lesliezhoa / dct-net.pytorch

unofficial implementation of DCT-Net: Domain-Calibrated Translation for Portrait

Hi there 👋

I am Leslie from China, nice to meet you all!

• 😄 Interst: computer vision especially Image generation
  
  
• 📫 How to reach me: send email to [email protected]

dct-net.pytorch's Issues

Why doesn't the pairing of data generated by get_tcc_input.py look like？

Trained for 9 epochs, 4000 times，

Does the ccn stage need to be trained for so long?

Are the following training results correct?

Can the CCN stage be directly adopted https://github.com/rosinality/stylegan2-pytorch?
to train, or what changes did you make in the CCN stage

In the TTN stage, do you use data pairs for p2p network model training? Data pairs are necessary?

error: name 'Engine' is not defined

Inference does not work correctly

Hi @LeslieZhoa, thanks for your implementation.
I tried running the inference.py script and I faced some problems.

I fixed the 'netG' key in the line:

DCT-NET.Pytorch/inference.py

Line 25 in ebdc7bd

self.net.load_state_dict(ckpt['netG'],strict=False)

to 'G' because final.pth does not contain 'netG'.

Also in the postprocess method you make the slice [..., ::-1], and you also make it after calling the run method:

DCT-NET.Pytorch/inference.py

Line 55 in ebdc7bd

oup = model.run(img)[...,::-1]

I removed the extra slice.
But I still get an unrealistic image (ran at 1024 and 256 resolutions). I'm attaching an example (first line is 1024, second line is 256) and waiting for a response/comment.

一个笔误在model/Pix2PixModule/module.py中的color_shift

b1, g1, r1 = torch.split(image, 1, dim=1) 应该是 r1,g1,b1

Training error

Hi @LeslieZhoa ! How can you compare the dictionary and the float?

DCT-NET.Pytorch/trainer/ModelTrainer.py

Line 124 in b303f41

if loss < mn_loss:

This method is called here:

DCT-NET.Pytorch/trainer/ModelTrainer.py

Line 96 in b303f41

acc_num,mn_loss,stop_flag = self.early_stop_wait(self.get_loss_from_val(val_loss),acc_num,mn_loss,epoch)

But the val_loss obtained in CCNTrainer.evalution is an empty dictionary:

DCT-NET.Pytorch/trainer/CCNTrainer.py

Lines 196 to 206 in b303f41

	def evalution(self,test_loader,steps,epoch):
	loss_dict = {}
	with torch.no_grad():
	fake_s,_ = self.netGs([self.sample_z])
	fake_t,_ = self.gt_ema([self.sample_z])
	if self.args.rank == 0 :
	self.val_vis.display_current_results(self.select_img([fake_s,fake_t]),steps)
	# self.val_vis.display_current_results(self.select_img([fake_t]),steps)
	return loss_dict

I want to understand how to fix this and what metrics to count? And as I understand it, you trained the network without early_stopping. How many epochs/iterations did it take?

The inferred effect is black？

What's the matter with these black heads?

背景问题

生成人脸都比较好，就是处理整图背景表现出有一块一块的黑色的

训练结果看起来不太正常

哈喽，大佬，我按照你的readme训练了自己的样式，但效果貌似不太令人满意，能告知我大概在哪出了问题吗
训练样式图：（1200张）

ccn训练了2000步
ttn训练了4w步的结果是：

非常好的开源代码

非常感谢作者伟大的工作，我尝试用ttn代码训练（ccn,face expression estimator等都没发现啥问题），大概训练了50epoch,但是效果不太理想。是训练时间过短还是有其他原因。

Training data size

What is the image size used for training CCN,
Why can't the matrix be multiplied when I input 1024 in training,
and The problem of ' 'Nan or Inf found in input tensor '' when inputting 256 during training

请问CCN模型和TTN模型在使用上有什么区别？为什么要训练两个模型？

可以跑训练了，取消issues

可以跑训练了，取消issues

关于ccn的blend和eometry expansion module

十分感谢你的贡献，但我有几个问题想要咨询一下。
我发现你在训练ccn的时候没有做论文中提到的blend操作，Geometry expansion module我也没有找到，这是有什么原因吗？
另外论文作者提到他用相同的模型对头部和背景进行了两次推理，只是对头部模型进行了额外的优化，但在你复现中似乎只有一个模型并且只做了一次推理，是有什么改动吗？
另外show中展示的效果似乎不是很好，有更好的效果图可以分享一下吗？谢谢

Great job!

thanks. a nice work!