Running YOLOv3 from Darknet on AWS Lambda FaaS (Function as a Service)
Big thanks to Redmon, Joseph and Farhadi, Ali for darknet https://pjreddie.com/darknet/yolo/ and Rustem Feyzkhanov for his https://github.com/ryfeus/lambda-packs repo which I used as inspiration for this.
Everything is already prepared in the zip file darknetmin.zip and it will use my S3 to pull the weights and sample image from. The code that will execute is in the service.py file in the handler(event, context) function.
- Create an AWS account.
- Create a Role in IAM, Under Services select IAM then select Roles from the sidebar. Click Create Role, Select Lambda, click next. Select permissions - I gave it AWSLambdaFullAccess, AmazonS3FullAccess and CloudWatchFullAccess which is probably more than needed but fine for a demo. Click next. Give the role a name calling it something sensible like 'LambdaFullPlusS3'. Click Create.
- Under Services select Lambda then 'Create Function'
- Author from scratch
- Give it a Name
- Runtime = Python 3.8
- Choose Existing Role
- Select the role you created in step two.
- Click 'Create Function'
- Under Function Code - Click 'Upload From' then select '.ZIP File'
- Under Function Package, click the Upload button. Select the zip file from the repo or the one you created to replace it (darknetmin.zip if using mine)
- Click save. Takes a second or two to upload.
- Under Runtime settings (you find this below the code editor) click edit and Change Handler to service.handler (this is because you want the code to execute service.py and the function called handler we defined)
- Set up a test event
- Click Test
- Create new test event
- Event Template - Hello World
- Event Name - 'BasicTest'
- Delete the key value pairs leaveing just this in the test event:
{
}
- Click Create.
- Configure Basic Settings
- Move the slider to 2048Mb (You can go lower but it goes slower)
- Change Timeout to 1 minute.
- Click Save
- Click Test
- Check your logs at the top and expand to see something like this:
26 x 256 0.177 BFLOPs
57 conv 512 3 x 3 / 1 26 x 26 x 256 -> 26 x 26 x 512 1.595 BFLOPs
58 res 55 26 x 26 x 512 -> 26 x 26 x 512
59 conv 256 1 x 1 / 1 26 x 26 x 512 -> 26 x 26 x 256 0.177 BFLOPs
...
103 conv 128 1 x 1 / 1 52 x 52 x 256 -> 52 x 52 x 128 0.177 BFLOPs
104 conv 256 3 x 3 / 1 52 x 52 x 128 -> 52 x 52 x 256 1.595 BFLOPs
105 conv 255 1 x 1 / 1 52 x 52 x 256 -> 52 x 52 x 255 0.353 BFLOPs
106 yolo
Loading weights from /tmp/yolov3.weights...Done!
Failed to write image predictions.png
/tmp/inputimage.jpg: Predicted in 11.440608 seconds.
motorbike: 57%
car: 99%
car: 94%
car: 76%
person: 97%
person: 97%
person: 95%
person: 94%
person: 90%
person: 84%
person: 63%
END RequestId: e13f4826-6dbc-11e8-af3c-1bdeaebb4bdf
REPORT RequestId: e13f4826-6dbc-11e8-af3c-1bdeaebb4bdf Duration: 19339.67 ms Billed Duration: 19400 ms Memory Size: 2048 MB Max Memory Used: 1210 MB
Nip over to cloud watch to debug if anything goes wrong. You now have darknet YOLO working on a platform that can scale to hundreds or thousands of concurrent instances and at a price that is so low. If you manage to get out of the 'free tier' it still works out at around 8k images processed for every dollar. Look up Youtube for guides on how to trigger this Lambda function from a file arriving in S3 and build an awesome on demand object detector. Redirect the output to a file? Lots of great options for using this.
To get this running locally you need a copy of the weights file for YOLO 3 - yolov3.weights as described on https://pjreddie.com/darknet/yolo/ or just run
wget https://pjreddie.com/media/files/yolov3.weights
You also need to create or unzip labels for the output image. You'll find them in data/labels/labels.zip and can either unpack directly into the data/labels or run the generation script:
/data/labels$ python make_labels.py
If you don't have the font required it still works or you can modify font = 'futura-normal' to any valid font.
To execute it locally with a sample image just run:
$ wget https://s3-eu-west-1.amazonaws.com/myqueuecounter/darknet/street.jpg
$ ./darknet detect cfg/yolov3.cfg yolov3.weights street.jpg
You should see output like:
street.jpg: Predicted in 26.313906 seconds.
motorbike: 57%
car: 99%
car: 94%
car: 76%
person: 97%
person: 97%
person: 95%
person: 94%
person: 90%
person: 84%
person: 63%
And it will generate a file called predictions.png which will look like:
To get it ready for upload just delete my zip file and then zip all the files and directories into a new zip file.
I'd encourage you to create your own S3 bucket to store the weights file and your input image. Lots of good guides on youtube showing how to set up a S3 trigger to kick off your lambda function. The main file to edit is service.py changing the bucket name and keys to match your bucket name.