AutoML

AutoML package implemented in Julia using Knet.

Getting Started

First, you have to install AutoML package. Run julia from the command line, set the environment for the package and install AutoML package using the repository address.

(v1.1) pkg> activate env
(env) pkg> add https://github.com/alperkesen/AutoML.jl

How to use

AutoML package provides a trained model using a few lines of code. In order to initialize a model, names and types of input features and output features should be given to constructor. Then, we can train our model using path to csv file and initialized model.

Home Rentals Example

Initializing the model

Home rentals dataset has seven input features:

• neighborhood
• number of bathrooms
• location
• days on market
• initial price
• number of rooms
• sqft

It has one output column:

• rental price

Lists containing tuples of feature names and types should be defined for input and output columns. Using those two lists, we can initialize our model before training it.

julia> import AutoML
julia> hrinputs = [("neighborhood", AutoML.CATEGORY),
                   ("number_of_bathrooms", AutoML.INT),
                   ("location", AutoML.CATEGORY),
                   ("days_on_market", AutoML.INT),
                   ("initial_price", AutoML.FLOAT),
                   ("number_of_rooms", AutoML.INT),
                   ("sqft", AutoML.FLOAT)]
julia> hroutputs = [("rental_price", AutoML.FLOAT)]
julia> model = AutoML.Model(hrinputs, hroutputs; name="homerentals")

Training the model

After initializing the model, we can use give our model and path of csv file to train it. Default number of epochs is 1, an arbitrary number can be given as a parameter to epochs. train method returns trained model and data.

julia> model, dtrn = AutoML.train(model, "data/home_rentals/home_rentals.csv"; epochs=50)

Calculating accuracy/loss

We can evaluate the performance of our model using accuracy/loss.

julia> err = model.model(dtrn.x, dtrn.y)
julia> println("Train error: $err")

julia> acc = AutoML.accuracy(model.model, dtrn)
julia> println("Train accuracy: $acc")

Predicting data

Trained model can be used to predict output value of an example. Example is defined as a dictionary where key indicates the feature name and value indicates the value of that column. Finally, we can use predictdata method to predict the output.

julia> example = Dict("neighborhood" => "downtown",
                      "number_of_bathrooms" => 1,
                      "location" => "good",
                      "days_on_market" => 13,
                      "initial_price" => 1378,
                      "number_of_rooms" => 0,
                      "sqft" => 113.0)
julia> AutoML.predictdata(model, example)
1×1 Array{Float64,2}:
 1347.9081253169932

Default of Credits Example

From command line, you can train a model with one line of code:

$ julia src/AutoML.jl --config configs/doc.yaml --train data/default_of_credit/train.csv --epochs 1 --model_name doc

After training, it prints the accuracy and error in training dataset. It saves model in the models folder in the library. So, we can use it to predict the outputs in different datasets.

$ julia src/AutoML.jl --predict data/default_of_credit/test.csv --model_path models/doc.jld2

Finally, it prints the accuracy and error in the test dataset.