YOLOv5 Object Detection for RarePlanes Tutorial

This repo contains four notebook tutorials in which you create a custom class using the RarePlanes data set, train a YOLOv5 model, perform inferences on the test set, and then evaluate performace.

How to use this repo:

You can either access the tutorial pipeline hosted on AWS by accessing the AMI here or reproduce the enviornment using your own GPUs by cloning this repository

A. AMI/EC2

For the AMI, all the relevant data and packages have been downloaded so you should easily be able to spin up and follow along with the tutorial. For more informaion on how to spin up an AMI, please visit the first half of this blog

1. Spin up the AMI instance from here
2. Ensure you are in the N. Virginia and then hit Launch Instance
3. Search for the pre-built AMI titled CosmiQ_YOLO_Planes
4. Select the p3.2xlarge instance type and hit the Review and Launch, Launch buttons succesively
5. Create a new key pair and download and launch this key pair
6. SSH into the machine using your address. The command should look like this ssh -i "cosmiq-yolo-planes-aws.pem" [email protected]
7. Navigate to the directory /home/ubuntu/src/yolo_planes/yolov5/
8. Launch the jupyter lab using jupyter lab --ip=0.0.0.0
9. Open a browser and insert the EC2 ip into the address bar; it should look like this ec2-3-235-146-223.compute-1.amazonaws.com:8888
10. The password for the jupyter lab is yoloplanes
11. Open the notebook titled 1_yolo_start.ipynb

B. On your own GPUs

1. Clone this repository
2. Download the data from here. You will only need the real data for this tutorial
3. You will need to sort the images from the PS-RGB_tiled directory per yolo specifications with the following hiearchy:

class_one (or any other name) 
|--images (these are the downloaded tiled .pngs) 
|     |--train 
|     |--val
|--labels (you will create these during the tutorial)
|     |--train
|     |--val

• The easiest way to do this is to mkdir class_one outside of your yolov5 directory
• cd class_one and then mkdir images, mkdir labels
• cd images and then mkdir train, mkdir val
• In your images directory, mv the PS_RGB_tiled from the downloaded train directory to the one you just created and the test directory to the val directory you just created
• cd ../lablels and then mkdir train, mkdir val
• If you are creating your own custom class, save the image directory paths for your data/class_one.yamlfile (they should look something like ../class_one/images/train/ and ../class_one/images/val/)

4. Create your docker image using the command nvidia-docker build -t <name_of_image> ./
5. Your image should now appear when you run docker images
6. Then run NV_GPU=0,1 nvidia-docker run -it -v /dir/to/yolov5:/yolov5/ -p 9002:9002 --shm-size=64g --name <name_of_container> <name_of_image>
7. Navigate to the directory yolov5 directory on your GPU
8. Launch conda enviornment using conda activate solaris
9. Launch the jupyter notebook using jupyter notebook --ip 0.0.0.0 --no-browser --allow-root --port=9002
10. Open a browser and insert your ip into the address bar; it should look like this http://gpu02:9002/ Use the token supplied in the terminal as your password.
11. Open the notebook titled 1_yolo_start.ipynb

This ML pipeline uses a modified implementation of the YOLOv5 implementation found here. The full RarePlanes dataset can be found here and helper functions for the dataset can be found here.

If you have any questions or errors, please don't hesitate to post an issue or email me here.