# Adding Models for General Use

Models implementing the MLJ model interface according to the instructions given here should import MLJModelInterface version 0.3 or higher. This is enforced with a statement such as `MLJModelInterface = "^0.3"`

under `[compat]`

in the Project.toml file of the package containing the implementation.

This guide outlines the specification of the MLJ model interface and provides detailed guidelines for implementing the interface for models intended for general use. See also the more condensed Quick-Start Guide to Adding Models.

For sample implementations, see MLJModels/src.

The machine learning tools provided by MLJ can be applied to the models in any package that imports the package MLJModelInterface and implements the API defined there, as outlined below. For a quick-and-dirty implementation of user-defined models see Simple User Defined Models. To make new models available to all MLJ users, see Where to place code implementing new models.

#### Important

MLJModelInterface is a very light-weight interface allowing you to *define* your interface, but does not provide the functionality required to use or test your interface; this requires MLJBase. So, while you only need to add `MLJModelInterface`

to your project's [deps], for testing purposes you need to add MLJBase to your project's [extras] and [targets]. In testing, simply use `MLJBase`

in place of `MLJModelInterface`

.

It is assumed the reader has read Getting Started. To implement the API described here, some familiarity with the following packages is also helpful:

MLJScientificTypes.jl (for specifying model requirements of data)

Distributions.jl (for probabilistic predictions)

CategoricalArrays.jl (essential if you are implementing a model handling data of

`Multiclass`

or`OrderedFactor`

scitype; familiarity with`CategoricalPool`

objects required)Tables.jl (if your algorithm needs input data in a novel format).

In MLJ, the basic interface exposed to the user, built atop the model interface described here, is the *machine interface*. After a first reading of this document, the reader may wish to refer to MLJ Internals for context.

## Overview

A *model* is an object storing hyperparameters associated with some machine learning algorithm. In MLJ, hyperparameters include configuration parameters, like the number of threads, and special instructions, such as "compute feature rankings", which may or may not affect the final learning outcome. However, the logging level (`verbosity`

below) is excluded.

The name of the Julia type associated with a model indicates the associated algorithm (e.g., `DecisionTreeClassifier`

). The outcome of training a learning algorithm is called a *fitresult*. For ordinary multivariate regression, for example, this would be the coefficients and intercept. For a general supervised model, it is the (generally minimal) information needed to make new predictions.

The ultimate supertype of all models is `MLJModelInterface.Model`

, which has two abstract subtypes:

```
abstract type Supervised <: Model end
abstract type Unsupervised <: Model end
```

`Supervised`

models are further divided according to whether they are able to furnish probabilistic predictions of the target (which they will then do by default) or directly predict "point" estimates, for each new input pattern:

```
abstract type Probabilistic <: Supervised end
abstract type Deterministic <: Supervised end
```

Further division of model types is realized through Trait declarations.

Associated with every concrete subtype of `Model`

there must be a `fit`

method, which implements the associated algorithm to produce the fitresult. Additionally, every `Supervised`

model has a `predict`

method, while `Unsupervised`

models must have a `transform`

method. More generally, methods such as these, that are dispatched on a model instance and a fitresult (plus other data), are called *operations*. `Probabilistic`

supervised models optionally implement a `predict_mode`

operation (in the case of classifiers) or a `predict_mean`

and/or `predict_median`

operations (in the case of regressors) although MLJModelInterface also provides fallbacks that will suffice in most cases. `Unsupervised`

models may implement an `inverse_transform`

operation.

## New model type declarations and optional clean! method

Here is an example of a concrete supervised model type declaration:

```
import MLJModelInterface
const MMI = MLJModelInterface
mutable struct RidgeRegressor <: MMI.Deterministic
lambda::Float64
end
```

Models (which are mutable) should not be given internal constructors. It is recommended that they be given an external lazy keyword constructor of the same name. This constructor defines default values for every field, and optionally corrects invalid field values by calling a `clean!`

method (whose fallback returns an empty message string):

```
function MMI.clean!(model::RidgeRegressor)
warning = ""
if model.lambda < 0
warning *= "Need lambda ≥ 0. Resetting lambda=0. "
model.lambda = 0
end
return warning
end
# keyword constructor
function RidgeRegressor(; lambda=0.0)
model = RidgeRegressor(lambda)
message = MMI.clean!(model)
isempty(message) || @warn message
return model
end
```

*Important.* The clean method must have the property that `clean!(clean!(model)) == clean!(model)`

for any instance `model`

.

