how to get the test time about centernet HOT 5 CLOSED

Hi,
Can you print the detailed time as in https://github.com/xingyizhou/CenterNet/blob/master/src/demo.py#L31 . The reported 19ms only includes network forwarding time, decoding time, and nms time (not applicable for centernet), as is indicated in the model zoo and following yolov3 (we move their time to after nms). This is also what we did for reporting the time for other methods. In your case, reading and resizing the image might take a lot of time, since the example images are larger than 512x512. If you really want to speedup the whole pipeline, you can use the multi-thread image pre-processing as in our testing script or use other data pre-processing library (e.g. NIVIDIA-dali) other than opencv for your application.

from centernet.

@xingyizhou ， thank you
i print the detailed time like in the demo.py, and i got it. the upper avg 90ms is contains the load and pre time, and the first debuger new time, which consume abount 0.5 second.

debugger = Debugger(dataset=self.opt.dataset, ipynb=(self.opt.debug==3),
                        theme=self.opt.debugger_theme)

the time:

18124840932_e42b3e377c_k.jpg
tot 0.039s |load 0.007s |pre 0.007s |net 0.022s |dec 0.001s |post 0.002s |merge 0.000s |
24274813513_0cfd2ce6d0_k.jpg
tot 0.039s |load 0.007s |pre 0.007s |net 0.022s |dec 0.001s |post 0.002s |merge 0.000s |
33823288584_1d21cf0a26_k.jpg
tot 0.039s |load 0.007s |pre 0.007s |net 0.022s |dec 0.001s |post 0.002s |merge 0.000s |

the operation net , deocde, post consume abount 25ms,

my env:
ubuntu 18.04
gtx 1080 ti
cuda 10.0
pytorch 1.1

from the quora disscuss, 1080ti is 11% slower than titan xp ([https://www.quora.com/Is-the-Titan-X-better-than-the-1080-ti-for-deep-learning-and-if-so-why]). so 22ms is expceted.

is your test time like following?
net 0.017s | dec 0.001s | post 0.001s

from centernet.

Good. This sounds like a reasonable difference between different machines (on our machine, 1080 Ti runs at 21ms). Also, make sure that you install opencv via pip install opencv-python. I observe conda install opencv gives a slower version ......

from centernet.

Oh the exact time we reported in the paper is measured with test.py. demo.py also includes visualization time. You can run python test.py ctdet --load_model ../models/ctdet_coco_dla_2x.pth --keep_res for a strict re-evaluation.

from centernet.

@xingyizhou thanks for your great work.
win 10.0
pytorch 1.0.3
gtx 1080 ti
cuda 9.0
cudnn 7.0
my test time：
tot 0.0501s |load 0.0000s |pre 0.0175s |net 0.0286s |dec 0.0020s |post 0.0015s |merge 0.0005s |
tot 0.0506s |load 0.0000s |pre 0.0180s |net 0.0281s |dec 0.0030s |post 0.0015s |merge 0.0000s |
tot 0.0521s |load 0.0000s |pre 0.0180s |net 0.0296s |dec 0.0025s |post 0.0020s |merge 0.0000s |
tot 0.0526s |load 0.0000s |pre 0.0191s |net 0.0291s |dec 0.0025s |post 0.0015s |merge 0.0005s |
tot 0.0501s |load 0.0000s |pre 0.0175s |net 0.0286s |dec 0.0020s |post 0.0015s |merge 0.0005s |
tot 0.0496s |load 0.0000s |pre 0.0175s |net 0.0271s |dec 0.0025s |post 0.0020s |merge 0.0005s |
tot 0.0516s |load 0.0000s |pre 0.0175s |net 0.0291s |dec 0.0030s |post 0.0020s |merge 0.0000s |
tot 0.0491s |load 0.0000s |pre 0.0175s |net 0.0276s |dec 0.0025s |post 0.0015s |merge 0.0000s |
tot 0.0496s |load 0.0000s |pre 0.0170s |net 0.0281s |dec 0.0025s |post 0.0015s |merge 0.0005s |
tot 0.0506s |load 0.0000s |pre 0.0180s |net 0.0281s |dec 0.0025s |post 0.0020s |merge 0.0000s |

from centernet.