Is the performance too bad?

Question

Is the performance too bad?

HW140701 opened this issue 3 years ago · comments

Is the performance too bad? It takes 1 - 2 seconds to predict a image on GTX 1070 8G. Is there a plan to optimize performance and increase the speed of predicit?

Hzzone · Answer 1 · Mon Apr 05 2021 22:03:59 GMT+0800 (China Standard Time)

well, I will update this repo with parallel processing and a new hand detection model. Waiting for the time I am free.

HW140701 · Answer 2 · Tue Apr 06 2021 08:54:55 GMT+0800 (China Standard Time)

well, I will update this repo with parallel processing and a new hand detection model. Waiting for the time I am free.

Thanks ! Expect

Risith Perera · Answer 3 · Wed Apr 07 2021 02:43:26 GMT+0800 (China Standard Time)

@Hzzone Can you please tell me which kind of optimizations did you used? When are you able to update the repo?
Thanks for you hard work!

Hzzone · Answer 4 · Wed Apr 07 2021 12:26:22 GMT+0800 (China Standard Time)

I mean that, the major consumption is the hand detection model which I have to use the pose to estimate. It would be better if I change it by training a mobile net hand detection model, and process the images with multi-thread.

Risith Perera · Answer 5 · Wed Apr 07 2021 12:47:53 GMT+0800 (China Standard Time)

@Hzzone In the body.py there are lots of calculations after using the model. These calculations take lots of time.

data = torch.from_numpy(im).float()
if torch.cuda.is_available():
    data = data.cuda()

with torch.no_grad():
    Mconv7_stage6_L1, Mconv7_stage6_L2 = self.model(data)

Do you have any idea to do these calculations on GPU? In my application, I crop each person in the frame using YOLO and send each person to this model. Can you give me any hint on how to speed up this little more.

HW140701 · Answer 6 · Wed Apr 07 2021 14:13:54 GMT+0800 (China Standard Time)

@Hzzone In the body.py there are lots of calculations after using the model. These calculations take lots of time.
data = torch.from_numpy(im).float()
if torch.cuda.is_available():
    data = data.cuda()

with torch.no_grad():
    Mconv7_stage6_L1, Mconv7_stage6_L2 = self.model(data)
Do you have any idea to do these calculations on GPU? In my application, I crop each person in the frame using YOLO and send each person to this model. Can you give me any hint on how to speed up this little more.

Alphapose maybe faster ? It gets 20 FPS on GTX 1080Ti.

Risith Perera · Answer 7 · Wed Apr 07 2021 14:22:16 GMT+0800 (China Standard Time)

@HW140701 Is there a PyTorch implementation of Alphapose?

HW140701 · Answer 8 · Wed Apr 07 2021 14:23:49 GMT+0800 (China Standard Time)

@HW140701 Is there a PyTorch implementation of Alphapose?

maybe you can see https://github.com/MVIG-SJTU/AlphaPose

notsong · Answer 9 · Mon Aug 02 2021 15:20:20 GMT+0800 (China Standard Time)

@Hzzone In the body.py there are lots of calculations after using the model. These calculations take lots of time.
data = torch.from_numpy(im).float()
if torch.cuda.is_available():
    data = data.cuda()

with torch.no_grad():
    Mconv7_stage6_L1, Mconv7_stage6_L2 = self.model(data)
Do you have any idea to do these calculations on GPU? In my application, I crop each person in the frame using YOLO and send each person to this model. Can you give me any hint on how to speed up this little more.

Hello, has your problem been solved?