atlas200dk / sample-imageinpainting-hifill Goto Github PK

When will the training code be open source?

Hi, thanks a lot for your research and your detailed explanations !

I still have question about the architecture of the attention computing branch. What does [P] stands for after the downsampling step ? It is described as a high-level feature map (compared to [P^l=3]) with several additional layers in Fig. 2, but not mentioned in the Appendix.

Hi,

Thanks for publishing this source code! It makes testing your model easy :)

I'm currently testing the model on center-crop masks; however, I find that the model is struggling with larger center crops?
The images I tested were 512x512, with a center mask of 256x256.

Here is an example image that I tested, where I got strange corruption in the final result.
Ground truth image:

Input image:

Generated image:

Do you know what might be the cause of this? Or are center-crops difficult for the model?

If you would like to reproduce the result, here is the original mask (256x256 center crop):

If the images are compressed, I can give you a link to the raw images :)

@duxingren14 @ascendhuawei

First of all, thank you for your great work? Can I train my own dataset on this network? Will you open your network code?

I added some functions for tf2.x, I tested on google colab and it's working, link is here @ascendhuawei

thanks for your contribution
I am asking how can I fine tunning your model
is there instruction
thanks

Hello, Great work!
Can you share the full pipeline including the training code?

thansks for your excellent work！when I test the example images you provide with your code, It perform very well. However, if we test the image shown above (400x600), it perform very bad. The result of completion will be a little blurry. I noticed that the minimum rate of change you tested in the experiment was 512x512. Does that mean that the algorithm needs to be on large images to get good results?

Hello,

I've been trying to use this project, but I use the 20.1 CANN with 1.7x. It changed from the 1.3x version and we haven't a hiai library anymore, everything is made using ACL. I can't see a hiai--acl documentation with more information about this transition. I'd like to know if there's anything being made to upgrade this application.

Thanks and Congratulations for the job.

cuda:10.0.130
cudnn:7.4
tensorflow-gpu:1.13.2
python:3.6

Hi, thanks for your excellent work.

I've read your paper and I'm really impressed by the results in your paper. But I am having some question when trying to test the model. I would appreciate it if you could provide some help.
In the experiment part, it's mentioned that the model is tested in several different resolutions, from 512512 to 8K. However, when I was testing the model, I found that the images in Places2 dataset are at most 512 pixels, and there seems to have no ground truth image whose size is really larger than 512. If the high resolution samples are produced by resizing the 512512 images, it seems that the evaluations on high resolution images are actually equivalent to that on 512*512 because each part of the model is essentially working on the 512 bases (including the ground truth). Am I misunderstanding something or there is another source of private testing samples?

Thanks for any help in advance :).

Is it possible to apply the model to images smaller than 512?

I noticed that a 7x7 kernel is used in the model to process the mask. Why do you want to do this operation? And how to train this module? In addition, attention score didn't work very well in my own training, because the score of each patch was always average. Are there any suggestions you can give me?

Hi! Thanks for your great work! But when I implementing the post process, I have no idea what the attention score is in your demo. In the paper, it said only one patch outside mask and one inside will have similarity. So I make all others in zero and do softmax to get the attention score. But it seems not the one in your demo, do you have any advices for me?

hi, @duxingren14 @ascendhuawei , you really did a great job. It is so amazing. It is also usefull for me.
but when I try your model in my own dataset, the result is bad. Can you help me?
the following is the input images:

and get the follow result:

Dear author, in the paper, the model was trained on 512x512 images. However, both the input and output of the generator are 512x512, which I think there is no need to resize the output to a high resolution to get the residuals. So my question is how do you train your CRA module on 512x512 images? Even there is no weights to be trained in CRA, it does affect the other part of the generator. And is the loss calculated on the generator output? Or the postprocess output (end-to-end training)?

Thanks for sharing the paper and model.
But we cannot get good results on our normal test cases
Most inpainting results have artifacts even in very low resolution cases especially when the hole is big.
Is there any update?

Hi, thanks for the great work. I would like to integrate your model on my task. SoI have a question about doing inference with batch data. The pretrained model seems only to provide inference for one image, i.e., the shape of the input is [1, 3, 512, 512] instead of [?, 3, 512, 512]. It means that I cannot pass batch data to the model. As a result, it took 10s to have the result for a batch input, e.g., BATCH_SIZE=16. Is there any solution to fasten inference? Thanks in advance!

Thanks for sharing code of your novel and inspiring work. However, when I runing your code, I got the following error:

F tensorflow/core/framework/tensor_shape.cc:44] Check failed: NDIMS == dims() (2 vs. 4)Asking for tensor of 2 dimensions from a tensor of 4 dimensions 

I'm using tensorflow 1.6 with TitanXp.

I test the post runtime beyound 6s.
But you paper pay more attention to talk the inpainting model runtime?

8 Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz

InvalidArgumentError (see above for traceback): logits must be 2-dimensional
[[Node: Softmax = SoftmaxT=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"]]

Who can tell me how to do 。。。

hi, @duxingren14 @ascendhuawei , you really did a great job. It is so amazing. It is also usefull for me.
but when I try your model in my own dataset, the result is bad. Can you help me?
the following is the input images:


and get the follow result: