search

ChrisWu1997 / PQ-NET

code for our CVPR 2020 paper "PQ-NET: A Generative Part Seq2Seq Network for 3D Shapes"
MIT License
118 stars 18 forks source link

some questions about results. #14

Open laodouza opened 3 years ago

laodouza commented 3 years ago

Dear Dr.Wu, I am a student of UESTC and I'm very interesetd about your work. Thanks for your code and data. I get the results of auoencoder of your work. It looks very well. But when I use the part of GAN, I meet some problems. The generated results is not look like an entire shape. And do you konw what happen?The dataset I used is chair. Looking forward to you reply. --------------- by a sincerely reader

laodouza commented 3 years ago

I don’t know why, I can’t upload pictures for now, but it seems that there is only one part, such as the part where the armrest and the seat are connected.

laodouza commented 3 years ago

微信图片_20201230173405

ChrisWu1997 commented 3 years ago

Did you use the pretrained model? How do you get the above voxel outputs?

laodouza commented 3 years ago

with h5py.File(save_path, 'r') as fp: voxel = fp['voxel'][:] np.save('0095.npy',voxel)

i use this code to save the voxel.

laodouza commented 3 years ago

and I did not use the pretrained model. I used the code released to train.

ChrisWu1997 commented 3 years ago

If the autoencoder works fine but the GAN doesn't work, then it means the problem comes from the latent GAN training. I'm not able to tell what the exact problem is without detailed informations.

Can you try the pretrained model first and see if that works? To make sure the current problem comes from the latent GAN training, rather than other parts.

laodouza commented 3 years ago

ok, I will try. Thank you for your prompt reply .