ggghsl / infoswap-master Goto Github PK

Hello, I have the following problem now. Could you please provide me with this checkpoint? Thank you

FileNotFoundError:No such file or directory: './checkpoints_512/w_kernel_smooth\ckpt_ks_G.pth'

Hi, thanks for your great work!

For the quantitative results in your work，I have questions about the correspondence of pair frames from source and target videos respectively.

(1) Did you randomly select 10 frames from each video or get the same pairs as FaceShifter?

(2) Could you provide the source-to-target pairs numbers for further fair comparison?

Thanks!

According to the pseudo code "Algorithm 1" in the supp.pdf file, there are 2 pretrained encoders, 3 decoders and 2 AII generators.

For the encoders and decoders, if I only use 1 module and use it several times during the training, I have this error:
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation
It's OK to copy a encoder as it's pretrained and I do not need to optimize it during training. But for the decoder, I think its not proper to copy it 3 times, as according to my understanding of the algorithm, the three parts should use decoder with the same parameter.

For the AII generator, I think I just need to have two generators with the same structure, just as the cycle gan has two generators. And I can use the Lcyc in line 41 to optimize these two generators together. Is this understanding right?

It would be awesome to test it out on Google Colab.
Thank you, sir.

hi. I'm implementing training code of this paper. Can you tell me about details of the training environment? (batch size, gpu ...)

我看38行是禁止了AII的参数更新，40行又恢复了AII的参数更新，是因为在生成Xs_cycle的时候传入到AII中的所有参数都是no_grad的吗？也就是Z_id_Yst也是detach出来的？

In inference_demo.py, R = encoder.features[i]; while in the forward function of IIB, R = readout_feats[i].
Which should I use for training?
And when I use R = readout_feats[i], Info is really large (5668875) at the beginning of training.

When I ran inference_demo.py as a test, I had the following problems while saving images:

could you tell me training method of pretrained face recognition model? (train dataset , loss etc..)

Do you have a plan to release the training code?

nice work!
Is it convenient to provide the pretrained face recognition model file model_128_ir_se50.pth?

Example in preprocessing is invalid, namely test_on_images.ipynb, could you please send it again

you can upload the pretrained models to google drive.
i can't access baidu outside of china.

In your experiment setup, you described alpha : beta = 1: 5 for IIB, is that means alpha = 1, beta = 5？

iib.py里面 191行，标准差为什么取的是均值呢？不太懂
m_s = torch.mean(Rs, dim=0) # [C, H, W]
std_s = torch.mean(Rs, dim=0)
Rs_params.append([m_s, std_s])
eps_s = torch.randn(size=Rt.shape).to(Rt.device) * std_s + m_s
feat_t = Rt * (1. - lambda_t) + lambda_t * eps_s
Xt_feats.append(feat_t) # only related with lambda

        m_t = torch.mean(Rt, dim=0)  # [C, H, W]
        std_t = torch.mean(Rt, dim=0)
        Rt_params.append([m_t, std_t])
        eps_t = torch.randn(size=Rs.shape).to(Rs.device) * std_t + m_t
        feat_s = Rs * (1. - lambda_s) + lambda_s * eps_t
        Xs_feats.append(feat_s)  # only related with lambda