Yolo V4 on Tensorflow First Inference is very slow take minutes on NVIDIA GPU A30

Description

I’m running yolov4 on NVIDIA A30 but the first inference is very slow take 15-20 minutes.

Environment

GPU Type: NVIDIA A30
Nvidia Driver Version: 512.78
CUDA Version: 10.1
CUDNN Version: 7.6
Operating System + Version: Windows Server 2016
Python Version (if applicable): Python 3.7
TensorFlow Version (if applicable): 2.3.0

Relevant Files

I’m use these code for testing: GitHub - theAIGuysCode/yolov4-deepsort: Object tracking implemented with YOLOv4, DeepSort, and TensorFlow.

Is anyone please give me a suggestion to accelerate first inference ?

Hi,

This looks more within the scope of the Tensorflow. We recommend you to please reach out on Issues · theAIGuysCode/yolov4-deepsort · GitHub or https://discuss.tensorflow.org/ to get better help.

Thank you.

Hi Spolisetty,

Thanks for your response.

It was Cuda version for Ampere architecture, I changed Cuda 10 Version to Cuda 11 and Cudnn 7.6 to Cuda 8.2. Now is fast less than 30 seconds the first inference, the others inference take 20 FPS aprox (I thought A30 would be more fast)

For reference I read in somewhere where recommend you will use Cuda 10 for Touring architecture and Cuda 11 for Ampere architecture, it worked for me

Environment

GPU Type: NVIDIA A30
Nvidia Driver Version: 512.78
CUDA Version: 11.0
CUDNN Version: 8.2
Operating System + Version: Windows Server 2016
Python Version (if applicable): Python 3.7
TensorFlow Version (if applicable): 2.4.0 (in specific tensorflow-gpu)

Reference

To select cuda version to tensorflow I guided GPU table of Build from source  |  TensorFlow

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.