zhaoyuzhi / deepfillv2 Goto Github PK

Wonderful work!
I am sorry to bother you,but I have a question about the value of opt.lr in train.py.

Traceback (most recent call last):
File "train.py", line 62, in
trainer.WGAN_trainer(opt)
File "/home/fl/deepfillv2/trainer.py", line 152, in WGAN_trainer
adjust_learning_rate(optimizer_g, (epoch + 1), opt)
File "/home/fl/deepfillv2/trainer.py", line 53, in adjust_learning_rate
lr = opt.lr * (opt.lr_decrease_factor ** (epoch // opt.lr_decrease_epoch))
AttributeError: 'Namespace' object has no attribute 'lr'

I would very appreciate if you could help me,thank you!

I have released all new code and models including colorful and grayscale image inpainting. Please refer to README

I used 10K images to train for 120 epochs, but the results were worse. :





where is wrong?

May I ask one question on discriminator output?
The code shows the discriminator are suppose to output 256 channels
x = self.block6(x) # out: [B, 256, 8, 8]

However, the definition of block6 are shown below
self.block6 = Conv2dLayer(opt.latent_channels * 4, 1, 4, 2, 1, pad_type = opt.pad_type, activation = 'none', norm = 'none', sn = True)
It seems it will output 1 channel instead of 256 channels

In the deepfillv2 paper, it is trying to output >1 channels. May I ask why the implementation changes this?

hi, the 'deepfill2' folder on your Onedrive seems empty, only 'deepfill2_grayscale' has some files

Hello, I really appreciate your elegant implementation of the gated convolution and the coarse to fine structure. I am very curious about how long does it take to get the comparable result as the original paper?

Hello there!
First of all, thank you for your code.
But, I have a simple problem. I really want to try RGB images, so is there any possible pre-trained model(RGB) to use that?

thanks!

Could you tell me if this works, because training a network cost so much time. Thanks!

In the trainer.py, line no. 53 and 54, you have weight_decay = opt.weight_decay. I tried to find the value of weight_decay in run.sh file, but I could not find it. Kindly let us know the value you used.

Hello. Nice repo!
But there are no pretrained weights for RGB?
This folder is empty https://portland-my.sharepoint.com/personal/yzzhao2-c_ad_cityu_edu_hk/_layouts/15/onedrive.aspx?originalPath=aHR0cHM6Ly9wb3J0bGFuZC1teS5zaGFyZXBvaW50LmNvbS86ZjovZy9wZXJzb25hbC95enpoYW8yLWNfYWRfY2l0eXVfZWR1X2hrL0V0NUJha25GZXdORWdPNkM1MlJ4TU4wQnJyb1ZaakVMX2lWd0taUTUzeVhjWFE_cnRpbWU9NFlldXFtQXQyRWc&id=%2Fpersonal%2Fyzzhao2-c_ad_cityu_edu_hk%2FDocuments%2Fgithub%2Fdeepfillv2%2Fdeepfillv2

Hi, could you add a license?

Hey,

Great work and really appreciate for sharing the code.
I want to train deepfillv2 for RGB images. But, when I tried to train it on custom data it raises an error in line number 39 in utils.py file saying that NoFileFoundError for "./vgg16_pretrained.pth".
I looked for the file in your repository but was not able to find it. Can you please share the link for the file?
Thank you in advance.

Please allow me expressing my gratitude for this beautiful re-implementation.

I suggest the author updates the README on section 1.1 about the parameter setting since I find 'perceptual_param' and 'gan_param' have been replaced with new names 'lambda_perceptual' and 'lambda_gan'.

Besides, may I ask why the 'lambda_gan' parameter has been setting that low(0.01) and only applying on generator loss instead of on both generator and discriminator loss? I am assuming it will make discriminator have little impact to guide generator training. The default value '1' shown in 'train.py' seems a more reasonable value for me. Exhibiting the exact parameters of training current model is very crucial for others to understand the system.

I will be more than appreciate if you can answer my question.

It is a great work, but it seems lack of "Two Branch Refinement Network with Contextual Attention (Stage II)" in deepfill_v2 corresponding to refinement network in deepfill_v1 for grayscale inpainting. I really want to know the reason. Will you please reply to my confusion? Thanks a lot!

Thanks for sharing your work. It seems like that you do not include optimizer.step in trainer.py.