YOLOv5 container on AWS Marketplace

[g0lemXIV]

Search before asking

• I have searched the YOLOv5 issues and found no similar feature requests.

Description

I want to write an article about SageMaker model training especially Object Detection models. The choice fell on yolov5 because you've done fantastic work! Is it possible to make this a feature for the repository?

I can make a complete pipeline for training the yolov5 model and deploying it as an endpoint/batch transform as containers available on AWS Marketplace.

Use case

Training:
A user wants to train the yolov5 model on AWS Sagemaker but doesn't have experience. Users can "fork" ready-to-use containers from the marketplace and use them for training.

Deploying:
A user can use the yolov5 container to deploy a custom or pretrained model with a few clicks.

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[github-actions]

👋 Hello @g0lemXIV, 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 support@ultralytics.com.

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]

@g0lemXIV this sounds like an awesome idea! I haven't used sagemaker myself, I've been to busy drinking coffee and making the world's best vision AI over here on GitHub haha.

We do have Docker images ready to go here. These are created with Docker Autobuild and are always up to date with master:
https://hub.docker.com/r/ultralytics/yolov5

We also have a quick AWS tutorial meant for using YOLOv5 on EC2:
https://docs.ultralytics.com/yolov5/environments/aws_quickstart_tutorial/

But no sagemaker or marketplace images! What do we need to do to get started? Could you help explain the steps and what's required from us? Thanks!

EDIT: I've reached out to our Amazon contact to see if they can help us get registered on Marketplace so we can provide an official container.

[g0lemXIV]

@glenn-jocher basically, SageMaker is a complete data science platform with data labelling, training, deploying and hyperparameters optimization. You also have standard notebooks to prototype. So it is possible to get a fully working model within the marketplace, train it or deploy it in a few minutes.

Because you've done fantastic work with torch hub for deployment, we need only one script similar to this aws-code

The training differs because we need a specific setting through the .sh / .py script, but I've done similar work in the past.

We can make deployment scripts and add them to the aws marketplace it may work similarly to mxnet-yolov3

Where in settings, users can:

• Use a custom model
• Choose architecture and any parameters
• Choose between return tagged images or JSON file

Then I think we can take care of the training container. What do you think?

[glenn-jocher]

@g0lemXIV I've reached out to AWS but am having trouble getting their team to return emails. I don't want to drag our feet on this since it seems like a great idea. Can you take the lead on this and let me know what next steps are and what I can do to help?

@g0lemXIV Hi, have you managed to deploy yolov5 to sagemaker ?

[g0lemXIV]

@sinatimelapse it is relatively easy with yolov5 torch hub and this script inference.py for pytroch

@glenn-jocher sorry that I didn't respond. I had a busy month...
I think I will do the commitment in the next two weeks for inference script and then for training.

[glenn-jocher]

@g0lemXIV great thanks!

@g0lemXIV Could you elaborate more please ? I've never used sagemaker !

[g0lemXIV]

@sinatimelapse please read a few examples about sagemaker, sagemaker SDK and then look at the inference.py again. It will be pretty easy tho to do this. Otherwise, you can wait for scripts for yolov5, I can notify you when I pr.

(edit) You have there a full example in a jupyter notebook with the code. code for pytorch inference
You need to implement image loading, loading model from torch hub, make inference and finally return JSON with bbox or image