Although not essential, try to avoid `Union`

types for model fields. For example, a field declaration `features::Vector{Symbol}`

with a default of `Symbol[]`

(detected with `isempty`

method) is preferred to `features::Union{Vector{Symbol}, Nothing}`

with a default of `nothing`

.

An alternative to declaring the model struct, clean! method and keyword constructor, is to use the `@mlj_model`

macro, as in the following example:

```
@mlj_model mutable struct YourModel <: MMI.Deterministic
a::Float64 = 0.5::(_ > 0)
b::String = "svd"::(_ in ("svd","qr"))
end
```

This declaration specifies:

- A keyword constructor (here
`YourModel(; a=..., b=...)`

), - Default values for the hyperparameters,
- Constraints on the hyperparameters where
`_`

refers to a value passed.

For example, `a::Float64 = 0.5::(_ > 0)`

indicates that the field `a`

is a `Float64`

, takes `0.5`

as default value, and expects its value to be positive.

You cannot use the `@mlj_model`

macro if your model struct has type parameters.

## Supervised models

The compulsory and optional methods to be implemented for each concrete type `SomeSupervisedModel <: MMI.Supervised`

are summarized below. An `=`

indicates the return value for a fallback version of the method.

### Summary of methods

Compulsory:

```
MMI.fit(model::SomeSupervisedModel, verbosity::Integer, X, y) -> fitresult, cache, report
MMI.predict(model::SomeSupervisedModel, fitresult, Xnew) -> yhat
```

Optional, to check and correct invalid hyperparameter values:

`MMI.clean!(model::SomeSupervisedModel) = ""`

Optional, to return user-friendly form of fitted parameters:

`MMI.fitted_params(model::SomeSupervisedModel, fitresult) = fitresult`

Optional, to avoid redundant calculations when re-fitting machines associated with a model:

```
MMI.update(model::SomeSupervisedModel, verbosity, old_fitresult, old_cache, X, y) =
MMI.fit(model, verbosity, X, y)
```

Optional, to specify default hyperparameter ranges (for use in tuning):

`MMI.hyperparameter_ranges(T::Type) = Tuple(fill(nothing, length(fieldnames(T))))`

Optional, if `SomeSupervisedModel <: Probabilistic`

:

```
MMI.predict_mode(model::SomeSupervisedModel, fitresult, Xnew) =
mode.(predict(model, fitresult, Xnew))
MMI.predict_mean(model::SomeSupervisedModel, fitresult, Xnew) =
mean.(predict(model, fitresult, Xnew))
MMI.predict_median(model::SomeSupervisedModel, fitresult, Xnew) =
median.(predict(model, fitresult, Xnew))
```

Required, if the model is to be registered (findable by general users):

```
MMI.load_path(::Type{<:SomeSupervisedModel}) = ""
MMI.package_name(::Type{<:SomeSupervisedModel}) = "Unknown"
MMI.package_uuid(::Type{<:SomeSupervisedModel}) = "Unknown"
```

`MMI.input_scitype(::Type{<:SomeSupervisedModel}) = Unknown`

Strongly recommended, to constrain the form of target data passed to fit:

`MMI.target_scitype(::Type{<:SomeSupervisedModel}) = Unknown`

Optional but recommended:

```
MMI.package_url(::Type{<:SomeSupervisedModel}) = "unknown"
MMI.is_pure_julia(::Type{<:SomeSupervisedModel}) = false
MMI.package_license(::Type{<:SomeSupervisedModel}) = "unknown"
```

If `SomeSupervisedModel`

supports sample weights, then instead of the `fit`

above, one implements

`MMI.fit(model::SomeSupervisedModel, verbosity::Integer, X, y, w=nothing) -> fitresult, cache, report`

and, if appropriate

```
MMI.update(model::SomeSupervisedModel, verbosity, old_fitresult, old_cache, X, y, w=nothing) =
MMI.fit(model, verbosity, X, y, w)
```

Additionally, if `SomeSupervisedModel`

supports sample weights, one must declare

`MMI.supports_weights(model::Type{<:SomeSupervisedModel}) = true`

### The form of data for fitting and predicting

The model implementer does not have absolute control over the types of data `X`

, `y`

and `Xnew`

appearing in the `fit`

and `predict`

methods they must implement. Rather, they can specify the *scientific type* of this data by making appropriate declarations of the traits `input_scitype`

and `target_scitype`

discussed later under Trait declarations.

