colorful-liyu / 3dqd Goto Github PK

Hello, your work is excellent. I would like to ask how to view the denoising results of paper 3.4 and where to see the generated visualization results. Looking forward to your reply, thank you very much.

Hello. Thank you for the great work!
I want to use the single-image 3D reconstruction model. When will you release the code for image-to-3D reconstruction?
And, could you release the demo for reconstruction from an arbitrary image at the same time? Thank you!

It seems that the definition of prepare_condtion function in df_model.py file is missing.

Hi, I'm studying the paper 3DQD, and I'm interested in the part called MFM. But I cannot find the specific code in this project. Could you tell me where this part of code is located? Appreciate for your help!

Thanks for your works!I wonder if there is code directly in place to process point cloud data. Because I saw experiments in the paper that generated point clouds but I didn't seem to find code that dealt with that data. By the way, your One Drive link is broken, can you fix it? I will be very grateful.

Hi, colorful-liyu!Great work!
I want to ensure some details about language-guided generation.

First, for each text in the test split of ShapeGlot, how many samples are generated (one text with one sample or one text with multiple samples)?

Second, how many samples in total are generated? Are texts from ShapeGlot fully used or just part of them is used?

At last, could you share me your hyperparameters config? I'm not sure that whether the config in the code is the same as yours. And I can not get the same metric values using the default config in your code.

Appreciate for your helps! Wish you all the best!

你好，我从disn下载sdf_v1数据集但是您的代码需要sdf_v1_64数据集，两者的数据尺寸不一样，请问后者应该在哪里下载呢

Unable to open the link to the Pix3D and ShapeNet preprocessing datasets, can you update the link?
Thank you very much~

In 'uncond_metrics.py' at line 41, the 'pairwise' in methods like LION, PF, PVD, etc., is centered around the test samples, whereas in your code, it is centered around the generated samples. To align with other methods for comparison, you should modify:
M_rs_cd,M_rs_emd = _pairwise_EMD_CD_(sample_pcs, ref_pcs, batch_size, EMD_flag, verbose=True)
to
M_rs_cd, M_rs_emd = _pairwise_EMD_CD_(ref_pcs, sample_pcs, batch_size, EMD_flag, verbose=True)
This might be the reason for the notable performance difference in the supplementary materials section of your paper compared to other papers.
Additionally, could you provide generated samples for the unconditional generation task or share the corrected evaluation results for quantitative comparison with your method?

Hello liyu,

I implemented and trained a VQVAE myself with a 512-slot codebook. The visualization of reconstruction is nice. However, when I came to test it for the diffusion process, I met a bug.
The log_z sampling from the diffusion model has length 513 (in LL553 of ./3DQD/models/networks/transformer_networks/diff_transformer/diffusion_transformer.py), which may generate an undesire code idx with value 512 and result in an Index overflow when querying features for the codebook.
I didn't notice any tools that prevent the generation of this undesired code value, but your unconditional generation demo works fine. It has taken me several days to solve this bug. Hope you can explain the rationale behind the design.

Why do we need the function unfold_to_cubes?

你好，请问这个json文件在哪里呀？

Thanks for your works! Can we use our own data to do point completion? How to do it?

Hi, I successfully run the eval_uncond.py which i believe it can unconditionally generate a 3d model like obj or other formats. But I only find a .npz file. Is it the output of unconditional gen?
Besides, where can I find the codes for just generating without evaluating, and then visualize the output file?
Thanks for your replying for my former issue, it's a really great work!

chamfer-3d==0.0.0， clip==1.0，emd-ext==0.0.0，pointnet2-ops==3.0.0
i can't pip install these packages, so i can't run the unconditional generation, where can i get these pkgs?

Thanks to the excellent work of open source, when running ‘eval_uncond.py’, the following error is prompted. Have you ever encountered a similar situation and how did you resolve it? Looking forward to your reply.

Pairwise EMD-CD: 0%| | 0/1311 [00:00<?, ?it/s]
Traceback (most recent call last):
File "/home/rain/3DQD/evaluation_scripts/eval_uncond.py", line 162, in
results = EVAL(opt, model, sdf_files, gen_num=len(model_list), gen_bs = 16, save_flag = True, guidew = opt.guidew, class_label=1, class_desc=cats) # 795, 1311, 626
File "/home/rain/anaconda3/envs/3dqd/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 28, in decorate_context
return func(*args, **kwargs)
File "/home/rain/3DQD/evaluation_scripts/eval_uncond.py", line 69, in EVAL
results = compute_all_metrics(opt.device, gen_list, sdf_files, 48, opt.mode, logger, opt.EMD_flag, num_points=opt.num_points, class_label=class_desc)
File "/home/rain/3DQD/metrics/uncond_metrics.py", line 41, in compute_all_metrics
M_rs_cd, M_rs_emd = pairwise_EMD_CD(sample_pcs, ref_pcs, batch_size, EMD_flag, verbose=True)
File "/home/rain/3DQD/metrics/uncond_metrics.py", line 202, in pairwise_EMD_CD
emd_batch = EMD(sample_batch_exp, ref_batch, transpose=False)
File "/home/rain/3DQD/metrics/pvd_metrics/PyTorchEMD/emd.py", line 44, in earth_mover_distance
cost = EarthMoverDistanceFunction.apply(xyz1, xyz2)
File "/home/rain/3DQD/metrics/pvd_metrics/PyTorchEMD/emd.py", line 11, in forward
match = emd_cuda.approxmatch_forward(xyz1, xyz2)
RuntimeError: Unrecognized tensor type ID: Autocast

Would you mind sharing the tensorboard of your vqvae training stage?

could you please tell me how to render the voxel grids?