Skip to content

Latest commit

 

History

History
34 lines (26 loc) · 1003 Bytes

README.md

File metadata and controls

34 lines (26 loc) · 1003 Bytes

Apex Trainer Example

Install Apex

  • apex: pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --upgrade 'git+https://github.com/NVIDIA/apex'

Results:

Trained with 2080Ti

faster rcnn

lr sched 1x

Name train time(s/iter) infer time(s/im) train mem(GB) box AP
R50-FPN-V100 0.210 0.055 3.0 37.9
R50-FPN 0.312 0.056 2.893 37.9
R50-FPN-Apex 0.322 0.053 2.793 38.0

RetineNet

lr sched 1x

Name train time(s/iter) infer time(s/im) train mem(GB) box AP
R50-V100 0.200 0.062 3.9 36.5
R50 0.285 0.079 3.934 36.5
R50-Apex 0.248 0.067 3.275 36.4

Mask R-CNN

lr sched 1x

Name train time(s/iter) infer time(s/im) train mem(GB) box AP mask AP
R50-V100 0.261 0.053 3.4 38.6 35.2
R50-FPN 0.348 0.063 3.44 38.5 35.1
R50-FPN-Apex 0.363 0.061 3.26 38.6 35.2