# Learning networks 2

Download the notebook, the raw script, or the annotated script for this tutorial (right-click on the link and save).

## Preliminary steps

Let's start as with the previous tutorial:

``````using MLJ
using StableRNGs
import DataFrames

rng = StableRNG(6616) # for reproducibility
x1 = rand(rng, 300)
x2 = rand(rng, 300)
x3 = rand(rng, 300)
y = exp.(x1 - x2 -2x3 + 0.1*rand(rng, 300))
X = DataFrames.DataFrame(x1=x1, x2=x2, x3=x3)

test, train = partition(eachindex(y), 0.8);``````

In this tutorial we will show how to generate a model from a network; there are two approaches:

• using the `@from_network` macro

• writing the model in full

the first approach should usually be the one considered as it's simpler.

Generating a model from a network allows subsequent composition of that network with other tasks and tuning of that network.

## Using the `@from_network` macro

Let's define a simple network

Input layer

``````Xs = source(X)
ys = source(y)``````
``Source @644 ⏎ `AbstractArray{Continuous,1}```

First layer

``````std_model = Standardizer()
stand = machine(std_model, Xs)
W = MLJ.transform(stand, Xs)

box_model = UnivariateBoxCoxTransformer()
box_mach = machine(box_model, ys)
z = MLJ.transform(box_mach, ys)``````
``````Node{Machine{UnivariateBoxCoxTransformer}} @347
args:
1:	Source @644
formula:
transform(
Machine{UnivariateBoxCoxTransformer} @156,
Source @644)``````

Second layer

``````ridge_model = RidgeRegressor(lambda=0.1)
ridge = machine(ridge_model, W, z)
ẑ = predict(ridge, W)``````
``````Node{Machine{RidgeRegressor}} @064
args:
1:	Node{Machine{Standardizer}} @070
formula:
predict(
Machine{RidgeRegressor} @310,
transform(
Machine{Standardizer} @498,
Source @912))``````

Output

``ŷ = inverse_transform(box_mach, ẑ)``
``````Node{Machine{UnivariateBoxCoxTransformer}} @605
args:
1:	Node{Machine{RidgeRegressor}} @064
formula:
inverse_transform(
Machine{UnivariateBoxCoxTransformer} @156,
predict(
Machine{RidgeRegressor} @310,
transform(
Machine{Standardizer} @498,
Source @912)))``````

No fitting has been done thus far, we have just defined a sequence of operations.

As we show next, a learning network needs to be exported to create a new stand-alone model type. Instances of that type can be bound with data in a machine, which can then be evaluated, for example. Somewhat paradoxically, one can wrap a learning network in a certain kind of machine, called a learning network machine, before exporting it, and in fact, the export process actually requires us to do so. Since a composite model type does not yet exist, one constructs the machine using a "surrogate" model, whose name indicates the ultimate model supertype (Deterministic, Probabilistic, Unsupervised or Static). This surrogate model has no fields.

``````surrogate = Deterministic()
mach = machine(surrogate, Xs, ys; predict=ŷ)

fit!(mach)
predict(mach, X[test[1:5], :])``````
``````5-element Array{Float64,1}:
0.22207406272038532
0.1145988140862213
0.5637023411242004
0.6208523052884072
0.36914116568152006``````

To form a model out of that network is easy using the `@from_network` macro.

Having defined a learning network machine, mach, as above, the following code defines a new model subtype WrappedRegressor <: Supervised with a single field regressor

``````@from_network mach begin
mutable struct CompositeModel
regressor=ridge_model
end
end``````

The macro defines a constructor CompositeModel and attributes a name to the different models; the ordering / connection between the nodes is inferred from `ŷ` via the `<= ŷ`.

Note: had the model been probabilistic (e.g. `RidgeClassifier`) you would have needed to add `is_probabilistic=true` at the end.

``````cm = machine(CompositeModel(), X, y)
res = evaluate!(cm, resampling=Holdout(fraction_train=0.8, rng=51),
measure=rms)
round(res.measurement[1], sigdigits=3)``````
``0.0136``

## Defining a model from scratch

An alternative to the `@from_network`, is to fully define a new model with its `fit` method:

``````mutable struct CompositeModel2 <: DeterministicNetwork
std_model::Standardizer
box_model::UnivariateBoxCoxTransformer
ridge_model::RidgeRegressor
end

function MLJ.fit(m::CompositeModel2, verbosity::Int, X, y)
Xs = source(X)
ys = source(y)
W = MLJ.transform(machine(m.std_model, Xs), Xs)
box = machine(m.box_model, ys)
z = MLJ.transform(box, ys)
ẑ = predict(machine(m.ridge_model, W, z), W)
ŷ = inverse_transform(box, ẑ)
mach = machine(Deterministic(), Xs, ys; predict=ŷ)
fit!(mach, verbosity=verbosity - 1)
return mach()
end

mdl = CompositeModel2(Standardizer(), UnivariateBoxCoxTransformer(),
RidgeRegressor(lambda=0.1))
cm = machine(mdl, X, y)
res = evaluate!(cm, resampling=Holdout(fraction_train=0.8), measure=rms)
round(res.measurement[1], sigdigits=3)``````
``0.0212``

Either way you now have a constructor to a model which can be used as a stand-alone object, tuned and composed as you would with any basic model.