Docker and Deep Learning for Pugs

Repo for the Strata+Hadoop World 2016 talk "Docker for Data Scientists".

Introduction

This repository comprises an end-to-end example of doing data science with docker. We begin by doing interactive analysis and modeling in a jupyter notebook. The task at hand is building a deep convolutional neural network that can recognize photos of pugs vs. photos of golden retrievers with transfer learning. That is, we take a pre-trained deep convolutional network and retrain the last layer for our particular pug-recognition task. Training can take place using docker on a CPU or on a GPU for speed.

Once the model is built and the weights are saved, we deploy a simple web app. We serve model scores with another container running a simple flask API wrapper around the neural network model. And we build a frontend using a container running R and Shiny. Finally, both containers are run and linked together using docker-compose.

The data comes from URL's from ImageNet. The /data directory of the project has the URL's as well as code for downloading them and normalizing the images. There's also a gzipped pickle file stored in Git LFS so the user doesn't need to download all of the original images.

Docker Images

• Modeling: python3 + theano + jupyter notebook for the CPU
• Modeling: python3 + theano + jupyter notebook for the GPU
• API: python3 + theano + jupyter notebook + flask
• Frontend: R + shiny

Interactive Notebook and Modeling on the CPU

Work locally, or create an ec2 instance:

docker-machine create --driver amazonec2 --amazonec2-access-key XXXX --amazonec2-secret-key XXXX --amazonec2-root-size 100 --amazonec2-instance-type m3.large awsnotebook

If you want the ec2 instance to be on a private VPN:

docker-machine create --driver amazonec2 --amazonec2-access-key XXXX --amazonec2-secret-key XXXX --amazonec2-root-size 100 --amazonec2-zone b --amazonec2-vpc-id vpc-XXXX --amazonec2-subnet-id subnet-XXXX --amazonec2-instance-type m3.large --amazonec2-private-address-only awsnotebook

Get our software:

git clone https://github.com/mdagost/pug_classifier.git
cd pug_classifier
curl http://pug-classifier.s3.amazonaws.com/cnn_pug_model_architecture.json > api/cnn_pug_model_architecture.json
curl http://pug-classifier.s3.amazonaws.com/cnn_pug_model_weights.h5 > api/cnn_pug_model_weights.h5
curl http://pug-classifier.s3.amazonaws.com/pugs_vs_golden_retrvrs_data.pkl.gz > data/pugs_vs_golden_retrvrs_data.pkl.gz
curl http://pug-classifier.s3.amazonaws.com/vgg16_weights.h5 > model/vgg16_weights.h5

Run the container:

eval $(docker-machine env awsnotebook)
docker run -d -p 8888:8888 -v /home/ubuntu/pug_classifier:/home/jovyan/work mdagost/pug_classifier_notebook

Get the IP of the instance:

docker-machine env awsnotebook

Visit http://{{IP}}:8888/ to use the notebook. Note: if you're using AWS you may have to add an inbound rule to the docker-machine security group opening up port 8888.

Interactive Notebook and Modeling on the GPU

Create a GPU instance:

docker-machine create --driver amazonec2 --amazonec2-access-key XXXX --amazonec2-secret-key XXXX --amazonec2-root-size 100 --amazonec2-instance-type g2.2xlarge --amazonec2-ami ami-76b2a71e awsgpunotebook

If you want the ec2 instance to be on a private VPN:

docker-machine create --driver amazonec2 --amazonec2-access-key XXXX --amazonec2-secret-key XXXX --amazonec2-root-size 100 --amazonec2-zone b --amazonec2-vpc-id vpc-XXXX --amazonec2-subnet-id subnet-XXXX --amazonec2-instance-type g2.2xlarge --amazonec2-private-address-only --amazonec2-ami ami-76b2a71e awsgpunotebook

SSH in:

docker-machine ssh awsgpunotebook

Set up the GPU following the instructions here.

Install nvidia-docker like so:

git clone https://github.com/NVIDIA/nvidia-docker
cd nvidia-docker
sudo make install
sudo nvidia-docker volume setup

Get our software:

git clone https://github.com/mdagost/pug_classifier.git
cd pug_classifier
curl http://pug-classifier.s3.amazonaws.com/cnn_pug_model_architecture.json > api/cnn_pug_model_architecture.json
curl http://pug-classifier.s3.amazonaws.com/cnn_pug_model_weights.h5 > api/cnn_pug_model_weights.h5
curl http://pug-classifier.s3.amazonaws.com/pugs_vs_golden_retrvrs_data.pkl.gz > data/pugs_vs_golden_retrvrs_data.pkl.gz
curl http://pug-classifier.s3.amazonaws.com/vgg16_weights.h5 > model/vgg16_weights.h5

Run the container:

sudo nvidia-docker run -d -p 8888:8888 -v /home/ubuntu/pug_classifier:/home/ubuntu mdagost/pug_classifier_gpu_notebook

Get the IP of the instance:

docker-machine env awsnotebook

Visit http://{{IP}}:8888/ to use the notebook. Note: if you're using AWS you may have to add an inbound rule to the docker-machine security group opening up port 8888.

Shiny App Hitting the Flask API

cd shiny/
docker-compose up

Get the IP of the docker VM:

docker-machine env default

Visit http://{{IP}}:3838/pugs/ and voila!

To run the app on AWS:

docker-machine create --driver amazonec2 --amazonec2-access-key XXXX --amazonec2-secret-key XXXX --amazonec2-root-size 100 --amazonec2-instance-type m3.large awsapp
eval $(docker-machine env awsapp)
cd shiny
docker-compose up

Get the IP of the AWS instance:

docker-machine env awsapp

Visit http://{{IP}}:3838/pugs/ and voila! Note: if you're using AWS you may have to add an inbound rule to the docker-machine security group opening up port 3838.