About Inference Speed Regarding Yolov5l

[callbarian]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hi, I have question on an inference speed regarding batch 1 and 32.
My machine is Rtx3090 (cuda11.0, pytorch 1.7.1), and the image is on size 640.

Yolov5l.pt on batch 1 is about 13ms, which is comparable to your 10ms on v100.
However, on batch 32 is about 250ms, and thus about 7.8ms per image, which is not comparable to your 2.7ms

I expected the batch inference would give accelerated inference time, but not much improvement compared to batch 1.

detect.py --weights yolov5l.pt --source imagesource

Is there anything I am missing?

Thank you.

[github-actions]

👋 Hello @callbarian, thank you for your interest in YOLOv5 🚀! Please visit our ⭐️ Tutorials to get started, where you can find quickstart guides for simple tasks like Custom Data Training all the way to advanced concepts like Hyperparameter Evolution.

If this is a 🐛 Bug Report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset images, training logs, screenshots, and a public link to online W&B logging if available.

For business inquiries or professional support requests please visit https://ultralytics.com or email Glenn Jocher at [email protected].

Requirements

Python>=3.6.0 with all requirements.txt installed including PyTorch>=1.7. To get started:

$ git clone https://github.com/ultralytics/yolov5
$ cd yolov5
$ pip install -r requirements.txt

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Google Colab and Kaggle notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training (train.py), validation (val.py), inference (detect.py) and export (export.py) on MacOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

@callbarian from v6.0 release PR #5141:

About Reported Speeds

mAP values are reproducible across any hardware, but speeds will vary significantly among V100 instances, and seem to depend heavily on the CUDA, CUDNN and PyTorch installations used. The numbers reported above were produced on GCP N1-standard-8 Skylake V100 instances running the v6.0 Docker image with:

• NVIDIA Driver Version: 460.73.01
• CUDA Version: 11.2

# Pull and Run v6.0 image
t=ultralytics/yolov5:v6.0 && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all $t

Our speed command is:

# Reproduce YOLOv5s batch-1 speeds in table
python val.py --data coco.yaml --img 640 --task speed --batch 1

We tried several options, including AWS P3 instances, pulling more recent base versions (FROM nvcr.io/nvidia/pytorch:21.09-py3), different pytorch install methods, and many returned slower speeds than the values reported in our table. batch-32 speeds were found to vary less across options.

[yanisIk]

@callbarian Just by curiosity, what OS are you using ?

[callbarian]

@yanisIk I am using Ubuntu 18.04

@glenn-jocher Thank you for your reply. I have figured out the gap between yours and my experiment.
val.py uses the model.half().
I used val.py for measuring time and replicated close to your result.

Also, onnxruntime copies the values from cpu to gpu for input, and gpu to cpu for output.
This takes quite amount of time, whereas pytorch input is already on gpu.

pytorch model.half() is comparable to tensorRt float16 inference.