search

Zhiyuan-R / Tiger-Diffusion

This is the official code for the CVPR 2024 Publication: Tiger: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process
15 stars 2 forks source link

Questions about paper(about normalization) #4

Closed JudgeLJX closed 3 months ago

JudgeLJX commented 3 months ago

Hi,

Thanks for the excellent work.

Could you figure out my confusion following:

I run your code on categories chair/car, and the output provides some visualization.

  1. According to the x.png/eval.png, it seems each point cloud is normalized into a cube, the chair looks nice, but the car looks like a mess, is that fine?
  2. Did you only use the global normalization across the dataset in your work?
  3. Did you calculate the metrics with output and the normalized original data distribution?
  4. If the generations are normalized into a cube, did you reshape them to the original scale (in your paper, the visualizations are normal scale)?

I would appreciate it if you could answer.

Zhiyuan-R commented 3 months ago
  1. Could you try a little bit longer training for the car? If you still have issues, i can try to reproduce again from my end.
  2. I think so.
  3. Yes! (Actually, I follow the same setting as the PointFlow )
  4. https://github.com/Zhiyuan-R/Tiger-Diffusion/blob/ca6aaf6a6a3ce084ad9d0252a3ec60555ccedd42/train_generation.py#L210 I think I visualize it after a clipping operation like this. Otherwise, some points will go out of the space.
JudgeLJX commented 3 months ago

Screenshot 2024-07-17 at 22 07 53 Screenshot 2024-07-17 at 22 09 33 Thanks for your quick reply, for the generation. I got the 850 epoch results (samples and x ), do you think this looks good?

If possible, could you please release your checkpoint and test part?

When I tried to test via $python test_generation.py --category chair --model 'output/train_generation/2024-07-10-19-40-38/epoch_1399.pth'; it shows that the model mismatch with the model in training. I checked the model in the training and testing file. In the training file: has model.module.transformer. .... but the testing file: has model.module.global_att. ... instead

Best

fivefingerhill commented 3 months ago

I meet the same problem that the Model for denoise in test_generation.py is different from that in train_generation.py, but not master. it's easy to debug.

JudgeLJX commented 3 months ago

I meet the same problem that the Model for denoise in test_generation.py is different from that in train_generation.py, but not master. it's easy to debug.

Hi, Did you reproduce the experimental values shown in the paper? I didn't achieve it, maybe I got something wrong.