search

CMU-Perceptual-Computing-Lab / caffe_rtpose

Realtime C++ code for multi-person pose estimation
Other
357 stars 207 forks source link

GPU Memory #18

Closed tladyman closed 7 years ago

tladyman commented 7 years ago

Hi,

When running rtpose.bin I'm running out of GPU memory.

116 syncedmem.cpp:64] Check failed: error == cudaSuccess (2 vs. 0) out of memory

Is there any way to decrease the amount of memory used or is the answer to use a GPU with more memory?

Cheers, Tom

gineshidalgo99 commented 7 years ago

Hi, we are working in a version that uses less memory (it should fit in 2GB GPUs). It will be released in around 3-5 weeks. However, if you are using a bigger GPU and it does not fit, you should reduce the number of scales and net resolution to the default values, as well as enable cuDNN. Thanks!

ZheC commented 7 years ago

4 Options to reduce memory usage: (1) Using MPI model instead of COCO model (2) Modifying the test.prototxt and reducing the stage number from 6 to 3. (3) Using one scale for testing (4) Using smaller --net_resolution for testing

madhawav commented 7 years ago

@gineshidalgo99 Is the version for 2GB VGAs available?

gineshidalgo99 commented 7 years ago

I am very sorry for the delay, it will be released within the next 2-3 weeks. Thank you for your patience.