Serverless Machine Learning
JeffConf - Hamburg - 2018
Part 1
Using an existing Machine Learning service
How to use a sentiment analysis API to find out which of your emails should be answered first
Cloud Natural Language: Google's ML API for Speech
Using Ultradox Serverless Workflows
Handles all authentication and API issues
Using the generated app
Calling the API from Ultradox Serverless Workflow
More APIs: E.g. Face Detection
Part 2
Train and Deploy a TensorFlow model
Predict the category of a new customer running on Google Cloud ML
Customer Data - Car Insurance
How would you rank me (47) for a car with 160 km/h top speed?
Create TensorFlow / Keras model by applying alchemy
inputs = Input(name='input', shape=(2, ))
x = Dense(100, name='hidden1', activation='relu')(inputs)
x = Dense(100, name='hidden2', activation='relu')(x)
x = Dense(100, name='hidden3', activation='relu')(x)
predictions = Dense(3, name='softmax', activation='softmax')(x)
Buzzword alarm: Deep Learning
Training Results
77% accuracy on test data (pretty good)
Deployment
# https://cloud.google.com/ml-engine/docs/deploying-models
# Copy model to bucket
gsutil cp -R tf/1 gs://booster_bucket
# https://console.cloud.google.com/mlengine: create model and version
Usage
{"inputs": [ 160, 47, 10]}
gcloud ml-engine predict --model=booster --version=v1
--json-instances=./sample_insurance.json
# SCORES, likelihood of category red, green, yellow
# [0.003163766348734498, 0.9321494698524475, 0.06468681246042252]
Try it yourself in an Azure Notebook
https://notebooks.azure.com/djcordhose/libraries/scipy2018-viz/html/5-keras-tensorflow-nn.ipynb
Wrapup
- Prebuilt ML Services are at your finger tips
- If standard services are not sufficient, train your custom model with free software (e.g. TensorFlow)
- Models can be deployed to managed clouds (Google Cloud ML, Microsoft Azure, AWS) with ease
Ping me for questions / help / comments: http://zeigermann.eu / @DJCordhose / floreysoft