*Important Note.* Unless it genuinely makes little sense to do so, the MLJ recommendation is to specify a `Table`

scientific type for `X`

(and hence `Xnew`

) and an `AbstractVector`

scientific type (e.g., `AbstractVector{Continuous}`

) for targets `y`

. Algorithms requiring matrix input can coerce their inputs appropriately; see below.

#### Additional type coercions

If the core algorithm being wrapped requires data in a different or more specific form, then `fit`

will need to coerce the table into the form desired (and the same coercions applied to `X`

will have to be repeated for `Xnew`

in `predict`

). To assist with common cases, MLJ provides the convenience method `MMI.matrix`

. `MMI.matrix(Xtable)`

has type `Matrix{T}`

where `T`

is the tightest common type of elements of `Xtable`

, and `Xtable`

is any table.

Other auxiliary methods provided by MLJModelInterface for handling tabular data are: `selectrows`

, `selectcols`

, `select`

and `schema`

(for extracting the size, names and eltypes of a table's columns). See Convenience methods below for details.

#### Important convention

It is to be understood that the columns of the table `X`

correspond to features and the rows to observations. So, for example, the predict method for a linear regression model might look like `predict(model, w, Xnew) = MMI.matrix(Xnew)*w`

, where `w`

is the vector of learned coefficients.

### The fit method

A compulsory `fit`

method returns three objects:

`MMI.fit(model::SomeSupervisedModel, verbosity::Int, X, y) -> fitresult, cache, report`

*Note.* The `Int`

typing of `verbosity`

cannot be omitted.

`fitresult`

is the fitresult in the sense above (which becomes an argument for`predict`

discussed below).`report`

is a (possibly empty)`NamedTuple`

, for example,`report=(deviance=..., dof_residual=..., stderror=..., vcov=...)`

. Any training-related statistics, such as internal estimates of the generalization error, and feature rankings, should be returned in the`report`

tuple. How, or if, these are generated should be controlled by hyperparameters (the fields of`model`

). Fitted parameters, such as the coefficients of a linear model, do not go in the report as they will be extractable from`fitresult`

(and accessible to MLJ through the`fitted_params`

method described below).

3. The value of `cache`

can be `nothing`

, unless one is also defining an `update`

method (see below). The Julia type of `cache`

is not presently restricted.

It is not necessary for `fit`

to provide type or dimension checks on `X`

or `y`

or to call `clean!`

on the model; MLJ will carry out such checks.

The method `fit`

should never alter hyperparameter values, the sole exception being fields of type `<:AbstractRNG`

. If the package is able to suggest better hyperparameters, as a byproduct of training, return these in the report field.

The `verbosity`

level (0 for silent) is for passing to learning algorithm itself. A `fit`

method wrapping such an algorithm should generally avoid doing any of its own logging.

*Sample weight support.* If `supports_weights(::Type{<:SomeSupervisedModel})`

has been declared `true`

, then one instead implements the following variation on the above `fit`

:

`MMI.fit(model::SomeSupervisedModel, verbosity::Int, X, y, w=nothing) -> fitresult, cache, report`

### The fitted_params method

A `fitted_params`

method may be optionally overloaded. It's purpose is to provide MLJ access to a user-friendly representation of the learned parameters of the model (as opposed to the hyperparameters). They must be extractable from `fitresult`

.

`MMI.fitted_params(model::SomeSupervisedModel, fitresult) -> friendly_fitresult::NamedTuple`

For a linear model, for example, one might declare something like `friendly_fitresult=(coefs=[...], bias=...)`

.

The fallback is to return `(fitresult=fitresult,)`

.

### The predict method

A compulsory `predict`

method has the form

`MMI.predict(model::SomeSupervisedModel, fitresult, Xnew) -> yhat`

Here `Xnew`

will have the same form as the `X`

passed to `fit`

.

