[Pose] Unsatisfactory training results

[Honlan]

Hi, thanks a lot for your awesome work.

I try to reproduce the results of pose in the paper. So I

• download 358 solo dance videos with simple backgrounds
• extract a slice of 30 seconds from each video
• obtain the poses with OpenPose and Densepose for all frames
• run train_g8.sh with 8 Telsa P40

However, after training for 108 epochs, I still cannot reproduce the results and the face regions are extremely terrible.

Cound you please give me some advice?

Another question. The option n_shot is set to 1 in base_option.py. Should I increase it so that the attention network can be trained during training?

Looking forward to your reply. Thanks again.

[Yilen-fan]

Could you provide me your training data ? I want to train it If the results are well , we can share the results , thanks my mail : mengmengboy@126.com

[azizsiyaev]

Could you please tell how much time did you spend for training with your data?

[Honlan]

Could you please tell how much time did you spend for training with your data?

0.19s per iteration, so 0.19 10000 100 seconds for the first 100 epochs.

[howardgriffin]

I set n_shot to 2 and learning rate to 0.0001, and I got the same result as you. I am also waiting for some advice.

[howardgriffin]

I used 2 GPUs to run the script and I doubt whether it was due to the multi-GPU?

[Honlan]

I used 2 GPUs to run the script and I doubt whether it was due to the multi-GPU?

I used 8 GPUs to run train_g8.sh

[howardgriffin]

I used 2 GPUs to run the script and I doubt whether it was due to the multi-GPU?

I used 8 GPUs to run train_g8.sh

Did you encounter the problem that as the number of iterations increases, the training stage is slower and slower?

[Honlan]

I used 2 GPUs to run the script and I doubt whether it was due to the multi-GPU?

I used 8 GPUs to run train_g8.sh

Did you encounter the problem that as the number of iterations increases, the training stage is slower and slower?

n_frame_total will double for every niter_step epochs after niter_single epochs

[centosrhel]

I used 2 GPUs to run the script and I doubt whether it was due to the multi-GPU?

I used 8 GPUs to run train_g8.sh

Did you encounter the problem that as the number of iterations increases, the training stage is slower and slower?

n_frame_total will double for every niter_step epochs after niter_single epochs

within niter_single epochs, say epoch 39, the training time is tens of that for epoch 1, and the time keeps growing for succeeding epochs. Did you encounter the problem?

[Honlan]

I used 2 GPUs to run the script and I doubt whether it was due to the multi-GPU?

I used 8 GPUs to run train_g8.sh

Did you encounter the problem that as the number of iterations increases, the training stage is slower and slower?

n_frame_total will double for every niter_step epochs after niter_single epochs

within niter_single epochs, say epoch 39, the training time is tens of that for epoch 1, and the time keeps growing for succeeding epochs. Did you encounter the problem?

I haven't encounter such problem. The training time of a single iteration keeps the same during niter_single epochs.

[howardgriffin]

I have the same problem as you when training the multi-frame stage and the 'DT_fake' loss is nearly 0. Did you solve it?

[Honlan]

I have the same problem as you when training the multi-frame stage and the 'DT_fake' loss is nearly 0. Did you solve it?

not yet :( looking forwards to suggestions from the authors

[AndroYD84]

I'm experiencing the same issue, my model started severely artifacting on the face region immediately and suddenly at epoch 25, before it was just a blurred face but there were no critical artifacts, I'm using the latest commit from 5 days ago (febb0b1), I started training using all default settings and running train_g1.sh on about 650.000 pictures split in 178 sequences/directories each with a different person.

[Peng2017]

This might because some of your openpose-face-keypoints' locations are out of you training images border(256x512). You can try any of the following solutions:

1. different "resize_or_crop" options,
2. carefully clean your dataset images, make sure all faces are located within 256x512 range.
3. Remove you extra-face-discriminator, only use the base generator. good luck!

[nihaomiao]

Hi, @Honlan , could you share the link to these solo dancing videos or give some keywords that I can search online?

Thanks!

[tcwang0509]

This repo is now deprecated. Please refer to the new Imaginaire repo: https://github.com/NVlabs/imaginaire.