How can performance be improved ?

Question

How can performance be improved ?

orgicus opened this issue 6 years ago · comments

Hi,

Thank you for sharing the repository.
It's great there's a demo ready to run already.

Unfortunately I ran into some (memory) issues with TensorFlow 1.6 with (a 2GB) GPU.
I could run the demo using TensorFlow 1.4 CPU, but I get about 1fps.
I've tried reducing the webcam resolution as much as possible, but got no visible improvements.

How did you manage to get ~40fps ?

Thank you,
George

asifzhcet11 · Answer 1 · Wed Jul 25 2018 22:52:04 GMT+0800 (China Standard Time)

Hi George,

First you should first install the GPU version of the TensorFlow and run it. For me with GPU version the FPS is around 45 :) and with CPU its ~1 fps

Thanks!

Asif

George Profenza · Answer 2 · Thu Jul 26 2018 04:49:10 GMT+0800 (China Standard Time)

Hi @asifzhcet11

Thank you very much for the feedback.
I'll try to compile TensorFlow 1.4 with GPU support on OSX and see how I fare.

By any chance have you tried exporting the model for Android ?

Thank you,
George

asifzhcet11 · Answer 3 · Thu Jul 26 2018 05:14:12 GMT+0800 (China Standard Time)

Hi George,

The macbook normally comes with an AMD GPU which is not supported by Tensorflow yet. So you have to test this on machines with Nvidia GPU.

And I haven't tried to work with this model for Android

Thanks!

Asif

George Profenza · Answer 4 · Thu Jul 26 2018 19:06:20 GMT+0800 (China Standard Time)

Hi Asif,

Thanks again for the clarifications.

I'm using an older macbook with a 2GB nVidia GPU.
It's a shame I got OOM errors with Tensorflow 1.6 GPU version.
(Will have to build Tensorflow 1.6 GPU from source)
What GPU are you using ?

Many thanks,
George

asifzhcet11 · Answer 5 · Thu Jul 26 2018 22:31:33 GMT+0800 (China Standard Time)

Hi George,
Did you install the correct CUDA and cuDNN version for your GPU. You can give it a try by using conda and the installing below package for the complete installation of TensorFlow 1.4 with CUDA and cuDNN

conda install -c cjj3779 tensorflow-gpu

https://anaconda.org/cjj3779/tensorflow-gpu

I am using GeForce GTX 1080

Thanks!

Asif

George Profenza · Answer 6 · Mon Jul 30 2018 23:51:55 GMT+0800 (China Standard Time)

Hi Asif,

I'm using CUDA 8.0.61 and cuDNN 5.1.10.

I can't use pip nor conda to install tensorflow-gpu as I'm using OSX and these options work only on Windows and Linux.

At the moment I can only think of compiling tensorflow 1.4 with GPU support from source just to test if I can get close to 40fps.

Thank you,
George

asifzhcet11 · Answer 7 · Tue Jul 31 2018 06:09:04 GMT+0800 (China Standard Time)

Hi George,

Please install conda as described below in the link and I hope you will be good to go :)

https://conda.io/docs/user-guide/install/macos.html

Thanks!

Asif

Kautuk Kundan · Answer 8 · Sun Jan 06 2019 03:15:49 GMT+0800 (China Standard Time)

Hello everyone
I am using a GTX 960 card but I still get about 4FPS

I have correctly set up CUDA and CUDnn
even nvidia-smi shows 3.7Gigs usage in GPU

can i do something to get better frame rates?

Qinwen · Answer 9 · Thu Jan 17 2019 12:45:15 GMT+0800 (China Standard Time)

Hello everyone
I am using a GTX 960 card but I still get about 4FPS

I have correctly set up CUDA and CUDnn
even nvidia-smi shows 3.7Gigs usage in GPU

can i do something to get better frame rates?
It is impossible, I am using GTX960M with about 12FPS, maybe you should Set FLAG to the correct parameter

DilaraAlbayrak · Answer 10 · Thu Feb 21 2019 16:20:57 GMT+0800 (China Standard Time)

Hi, I am running the code on ubuntu 16.04, I can run the code with tensorflow. I download tensorflow-gpu by using conda install -c cjj3779 tensorflow-gpu but, I have 'ImportError: libcublas.so.8.0: cannot open shared object file: No such file or directory' error. How can I run the code with tensorflow gpu?

asifzhcet11 · Answer 11 · Thu Feb 21 2019 16:47:25 GMT+0800 (China Standard Time)

Hi @DilaraSina

It could be because of the path and library path is not set to the cuda 8.0 specific directories. If not present just add the path and give it a try.

Thanks,
Asif

DilaraAlbayrak · Answer 12 · Thu Feb 21 2019 21:03:35 GMT+0800 (China Standard Time)

Thanks, I run on gpu and reached ~35 fps

Lang · Answer 13 · Sun Mar 03 2019 17:06:51 GMT+0800 (China Standard Time)

Hi, @asifzhcet11
I'm sorry to trouble you. Do you know where I can get the training data ,I want to retraining cpm_hand model.

asifzhcet11 · Answer 14 · Sun Mar 03 2019 19:11:26 GMT+0800 (China Standard Time)

Hi @liang0724s,

Below is the dataset I believe is used. I am not certainly sure but the guys who developed the convolution pose machines are from the same university and used datasets from the below resource.

http://domedb.perception.cs.cmu.edu/handdb.html

Thanks,
Asif

Lang · Answer 15 · Sun Mar 10 2019 16:55:33 GMT+0800 (China Standard Time)

Hi @liang0724s,

Below is the dataset I believe is used. I am not certainly sure but the guys who developed the convolution pose machines are from the same university and used datasets from the below resource.

http://domedb.perception.cs.cmu.edu/handdb.html

Thanks,
Asif

Hi @asifzhcet11

Thank you very much for the feedback.

Leo

Lang · Answer 16 · Tue Mar 12 2019 14:33:48 GMT+0800 (China Standard Time)

Hello everyone
I am using a GTX 960 card but I still get about 4FPS

I have correctly set up CUDA and CUDnn
even nvidia-smi shows 3.7Gigs usage in GPU

can i do something to get better frame rates?

Hello everyone
I am using a GTX 960 card but I still get about 4FPS
I have correctly set up CUDA and CUDnn
even nvidia-smi shows 3.7Gigs usage in GPU
can i do something to get better frame rates?
It is impossible, I am using GTX960M with about 12FPS, maybe you should Set FLAG to the correct parameter

Hi @kautukkundan ,
Did get a better frame rates by setting FLAG to the correct parameter ?

Hi @jgjxzqw ,
The compute capability of GTX 960 and GTX960M respectively are 5.2 and 5.0, the performance is not very bad, why the frame rate is so low?
I am using GT720M with about 2FPS, and its compute capability is 2.1.