search

sacmehta / ESPNet

ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation
https://sacmehta.github.io/ESPNet/
MIT License
541 stars 112 forks source link

inference time #11

Closed wldeephi closed 6 years ago

wldeephi commented 6 years ago

@sacmehta , thanks very much for your work, how about the inference time for one image with input_size=1024x512 ? I saw the time is 0.0089s per image shown in the cityscapes benchmark, but I get the time is 0.132s with input_size=1024x512 on GPU, and my GPU is Titanx 1080.

sacmehta commented 6 years ago

Could you please try to measure the model execution time only i.e. without image loading and image writing time?

Note: For the very first iteration, inference time will be low.

MrLinNing commented 5 years ago

Hi, @sacmehta I test the 1024x512 shape of image on TITAN xp, and the FPS is about 100. Your paper report 112 fps on the NVIDIA TitanX. Do you have any skill, like absorving the batchnorm layers to conv layer? Besides, my cuda version is 9.0 and cudnn is 7.1.

sacmehta commented 5 years ago

No, we don’t combine batch norm withconv layers. Are you discarding the first iteration?