Benchmarking

[IanButterworth]

It would be great to check cross-platform benchmarking, for which a basic test can be done through:

pkg> add https://github.com/ianshmean/YOLO.jl
pkg> test YOLO

For my modern Macbook Pro (no CUDA GPU):

[ Info: YOLO_v2_tiny inference time per image: 0.1411 seconds (7.09 fps)
[ Info: YOLO_v2_tiny postprocess time per image: 0.0024 seconds (423.61 fps)
[ Info: Total time per image: 0.1434 seconds (6.97 fps)

cc. @Ybakman @iuliancioarca @stemann

[iuliancioarca]

Test Summary: | Pass Total
Example datasets have been downloaded | 1 1
Test Summary: | Pass Total
Example pretrained weights files have been downloaded | 1 1
[ Info: YOLO_v2_tiny inference time per image: 0.2287 seconds (4.37 fps)
[ Info: YOLO_v2_tiny postprocess time per image: 0.0048 seconds (207.02 fps)
[ Info: Total time per image: 0.2336 seconds (4.28 fps)
Test Summary: | Pass Broken Total
Loading and running YOLOv2_tiny_voc pretrained model | 9 2 11
Testing YOLO tests passed

Tested on Windows 10 with Julia 1.1.1 on i7 3740qm @3.6GHz, 16GB DDR3.
Runtime varies between roughly 200ms and 250ms.
Not sure what the two broken tests refer to...

[IanButterworth]

Thanks! I overlooked the broken tests. Should be fixed on master now

[Ybakman]

Hi guys,
I have been sick for a week so I couldn't answer. I tested the code on Amazon's p2x.large machine with CUDA GPU. Here is the result:

 % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100   621    0   621    0     0   4167      0 --:--:-- --:--:-- --:--:--  4167
100  423M  100  423M    0     0  66.5M      0  0:00:06  0:00:06 --:--:-- 67.5M
Test Summary:                         | Pass  Total
Example datasets have been downloaded |    1      1
Test Summary:                                         | Pass  Total
Example pretrained weights files have been downloaded |    1      1
Loading images...100%|██████████████████████████████████| Time: 0:00:19
[ Info: YOLO_v2_tiny inference time per image: 0.0047 seconds (213.23 fps)
[ Info: YOLO_v2_tiny postprocess time per image: 0.0057 seconds (175.35 fps)
[ Info: Total time per image: 0.0104 seconds (96.22 fps)
Test Summary:                                        | Pass  Total
Loading and running YOLOv2_tiny_voc pretrained model |   10     10
   Testing YOLO tests passed 

Also tested on my MacBookPro14,1.

Test Summary:                         | Pass  Total
Example datasets have been downloaded |    1      1
Test Summary:                                         | Pass  Total
Example pretrained weights files have been downloaded |    1      1
[ Info: YOLO_v2_tiny inference time per image: 0.1341 seconds (7.46 fps)
[ Info: YOLO_v2_tiny postprocess time per image: 0.0023 seconds (426.58 fps)
[ Info: Total time per image: 0.1364 seconds (7.33 fps)
Test Summary:                                        | Pass  Total
Loading and running YOLOv2_tiny_voc pretrained model |   10     10
   Testing YOLO tests passed 

[iuliancioarca]

Hi! I hope you're feeling well! I think the 2x performance decrease on my side is due to the older cpu with avx (while both of you have avx2). Keeping this in mind,the performance should be similar on mac/windows.