Deep Learning with Keras on Google Compute Engine

Inception V3 architecture

Inception, a model developed by Google is a deep CNN. Against the ImageNet dataset (a common dataset for measuring image recognition performance) it performed top-5 error 3.47%.

In this tutorial, you’ll use the pre-trained Inception model to provide predictions on images uploaded to a web server.

Demo

You’ll use:

• Keras for image prediction running on Google Compute Engine
• Google Cloud Storage to store the uploaded images
• App Engine w/Flask for your front-end web server.

Overview

• Keras Inception V3 Prediction
• Deploy on Compute Engine w/ Docker & Google Container Registry
• Flask Front-end Webserver

Prereqs:

• Basic machine learning understanding
• Basic Flask understanding
• Google Cloud SDK Installed (https://cloud.google.com/sdk/)

Keras

Keras is a high-level neural networks library, written in Python and capable of running on top of either TensorFlow or Theano. It was developed with a focus on enabling fast experimentation. Being able to go from idea to result with the least possible delay is key to doing good research.

Below, the Inception model is loaded with Keras. Keras holds a cache directory of the models pre-trained weights. On first use, Keras will download these weights into ~/.keras/models/.

You’ll create a predict function with accepts a base64 encoded image file. Inception V3 requires images to be 299 x 299. After loading the image, it is expanded into a vector and pre-processed.

First, install the dependencies

$ pip install -r requirements.txt

Serving Predictions

With the model loaded, create a web server that can accept base64 encoded images using flask.

You now have image-recognition as a service! Let’s test it out locally before deploying. First, start the server. Then make a POST request to the prediction service. Sample image

$ python main.py
(in another terminal window)
$ (echo -n '{"data": "'; base64 cat.jpg; echo '"}') |
curl -X POST -H "Content-Type: application/json" -d @- http://0.0.0.0:8080

If everything is working correctly, you’ll see a response back

{
 “predictions”: [
 {
 “description”: “tiger_cat”,
 “label”: “n02123159”,
 “probability”: 55.242210626602173
 },
 {
 “description”: “tabby”,
 “label”: “n02123045”,
 “probability”: 25.407189130783081
 },
 {
 “description”: “Egyptian_cat”,
 “label”: “n02124075”,
 “probability”: 10.042409598827362
 }
 ]
}

Deployment

You’ve got the model created and generating predictions. Time to deploy the model to Google Compute Engine using Docker. You’ll use a few different technologies Gunicorn, Nginx, and Supervisor. Below, a dockerfile and a few configuration files will set up and serve the prediction api.

After creating the docker image, you’ll push it to Google Container Registry.

For docker to copy the files correctly, structure the directory:

|____Dockerfile
|____nginx.conf
|____requirements.txt
|____service
| |____main.py
| |____model.py
|____supervisord.conf

You’ve built and pushed the docker image to Google Container Registry. From here, create the server and pull down the previously created docker image. First, you’ll enable the API

https://console.developers.google.com/apis/api/compute_component/overview?project=PROJECT_ID

$ gcloud compute firewall-rules create default-allow-http --allow=tcp:80 --target-tags http-server
$ gcloud compute instances create predict-service --machine-type=n1-standard-1 --zone=us-central1-c --tags=http-server
$ gcloud compute ssh predict-service --zone=us-central1-c
$ curl -sSL https://get.docker.com | sh
$ sudo gcloud docker pull gcr.io/PROJECT_ID/predict_service:latest

Finally, run it!

$ sudo docker run -td -p 80:80 gcr.io/PROJECT_ID/predict-service

If all went well, after a short bit, you’ll have a running prediction service. You can use the curl command above to confirm it’s working. Let’s build a quick front-end to visualize our predictions.

Google App Engine Front End Server

Cloud Storage

Create a bucket to upload images received from the front-end. We modify the permissions of the bucket to give read access to anyone on the internet.

$ gsutil mb gs://my-unique-bucket-identifier

Add this snippet to your project. It is an abstraction around Google Cloud Storage you’ll use when uploading files. First, we’ll install our dependencies

requirements.txt

App engine requires libraries to be installed into a folder for deployment. You’ll download the Gcloud Storage Client as well.

$ git clone https://github.com/GoogleCloudPlatform/appengine-gcs-client.git
$ pip install -r requirements.txt -t lib
$ pip install GoogleAppEngineCloudStorageClient -t lib

Upload File & Predict

First, let’s create our config file. Here, you’ll insert your project and storage bucket you created earlier. You’ll also need to insert your prediction services IPv4 address. In the networking tab, click reserve static instance address on our machine.

Reserve static address for prediction service

config.py

PROJECT_ID = 'PROJECT_ID'
MAX_CONTENT_LENGTH = 8 * 1024 * 1024
ALLOWED_EXTENSIONS = set(['png', 'jpg', 'jpeg', 'gif'])
PREDICTION_SERVICE_URL = 'http://YOUR.SERVER.IP.HERE:80'

app.yaml

appengine_config.py

Below, you’ll create the front-end web server using flask. This web server will be deployed to Google App Engine.

Copy the below templates into /templates:

base.html

form.html

view.html

Last step, let’s deploy to google app engine.

gcloud app deploy

Conclusion

That’s it! You’ve created a flask web server that will present the user with a form to upload photos. These photos are uploaded to google cloud storage and sent to our image prediction API.

Complete Code here:

ColeMurray/keras-inception-v3-tutorial

git clone https://github.com/ColeMurray/keras-inception-v3-tutorial.git --recursive

If you liked the tutorial, follow & recommend!

Interested in recommendation systems, node, android, or react? Check out my other tutorials:

- Recommendation Systems with Spark on Google DataProc
- Deploy Node to Google Cloud
- Android Impression Tracking in RecyclerViews
- React & Flux in ES6

Other places to find me:

Twitter: https://twitter.com/_ColeMurray