Can the pytorch implementation be as fast as the auther's implementaion（42 FPS）?

[sisrfeng]

I convert the demo.ipynb to demo.py and test an image, it cost 9s if just use cpu and 12s if using gpu.
Can the pytorch implementation be as fast as the auther's implementaion which can get a detection speed of more than 42 FPS?
I know the visiualizion may be time consuming.

I get the time by following:

import datetime
starttime = datetime.datetime.now()

...
original code
...

endtime = datetime.datetime.now()
print ((endtime - starttime).seconds)

I am not sure whether it is correct. Should I put starttime = datetime.datetime.now() before loading a picture and ignore the time cost on loading the model and other place?

I am a beginner and hope not to bother motokimura. Could anyone knows the answer give me some help?
Many thanks!

[sisrfeng]

I find that testing on thousands of pictures, the time cost on loading the model and other place can be ignore.