search

Shimingyi / MotioNet

A deep neural network that directly reconstructs the motion of a 3D human skeleton from monocular video [ToG 2020]
https://rubbly.cn/publications/motioNet/
BSD 2-Clause "Simplified" License
554 stars 82 forks source link

Question about Training Time #27

Closed JinchengWang closed 3 years ago

JinchengWang commented 3 years ago

What is your best estimate of the amount of time it takes to train the network? And what kind of GPU did you use?

Shimingyi commented 3 years ago

Hi @JinchengWang , I haven't recorded the training time. But in genral, you can reach the performance about 55mm in one hour, and a complete training will take extra few hours. The model can be trained on all Nvidia GPUs which support Cuda and cudnn, you just need to install a suitable pytorch-gpu. I have experiences to train the model on gtx 1080, titan xp and rtx 3070 which can reach same performance on the accuracy.

JinchengWang commented 3 years ago

Thank you! That was fast oдo

Shimingyi commented 3 years ago

Because the network doesn't use image feature as input, so the tensor dimension is flexible. And also, we haven't predicted the camera parameters which requires lots of optimization steps in the training : ) Feel free to contact me(mailto: irubbly@gmail.com) if you have any idea on the extension if I can help.