#### Prediction types for deterministic responses.

In the case of `Deterministic`

models, `yhat`

should have the same scitype as the `y`

passed to `fit`

(see above). Any `CategoricalValue`

or `CategoricalString`

elements of `yhat`

**must have a pool == to the pool of the target y presented in training**, even if not all levels appear in the training data or prediction itself. For example, in the case of a univariate target, such as

`scitype(y) <: AbstractVector{Multiclass{3}}`

, one requires `MLJ.classes(yhat[i]) == MLJ.classes(y[j])`

for all admissible `i`

and `j`

. (The method `classes`

is described under Convenience methods below).Unfortunately, code not written with the preservation of categorical levels in mind poses special problems. To help with this, MLJModelInterface provides three utility methods: `int`

(for converting a `CategoricalValue`

or `CategoricalString`

into an integer, the ordering of these integers being consistent with that of the pool), `decoder`

(for constructing a callable object that decodes the integers back into `CategoricalValue`

/`CategoricalString`

objects), and `classes`

, for extracting all the `CategoricalValue`

or `CategoricalString`

objects sharing the pool of a particular value. Refer to Convenience methods below for important details.

Note that a decoder created during `fit`

may need to be bundled with `fitresult`

to make it available to `predict`

during re-encoding. So, for example, if the core algorithm being wrapped by `fit`

expects a nominal target `yint`

of type `Vector{<:Integer}`

then a `fit`

method may look something like this:

```
function MMI.fit(model::SomeSupervisedModel, verbosity, X, y)
yint = MMI.int(y)
a_target_element = y[1] # a CategoricalValue/String
decode = MMI.decoder(a_target_element) # can be called on integers
core_fitresult = SomePackage.fit(X, yint, verbosity=verbosity)
fitresult = (decode, core_fitresult)
cache = nothing
report = nothing
return fitresult, cache, report
end
```

while a corresponding deterministic `predict`

operation might look like this:

```
function MMI.predict(model::SomeSupervisedModel, fitresult, Xnew)
decode, core_fitresult = fitresult
yhat = SomePackage.predict(core_fitresult, Xnew)
return decode.(yhat) # or decode(yhat) also works
end
```

For a concrete example, refer to the code for `SVMClassifier`

.

Of course, if you are coding a learning algorithm from scratch, rather than wrapping an existing one, these extra measures may be unnecessary.

#### Prediction types for probabilistic responses

In the case of `Probabilistic`

models with univariate targets, `yhat`

must be an `AbstractVector`

whose elements are distributions (one distribution per row of `Xnew`

).

Presently, a *distribution* is any object `d`

for which `MMI.isdistribution(::d) = true`

, which is the case for objects of type `Distributions.Sampleable`

.

Use the distribution `MMI.UnivariateFinite`

for `Probabilistic`

models predicting a target with `Finite`

scitype (classifiers). In this case the eltype of the training target `y`

will be a `CategoricalValue`

.

For efficiency, one should not construct `UnivariateDistribution`

instances one at a time. Rather, once a probability vector or matrix is known, construct an instance of `UnivariateFiniteVector <: AbstractArray{<:UnivariateFinite},1}`

to return. Both `UnivariateFinite`

and `UnivariateFiniteVector`

objects are constructed using the single `UnivariateFinite`

function.

For example, suppose the target `y`

arrives as a subsample of some `ybig`

and is missing some classes:

```
ybig = categorical([:a, :b, :a, :a, :b, :a, :rare, :a, :b])
y = ybig[1:6]
```

Your fit method has bundled the first element of `y`

with the `fitresult`

to make it available to `predict`

for purposes of tracking the complete pool of classes. Let's call this `an_element = y[1]`

. Then, supposing the corresponding probabilities of the observed classes `[:a, :b]`

are in an `n x 2`

matrix `probs`

(where `n`

the number of rows of `Xnew`

) then you return

`yhat = UnivariateFinite([:a, :b], probs, pool=an_element)`

This object automatically assigns zero-probability to the unseen class `:rare`

(i.e., `pdf.(yhat, :rare)`

works and returns a zero vector). If you would like to assign `:rare`

non-zero probabilities, simply add it to the first vector (the *support*) and supply a larger `probs`

matrix.

If instead of raw labels `[:a, :b]`

you have the corresponding `CategoricalElement`

s (from, e.g., `filter(cv->cv in unique(y), classes(y))`

) then you can use these instead and drop the `pool`

specifier.

In a binary classification problem it suffices to specify a single vector of probabilities, provided you specify `augment=true`

, as in the following example, *and note carefully that these probablities are associated with the* **last** *(second) class you specify in the constructor:*

```
y = categorical([:TRUE, :FALSE, :FALSE, :TRUE, :TRUE])
an_element = y[1]
probs = rand(10)
yhat = UnivariateFinite([:FALSE, :TRUE], probs, augment=true, pool=an_element)
```

The constructor has a lot of options, including passing a dictionary instead of vectors. See `UnivariateFinite`

for details.

