yang7879 / 3d-recgan Goto Github PK

I tried to train this network on my own dataset and met the issue with main .py file.
Whenever I try to change certain parameters as batch size, the value of i iterator for the moment of when I want to save the model or test the model, I receive the predicted results of 0 matrices or very strange values. But, while using vanilla code everything seems to work fine. Are those values somehow related to the code structure?
For the training I'm using DGX stations w/ Tesla V100.
CUDA 9, cudNN 7
Tensorflow 1.12.0

Unable to visualize the generated output (.mat files). Any recommendations/alternative softwares to visualize the chair result .mat files.

Hi!
Thank you for sharing the code! I notice in the paper the output of the discriminator is a vector. I am wondering if the loss of the discriminator is the different between two sum of the output vector(one for real and one for fake)?
Thanks!

Hi, because I can't use Google search engine, you provide the url “https://drive.google.com/open? Id =1n4qQzSd_S6Isd6WjKD_sq6LKqn4tiQm9 ”cannot be used. Can you improve other ways of sharing?
Looking forward to your reply.

Hi,
Thanks for sharing your great work!
I tried implementing your solution and I was wondering what is the expected behavior of the discriminator loss (the gan_d_loss in your code).
Is it supposed to exponentially decrease and converge to zero? I was assuming that if removing the gradient penalty the expected loss for a perfect discriminator should be -1. What should I expect with the gradient penalty? Also, is the gan_g_loss supposed to vary accordingly? In my test case, it is stuck around -0.5.
Thanks in advance for your input!

Hello, when I was studying your thesis(《3D Object Reconstruction from a Single Depth View with Adversarial Learning》), I found that the code could not be run. Can you describe how to use it? I downloaded your github code and installed the conda environment on win10. After running, nothing is generated in the folder train_mod. In addition, which parameter can be modified to be smaller, the computer speed is very slow. Thank you for your answer!