮助，我将不胜感激。

作者你好 请问批量大小能不能设置大一点呢 , 代码说需要设置为1 ,如果想设置大一点应该怎么改代码

Your work is very helpful to me. Can the complete code of this paper be uploaded? thx

Can you please specify the license? Preferably MIT

想请教一下，在celeba和paris 上分别需要设置epoch为多少？，总的训练epoch数是不是niter+niter_decay?

我想问下，这个flatten_offsets作用是什么，谢谢。

Is there any standard mask file source? Because the network issue, I can not download the specified three dataset, can you provide the mask file only?

when the batchsize is 1 it just use 30% GPU memory in 2080Ti( similar in 1080Ti)

and even I set batchsize as 6 it just use 5385 MB (which I don't understand)

is there any reason ?

pic = (torch.cat([real_A, real_B,fake_B], dim=0) + 1) / 2.0
torchvision.utils.save_image(pic_1, '%s/Epoch_(%d)_(%dof%d).jpg' % (
save_dir, epoch, total_steps + 1, len(dataset_train)), nrow=2)

pic_1 not defined, 是不是要把pic_1改成pic？
按如上修改后，save_dir not defined 这是为啥呢？是代码执行的先后顺序有问题吗？

Hi, I'm reading your CSA-inpainting work recently. The most impressive novelty of this work is that CSA layer takes use of information from not only known regions, but also generated contents. A good work~

But I'm a little confused with the feature patch discriminator. Could you please help me?

1. Pretrained? Are shallow layers of the feature patch discriminator pretrained on ImageNet, i.e., you use the shallow layers from pretrained VGG Net?
2. Does the network directly calculate the adversarial loss in 14x14x512 feature map? How do you set the ground truth? a 14x14x512 tensor filled with 0/1? Dose this idea/operation similar with the paper Learning Pyramid-Context Encoder Network for High-Quality Image Inpainting, CVPR2019 . Why you would like to do so, avoid heavy parameters in fully connected layers?

Thx!

Hello, thank you for sharing this great topic. Can you sharing the code for evaluating the results ?
I'm looking forward to your responding. Thank you

请问您在用Irregular Mask Dataset训练的时候使用那个网页上说的随机平移、旋转、膨胀、裁剪等方法来扩充数据集了吗 Places2, CelebA and Paris Street-View这三个数据集您用了多少呢

能提供代码吗？邮箱地址[email protected]

could you please offer an pre-trained model for your network?

Thanks for your great work.
I am wondering if you could share your Places2 full training config, e.g. the training epoch, lr_decay, etc..
Otherwise, did you train your model on Places2 low resolution(256*256) one or high resolution one?

The code, looks extremely clean but any ideas how to prepare the dataset. I want to use it on my custom datset.

请问这个文件是用来存放什么的，还有在opin类中，有个变量 self.checkpoints_dir=r'.\checkpoints' # 请问这个变量的作用是什么呢？ 十分期待您的回复！万分感谢，祝您生活愉快！

Hello, thank you very much for the code.
Now,I use celeba-hq data to train model. In the program, according to the parameter opt.display_freq, the intermediate results can be displayed. However, from the 14th epoch, the output results of the model become all white,such as :



I don't know what happened. Have you ever met this situation? I look forward to your reply!

i used CelebA -> img_align_celeba.zip dataset, then image size = 178x 218.

When i run

error...

Thx

Hi,
Do you have plan to release pre-trained model?
We can try the result, thank you.

您好，您论文中说Place2数据集训练完只需要2天。但我在RTX TITAN上训练的时候，单帧图像训练耗时在500ms~1s之间，例如训练30个epoch，则训练完的时间远远不是2天（2个月甚至都训练不完），请问是我速度哪儿预估不对么。

Dear Mr. Liu,

Thank you for your work and code. They are fantastic.
I have a question. How about if the dataset has different size images or rectangular images? Have you tried that?

Sally Huang

Please suggest how to port the deep learning model to Android.

I was wondering if there is any way to visualize the attention weights over the original inputs. I see such figures in papers.

1.为什么batchsize只能设置为1？我尝试过设置为6，在3个epoch之后loss都变成了Nan，请问调大batchsize需要改代码的哪些地方？
2，GPU利用率不稳定的问题，当batchsize为1时，GPU的利用率在20%-90%之间波动，我试着调大了batchsize该问题仍然存在，请问该如何解决？修改dataloader的num_workers以及pin_memory也都无效。

Hi there, I'm also working on inpainting, and find attention mechanism very interesting. Yet I found it difficult to make it work on inpainting problems. Is there any plan on releasing your code? Thanks a lot.