Skip to content

Latest commit

 

History

History
143 lines (111 loc) · 4.31 KB

README.md

File metadata and controls

143 lines (111 loc) · 4.31 KB
Raw

Classification

This examples uses Support-vector Machines (SVM) to do multinomial classification.

Classification is predicting which class a new observation belongs to. The prediction is based on a model trained on a set of observations whose class membership is known.

Multinomial problems have more then 2 possible classes. We use the Iris flower data set that contains measurements for 3 Iris species.

We will create a function that takes a JSON array of measurements as input and returns JSON with the predicted class and class membership probabilities corresponding to the measurements.

You will learn how to pre-load trained model objects to make predictions.

You'll need the prerequisites listed here.

Create a new function using a template

Create a new function called r-iris.

faas-cli new --lang rstats-base-plumber r-iris

Customize the function

Edit the ./r-iris/DESCRIPTION file.

Package: OpenFaaStR
Version: 0.0.1
Imports:
  e1071
Remotes:
SystemRequirements:
VersionedPackages:

Open R and perform model training, then save the trained model into ./r-iris/model.rda (see the file model.rda file):

library(e1071) # library with SVM function
data(iris)     # Iris data set

str(iris) # see the measured variables
'data.frame': 150 obs. of  5 variables:
# $ Sepal.Length: num  5.1 4.9 4.7 4.6 5 5.4 4.6 5 4.4 4.9 ...
# $ Sepal.Width : num  3.5 3 3.2 3.1 3.6 3.9 3.4 3.4 2.9 3.1 ...
# $ Petal.Length: num  1.4 1.4 1.3 1.5 1.4 1.7 1.4 1.5 1.4 1.5 ...
# $ Petal.Width : num  0.2 0.2 0.2 0.2 0.2 0.4 0.3 0.2 0.2 0.1 ...
# $ Species     : Factor w/ 3 levels "setosa","versicolor",..: 1 1 ...
levels(iris$Species) # the 3 Iris species
# [1] "setosa"     "versicolor" "virginica" 

## train model with probability=TRUE
model <- svm(Species ~ ., iris, probability=TRUE)
model # print our model info
#
# Call:
# svm(formula = Species ~ ., data = iris)
#
#
# Parameters:
#    SVM-Type:  C-classification 
#  SVM-Kernel:  radial 
#        cost:  1 
#
# Number of Support Vectors:  51

## save the trained model
saveRDS(model, "./r-iris/model.rda")

Change the ./r-iris/handler.R file. Note: loading libraries is good practice, it makes trouble shooting installation related issues much easier (i.e. when shared objects are not found doe to not building the package against specific libraries). Startup messages can also be useful.

When reading in the rda file, we don't need the directory because the file will be moved into the function's root directory:

library(e1071)
model <- readRDS("model.rda")
#* Iris
#* @serializer unboxedJSON
#* @post /
function(req) {
  x <- as.data.frame(
    jsonlite::fromJSON(paste(req$postBody))
  )
  p <- predict(model, x, probability=TRUE)
  list(
      species=as.character(p),
      probabilities=as.list(drop(attr(p,"probabilities")))
  )
}

Edit the r-iris.yml file as required, see configuration options.

Build, push, deploy the function

faas-cli up is a shorthand for automating faas-cli build, faas-cli push, and faas-cli deploy.

faas-cli up -f r-iris.yml

Testing

Test the Docker image locally after docker run -p 5000:8080 $OPENFAAS_PREFIX/r-iris:

curl http://localhost:5000/ -H \
  "Content-Type: application/json" -d \
  '{"Sepal.Length":5.2,"Sepal.Width":3.4,"Petal.Length":1.5,"Petal.Width":0.2}'

Test the deployed instance:

curl $OPENFAAS_URL/function/r-iris -H \
  "Content-Type: application/json" -d \
  '{"Sepal.Length":5.2,"Sepal.Width":3.4,"Petal.Length":1.5,"Petal.Width":0.2}'

The output should include the predicted species name and the probabilities:

{
  "species": "setosa",
  "probabilities": {
    "setosa": 0.9781,
    "versicolor": 0.0126,
    "virginica": 0.0093
  }
}