Creative Prediction with Neural Networks

Why cloud for ML/AI?

• Not always convenient/cost effective to use big workstation.
• We like small laptops without hot GPUs and processors.
• We might want to move from research to product!
• The internet is cool/fun?

Cloud Models

src: www.comgt.com/lib/sw/deliverymodels/

Pizza as a Service

src: Albert Barron

What do we need?

• GPUs: Nvidia [GTX1080TI, K80, P100, V100] or… “Tensor Processing Units”
• OS: Linux?
• CUDA + CUDnn
• Python
• Python libraries: Tensorflow, Keras, SKLearn, etc.
• Jupyter

On Premises: Workstations

• Workstations (15-50KNOK)
• Pro: fun to play with
• Pro: good for small number of users
• Pro: one-time cost
• Con: not practical for many users
• Con: have to keep setting up eduroam
• Con: I don’t like sharing?

Infrastructure-as-a-Service

• Virtual servers
• Set up server, access via Linux shell
• Amazon Web Services (AWS)
• Google Cloud Platform (GCP)
• DigitalOcean (DO)
• UH Cloud (UiO)

Platform-as-a-Service

• Google Cloud Kubenetes Engine
• Deploy “Containerised” application to servers.
• (Deploy DL to Kubernetes)
• Sigma2 (UiO)

Software-as-a-Service

• Google Colaboratory (👏🏼)
• Kaggle Kernels

Example: Robojam

Example: Robojam

RoboJam is a Keras project, now deployed as a Flask web application.

Starting point: Local + SaaS

• Developed on local shared workstation
• Also worked on Colaboratory
• Tips:
• keep jupyter sessions around with screen
• tunnel jupyter port with ssh -L 8888:localhost:8888
• Could also use Google Cloud VMs with GPUs for short training runs

Starting point: IaaS + PaaS

• Can use VMs and Containers for DL development
• Google’s “Cloud Deep Learning VM Image”
• Comes with jupyterhub running to do development in a browser.
• Expensive for a good machine: K80 GPU 0.45USD/h

SaaS Architecture on Colab

Production: Turning into a web service

• Used flask framework to create a RESTful web API
• Just one endpoint: https://0.0.0.0:5000/api/predict
• Send performance as JSON to that endpoit
• robojam RNN model is conditioned with input, then a continuation is predicted.
• prediction returned as JSON
• “Deploying DL models with Flask”

Production: Deploying to DigitalOcean

• Using cheapest DigitalOcean VM: 1vCPU, 1GB, $5 per month.
• Login, clone git repo, run server in a detached screen.
• Works! Deployed for about a year.
• Predictions take about 1.0s-1.5s, not too bad.
• Problem: what if the app gets popular?

IaaS Architecture on DigitalOcean

Containerising: Docker

• We want to make a “container” that includes Robojam and all necessary libraries to run on any Docker installation.
• We’ll start with the tensorflow docker which includes a development environment for tensorflow.

Containerising: Dockerfile

FROM tensorflow/tensorflow:latest-py3
MAINTAINER Charles Martin "charlepm@ifi.uio.no"

COPY requirements.txt /tmp/
RUN pip install --requirement /tmp/requirements.txt
COPY . /tmp/
WORKDIR /tmp
CMD [ "python", "./serve_tiny_performance_mdrnn.py" ]

Containerising: Building the container

sudo docker build -t robojam:latest .
docker tag robojam:latest charlepm/robojam:latest
docker push charlepm/robojam:latest

Containerising: Running the Container

docker run -d -p 5000:5000 robojam:latest

Containerised Architecture with Docker

Deploying to Kubenetes

• Kubernetes is a system to run docker containers on multiple computers simultaneously.
• Let’s set up a little cluster on Google Cloud Platform and deploy Robojam.
• Need to set up computers through the web interface
• Then use command interface to start Robojam.

Deploying Robojam to Cluster

kubectl run robojam-cluster --image=charlepm/robojam:latest --port 5000
kubectl get pods
kubectl expose deployment robojam-cluster --type=LoadBalancer --port 5000 --target-port 5000
kubectl get service

Micro-Service Architecture with Kubernetes

Conclusion

• ML/AI isn’t just for research, we can make cool applications too!
• Cloud resources very available:
• Not too expensive to use powerful servers for short time (training)
• can do a lot with cheap servers for production
• Try out docker etc, makes life much easier.
• Try out Jupyterhub for development. Might be way of future?

Extra links

• Maximise your GPU Dollars