search

facebookresearch / mvp

Training and Evaluation Code for "Mixture of Volumetric Primitives for Efficient Neural Rendering"
Other
191 stars 16 forks source link

Cannot replicate experiments/neuralvolumes results: Completely vague output and vanishing kldiv #13

Open RogerDAI1217 opened 1 year ago

RogerDAI1217 commented 1 year ago

Hi, could someone kindly help me? I cannot replicate the output results in experiments/neuralvolumes. I have successfully built the extension and download the experiment.zip in the latest release. However, the output images are completely vague. I found that my kldiv term is quickly vanishing to 0, while in the example log.txt file given in experiment.zip, the kldiv terms remain larger than 0.3.

Below are my outputs after 79579, 92139, and 106682 iterations. prog_079579 prog_092139 prog_106682

Here I append my log.txt file which contains my configuration information and training statistics: log.txt

I'm not sure where is the problem. Incorrect camera pose? Or the code in this repository has some bugs. Really hope some kind guy could help me. Thanks! :)

RogerDAI1217 commented 1 year ago

Also, my irgbmse remains larger, while the irgbmse in the sample file in experiments.zip decreases quickly. :(

stephenlombardi commented 1 year ago

Hi,

Please post your version of python, pytorch, gcc, and cuda. Also please post the output of building the CUDA extensions. I'm guessing this is a problem with your environment because I just built the extensions and tested the training code on an instance of Google colab and it appears to be training fine:

prog_000406

(the Colab instance is running with Python 3.7.14, GCC 7.5.0, PyTorch 1.12.1+cu113)