See LinearBinaryClassifier for an example of a Probabilistic classifier implementation.

*Important note on binary classifiers.* There is no "Binary" scitype distinct from `Multiclass{2}`

or `OrderedFactor{2}`

; `Binary`

is just an alias for `Union{Multiclass{2},OrderedFactor{2}}`

. The `target_scitype`

of a binary classifier will generally be `AbstractVector{<:Binary}`

and according to the *mlj* scitype convention, elements of `y`

have type `CategoricalValue`

, and *not* `Bool`

. See BinaryClassifier for an example.

### Trait declarations

Two trait functions allow the implementer to restrict the types of data `X`

, `y`

and `Xnew`

discussed above. The MLJ task interface uses these traits for data type checks but also for model search. If they are omitted (and your model is registered) then a general user may attempt to use your model with inappropriately typed data.

The trait functions `input_scitype`

and `target_scitype`

take scientific data types as values. We assume here familiarity with MLJScientificTypes.jl (see Getting Started for the basics).

For example, to ensure that the `X`

presented to the `DecisionTreeClassifier`

`fit`

method is a table whose columns all have `Continuous`

element type (and hence `AbstractFloat`

machine type), one declares

`MMI.input_scitype(::Type{<:DecisionTreeClassifier}) = MMI.Table(MMI.Continuous)`

or, equivalently,

`MMI.input_scitype(::Type{<:DecisionTreeClassifier}) = Table(Continuous)`

If, instead, columns were allowed to have either: (i) a mixture of `Continuous`

and `Missing`

values, or (ii) `Count`

(i.e., integer) values, then the declaration would be

`MMI.input_scitype(::Type{<:DecisionTreeClassifier}) = Table(Union{Continuous,Missing},Count)`

Similarly, to ensure the target is an AbstractVector whose elements have `Finite`

scitype (and hence `CategoricalValue`

machine type) we declare

`MMI.target_scitype(::Type{<:DecisionTreeClassifier}) = AbstractVector{<:Finite}`

#### Multivariate targets

The above remarks continue to hold unchanged for the case multivariate targets. For example, if we declare

`target_scitype(SomeSupervisedModel) = Table(Continuous)`

then this constrains the target to be any table whose columns have `Continous`

element scitype (i.e., `AbstractFloat`

), while

`target_scitype(SomeSupervisedModel) = Table(Continuous, Finite{2})`

restricts to tables with continuous or binary (ordered or unordered) columns.

For predicting variable length sequences of, say, binary values (`CategoricalValue`

s) with some common size-two pool) we declare

`target_scitype(SomeSupervisedModel) = AbstractVector{<:NTuple{<:Finite{2}}}`

The trait functions controlling the form of data are summarized as follows:

method | return type | declarable return values | fallback value |
---|---|---|---|

`input_scitype` | `Type` | some scientfic type | `Unknown` |

`target_scitype` | `Type` | some scientific type | `Unknown` |

Additional trait functions tell MLJ's `@load`

macro how to find your model if it is registered, and provide other self-explanatory metadata about the model:

method | return type | declarable return values | fallback value |
---|---|---|---|

`load_path` | `String` | unrestricted | "unknown" |

`package_name` | `String` | unrestricted | "unknown" |

`package_uuid` | `String` | unrestricted | "unknown" |

`package_url` | `String` | unrestricted | "unknown" |

`package_license` | `String` | unrestricted | "unknown" |

`is_pure_julia` | `Bool` | `true` or `false` | `false` |

`supports_weights` | `Bool` | `true` or `false` | `false` |

**New.** A final trait you can optionally implement is the `hyperparamter_ranges`

trait. It declares default `ParamRange`

objects for one or more of your model's hyperparameters. This is for use (in the future) by tuning algorithms (e.g., grid generation). It does not represent the full space of *allowed values*. This information is encoded in your `clean!`

method (or `@mlj_model`

call).

The value returned by `hyperparamter_ranges`

must be a tuple of `ParamRange`

objects (query `?range`

for details) whose length is the number of hyperparameters (fields of your model). Note that varying a hyperparameter over a specified range should not alter any type parameters in your model struct (this never applies to numeric ranges). If it doesn't make sense to provide a range for a parameter, a `nothing`

entry is allowed. The fallback returns a tuple of `nothing`

s.

