weiliu89 / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
4.77k stars 1.68k forks source link

Inference time for single image without lmdb format #363

Open XiongweiWu opened 7 years ago

XiongweiWu commented 7 years ago

Hi, thx for your excellent work! When I use SSD-512 for inference with lmdb format, the test time is almost identical to your paper report(19FPS), however, when I use your ssd_detect_ipynotebook script for a single image inference instead of transform it into lmdb format, the time is much larger (136s for 1000iteras) even excluding time for transforming jpg to blob format. I wonder whether lmdb can largely improve the inference speed?

weiliu89 commented 7 years ago

I guess there are some overhead of calling the Forward() function in python. I don't know the answer. You could write a simple python script to try again (instead of using the .ipynb)