test about pytorch-inpainting-with-partial-conv

[wangcx2018]

To run train.py,how to set the datasets of Places2? Need to download in advance？if yes, where do the datasets place? thanks!

[naoto0804]

@an1018: This is the unofficial implementation by the third party and I admit that there still remains some bugs. I really appreciate it if you could report them.

[naoto0804]

@wangcx2018 see #15

[an1018]

@naoto0804 So why do you think this is happening?

[naoto0804]

@an1018 The biggest difference is the network itself. It might be better to strictly follow the paper (they use unet_256 used in pix2pix), but I have difficulty in interpreting some parts of the paper (e.g., The last partial convolution layer’s input will contain the concatenation of the original input image with hole and original mask.), since unet_256 does not concatenate the original input image .

[naoto0804]

Another possibility is that the mask that I used for this experiment is not suitable.
Recently, the authors make some part of the masks available on their project site.
Splitting that masks into the training and test subset, and then train the model may be better.

[jeejeelee]

@an1018 Hi, I also encountered this issue, In my personal opinion ,the red region is checkboard artifacts（ according to the original paper(page 7-8),Have you solved this issue?

[sunjunlishi]

@an1018 how do you use test.py;i don't know how to use

[an1018]

@sunjunlishi Just like this:
CUDA_VISIBLE_DEVICES=2 python test.py

[sunjunlishi]

oh,i have change the img path directly.but the result is not good as author's article

[sunjunlishi]

could author supply a better model.?it is important for me.My project is in need of watermarking and recognition

[naoto0804]

First, I'm not the author.
Second, as you can see in the codes, the current checkpoint that I share is trained on Places2 dataset. It is natural that you cannot get as good results as the paper if you try on faces.

[sunjunlishi]

oh.thank you verymuch for your reply