README LambdaVision

This is a project to do image recognition using AWS Lambda and the Serverless Framework.

It's totally derived from a blog post by Michael Diets entitled "Serverless deep/machine learning in production -- the pythonic way".

Building Prerequisites: Makefile, Docker

In the post, Michael uses a script to build PyTorch and its prerequisites, some of which are C-based. Those require compilation on a Lambda-compatible machine. I wanted something a bit more automated, so I've turned that build into a Makefile which builds inside the enormously useful lambci docker image.

Warning: The build needs over 2GB RAM, and macOS Docker ran out of memory while copiling, with GCC reporting this inscrutable message:

error: command 'gcc' failed with exit status 4

I increased it by going to the Docker icon in my menubar, Preferences, Advanced, and setting the RAM to 4GB, then restarting Docker.

I wanted to commit the binaries and built Python libraries to this Git repo, but one of the compiled libraries was too big for Git:

remote: error:
File lambdavision/torch/_C.cpython-36m-x86_64-linux-gnu.so is 105.86 MB;
this exceeds GitHub's file size limit of 100.00 MB

So you're stuck building the prerequisites yourself. Since this build can take 11 minutes on my MacBookPro, the Makefile tries to be smart about not building stuff it sees is already there.

After building each piece, the Makefile puts everything in the Serverless Framework service directory, lambdavision.

Use the Serverless Framework to Make Life Easier

Michael's build process creates a ZIP bundle which can be uploaded to AWS Lambda, but then you're on your own to create S3 buckets, triggers, permissions and so on. I'm lazy, so I specify all this stuff in my serverless.yml file. Then I can deploy my Lambda function along with all the requisite PyTorch binaries and libraries in one step:

sls deploy

Normally, with Serverless, you can redeploy the function by itself, without having to rebuild any of the infrastructure like:

sls deploy function --function lambdavision

Unfortunately, AWS has limit on size which we exceed:

Serverless: Packaging function: lambdavision...
Serverless: Excluding development dependencies...
Serverless: Uploading function: lambdavision (66.79 MB)...
Serverless Error ---------------------------------------
Request must be smaller than 69905067 bytes for the UpdateFunctionCode operation

So you have to do the full deploy each time; it's not a big deal, because AWS knows the CloudFormation-based infrastructure hasn't changed.

Update the Lambda Function to be More Descriptive

Michael's original Lambda function simply returned the model_output, which is a 1000-element tensor. But in Lambda, async functions don't "return" anything to the user so there was no output. I printed it, and the prints get sent to the AWS CloudWatch logs, which Serverless lets you watch easily with:

sls logs --tail --function lambdavision &

That output tensor, truncated in the display, didn't tell me anything useful; it appeared to be just a list of scores -- how well each pre-trained model classification matched the submitted image. I dug around a while and determined that the model was pre-trained on an ImageNet sample of 1000 images; some more sleuthing turned up a useful class-id to human-name mapping. I've never played with PyTorch or tensors before so it took me a while to figure out how to index the best score, but now it displays the score, index, and human friendly class name.

We can use the AWS CLI to upload a picture to S3:

s3 cp ~/Downloads/squirrel.jpeg    s3://chris-lambdavision/

and watch what it reports in the tailed logs:

score=11.207955360412598 classid=335 imagenet human=fox squirrel, eastern fox squirrel, Sciurus niger

If upload a picture of myself wearing a big white hat and it says:

score=9.719132423400879 classid=515 imagenet human=cowboy hat, ten-gallon hat

I can upload a picture of an audio tape cassette found on the net and it says:

score=13.65023136138916 classid=481 imagenet human=cassette

It took about 8 seconds of Lambda-time to do this:

REPORT RequestId: bd4c7ded-23cb-11e8-ba21-2116960e04ca
Duration: 8073.79 ms  Billed Duration: 8100 ms
Memory Size: 2048 MB  Max Memory Used: 276 MB

It's not AWS Rekognition but it works.

Our future work will require training models based on specific and unusual subject matter which our client deems sensitive, so Rekognition may not be a viable option for us, and this Lambda-based classifier shows some promise... or at least a start.