Open WilliamHBW opened 1 year ago
Random seeds used in the code are available at Google Drive.
Hello, why does the size of the bpp increase as the epoch increases when training the network, does it tend to a steady value? And what are the lamada values for the five points of your method on the R-D curve?
I also observed that the bpp tends to increase when the training continues. I currently don't have an explanation for it. My intuition: using Gaussian entropy model to control the rate of quantized network parameters with Adam optimization is not a perfect solution. Hence, as mentioned in the paper, I also need to change the width of the architecture to control the rate.
I will be posting details on the requested lambda and network arch configurations later today or tomorrow.
Thank you, I have another question, your method was compared with the R-D performance of G-PCC, I used the official G-PCC code and found that only bpp was counted, not D1-PSNR. If the statistics part is your implementation, can you share the corresponding code?
G-PCC always losslessly encodes a point cloud. It achieves different rate-distortion tradeoffs by setting different positionQuantizationScale in the configuration (e.g. cfg/octree-predlift/lossy-geom-lossy-attrs/longdress_vox10_1300/r05/encoder.cfg).
Thank you. Do you remember the lamda value mentioned before? Thank you for taking the time to reply me!
I was using lambda from 150 to 600, with network width ranging from 8 to 32. Sorry that the exact settings to generate the R-D curves are pending due to technical issues.
Does the network width setting refer to this --chanstr 8,16,8,8; How is the range from 8 to 32 modified
Try the following: 8,8,8,8; 8,16,8,8; 8,32,8,8;
The R-D curve in the paper has five points for a frame point cloud. How do you control the two variables lamda and charstr
I am having difficulties retrieving my records so I cannot provide the exact settings for now. My impression tells me that they are the following: --lambda 600 --chanstr 8,8,8,8 --ch 3 --lambda 400 --chanstr 8,8,8,8 --ch 3 --lambda 200 --chanstr 8,8,8,8 --ch 3 --lambda 200 --chanstr 8,16,8,8 --ch 3 --lambda 100 --chanstr 8,32,8,8 --ch 3 I am not 100% percent sure since this is what I recall. Actually one can freely choose different settings for different R-D tradeoffs.
thank you very much!!
Hello, in the utils/network.py, 'SEED3.npy' and 'SEED4_Gaussian.npy' are loaded but are not found in the repository. Could you upload these two files? Thanks.