Tensorflow-Lambda

What's going on here?

This is a script to create an AWS Lambda zipfile that can run keras models.

Lambda is a really great target for small, lightweight machine learning inference. However, there's a 250 MB limit on lambda zipfiles, and tensorflow passes that limit in its natural state. This script builds an AWS instance that builds a minified version of TF and all dependencies that weighs about 195 MB, which leaves you about 55 MB for your code and model.

How do I use it?

First, if your dependencies are only keras and TF, try the prebuilt tar archive

If that works, great. If not, try running the script manually as described below (requires an AWS account, and uses some non-free EC2 compute).

Then what?

If you used the tar archive,

zip -r9 lambda.zip *
aws s3 cp lambda.zip s3://mybucket/lambda.zip

If you used the script, your zipfile will already be uploaded to a bucket.

Then, you go to the lambda console, create your function, point it at the S3 zip, and you're off!

At that point, you might want to consider [exposing your function with an API gateway endpoint] (https://docs.aws.amazon.com/apigateway/latest/developerguide/api-gateway-create-api-as-simple-proxy-for-lambda.html)

Running the script

Note: this will cost about 10 minutes of EC2 t2.small

I've found it runs out of memory compiling on anything smaller

1. Install AWS-cli and jq
2. Clone this repository
3. Edit aws-build-lambda.sh at the top to set 4 environment variables, describing your environment and how you'll connect to Amazon
4. Edit aws-build-lambda.sh at line 180 if you need any python libraries not included
5. Place your code, model, and any other requirements in the src repository -- there's an example classify.py already there
6. Create test.py, which should cover your code. This script will read the files accessed when test.py is run, so try to cover everything you'll be using
7. Run ./aws-build-lambda.sh
8. After it's done, whether it's succeeded or not, REMEMBER TO TERMINATE YOUR INSTANCE. This script spawns a t2.small, which costs about $20 per month if you forget to turn it off.

How does it work?

If running it automatically doesn't work, you could try following along the script and running the commands manually.

This script:

• Sets up an EC2 instance / security group
• Sends your code to that instance
• Installs Python
• Installs Tensorflow and Keras
• Runs test.py to run your code while watching all the files Python accesses (inotify)
• Copies over all the accessed files to the lambda zip root
• Copies over all ".py" sources to the lambda zip root
• Strips unneeded symbols from copied ".so" files
• Zips the zip root and sends it to S3

As long as you can do all of the above, it should work out.

It broke!

ELF load command address/offset not properly aligned

• Add the broken .so file to aws-build-lambda.sh line 255. It won't be shrunk, but that might fix this error.

Some other error

• Let me know! No guarantees, but I might be able to help.