Training at 1280x720 about yolov5

[github-actions]

👋 Hello @miles-codes, 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 [email protected].

Requirements

Python>=3.7.0 with all requirements.txt installed including PyTorch>=1.7. To get started:

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

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]

@miles-codes training settings are optimized for best results, I would leave them alone. If you want to train at native resolution then just do python train.py --img 1280.

YOLOv5 PyTorch model inference already exploits rectangular shapes, so if you run python detect.py --img 1280 your image short side will automatically use 720.

[miles-codes]

I am aware of detect.py which uses letterbox to resize/pad input image to 1280x1280.
But my problem is that resizing/padding is slow, so I wanted to get rid of it by training on 1280x720

Thanks for a quick response anyway :)

[glenn-jocher]

@miles-codes you're not understanding. If you pass a 1280x720 image to python detect.py --img 1280 then it already by default exploits rectangular inference for the minimum possible FLOPs.

[glenn-jocher]

@miles-codes also this is already clearly displayed in the console:

[miles-codes]

okay, thank you for the clarification :)

[HeChengHui]

@glenn-jocher
if my detection input is 1920x1080 but my training images are of different sizes, is --imgsze 1920 --rect correct?

[glenn-jocher]

@HeChengHui if you want to detect at --img 1920 then train at --img 1920. It's very simple. Forget about any other options.

[HeChengHui]

@glenn-jocher Thank you!

A side thought, if detecting at 1920 is too slow and I changed detect size to 960, what would be the performance impact with a model trained at 1920?

[glenn-jocher]

@HeChengHui README Table curves shows models trained at 1280 running inference from 320 to 1280 image size.

[HeChengHui]

@glenn-jocher
Is it correct to read it as each dot of each curve representing an inference at different img-size starting from 320 at the bottom and 1280 at the top?

[glenn-jocher]

@HeChengHui yes!

[HeChengHui]

@glenn-jocher Thank you for the guidance!

[wwcc1107]

if I want to detect at --img 1024 ,but train at --img 256，does it work?

[glenn-jocher]

@wwcc1107 depends on your definition of work

[stphtan94117]

i have a dataset from a lot of different img size.
their size are 640x480、1280x720 and 1920x1080 .
if i want to detect 1920x1080 size video, what do i set --img size in train.py ?
--imgsize 1920 is more better than 1280 or 640?

another question
if i want to detect 1920x1080 size video,so i set --imgsize 1920 in detect.py , right?
thanks.

[glenn-jocher]

@stphtan94117 if your objects are large then training at smaller images sizes will be fine, if they are small then larger images sizes up to native may be better. You can set any training --imgsz you want like this:

python train.py --imgsz 640