For example, a three parameter model of the form

```
mutable struct MyModel{D} <: Deterministic
alpha::Float64
beta::Int
distribution::D
end
```

you might declare (order matters):

```
MMI.hyperparameter_ranges(::Type{<:MyModel}) =
(range(Float64, :alpha, lower=0, upper=1, scale=:log),
range(Int, :beta, lower=1, upper=Inf, origin=100, unit=50, scale=:log),
nothing)
```

Here is the complete list of trait function declarations for `DecisionTreeClassifier`

(source):

```
MMI.input_scitype(::Type{<:DecisionTreeClassifier}) = MMI.Table(MMI.Continuous)
MMI.target_scitype(::Type{<:DecisionTreeClassifier}) = AbstractVector{<:MMI.Finite}
MMI.load_path(::Type{<:DecisionTreeClassifier}) = "MLJModels.DecisionTree_.DecisionTreeClassifier"
MMI.package_name(::Type{<:DecisionTreeClassifier}) = "DecisionTree"
MMI.package_uuid(::Type{<:DecisionTreeClassifier}) = "7806a523-6efd-50cb-b5f6-3fa6f1930dbb"
MMI.package_url(::Type{<:DecisionTreeClassifier}) = "https://github.com/bensadeghi/DecisionTree.jl"
MMI.is_pure_julia(::Type{<:DecisionTreeClassifier}) = true
```

Alternatively these traits can also be declared using `MMI.metadata_pkg`

and `MMI.metadata_model`

helper functions as:

```
MMI.metadata_pkg(DecisionTreeClassifier,name="DecisionTree",
uuid="7806a523-6efd-50cb-b5f6-3fa6f1930dbb",
url="https://github.com/bensadeghi/DecisionTree.jl",
julia=true)
MMI.metadata_model(DecisionTreeClassifier,
input=MMI.Table(MMI.Continuous),
target=AbstractVector{<:MMI.Finite},
path="MLJModels.DecisionTree_.DecisionTreeClassifier")
```

*Important.* Do not omit the `path`

specifcation.

`MLJModelInterface.metadata_pkg`

— Function`metadata_pkg(T; args...)`

Helper function to write the metadata for a package providing model `T`

. Use it with broadcasting to define the metadata of the package providing a series of models.

**Keywords**

`name="unknown"`

: package name`uuid="unknown"`

: package uuid`url="unknown"`

: package url`julia=missing`

: whether the package is pure julia`license="unknown"`

: package license`is_wrapper=false`

: whether the package is a wrapper

**Example**

```
metadata_pkg.((KNNRegressor, KNNClassifier),
name="NearestNeighbors",
uuid="b8a86587-4115-5ab1-83bc-aa920d37bbce",
url="https://github.com/KristofferC/NearestNeighbors.jl",
julia=true,
license="MIT",
is_wrapper=false)
```

`MLJModelInterface.metadata_model`

— Function`metadata_model(`T`; args...)`

Helper function to write the metadata for a model `T`

.

**Keywords**

`input=Unknown`

: allowed scientific type of the input data`target=Unknown`

: allowed sc. type of the target (supervised)`output=Unknown`

: allowed sc. type of the transformed data (unsupervised)`weights=false`

: whether the model supports sample weights`descr=""`

: short description of the model`path=""`

: where the model is (usually`PackageName.ModelName`

)

**Example**

```
metadata_model(KNNRegressor,
input=MLJModelInterface.Table(MLJModelInterface.Continuous),
target=AbstractVector{MLJModelInterface.Continuous},
weights=true,
descr="K-Nearest Neighbors classifier: ...",
path="NearestNeighbors.KNNRegressor")
```

You can test all your declarations of traits by calling `MLJBase.info_dict(SomeModel)`

.

### Iterative models and the update! method

An `update`

method may be optionally overloaded to enable a call by MLJ to retrain a model (on the same training data) to avoid repeating computations unnecessarily.

```
MMI.update(model::SomeSupervisedModel, verbosity, old_fitresult, old_cache, X, y) -> fit
result, cache, report
MMI.update(model::SomeSupervisedModel, verbosity, old_fitresult, old_cache, X, y, w=nothing) -> fit
result, cache, report
```

Here the second variation applies if `SomeSupervisedModel`

supports sample weights.

If an MLJ `Machine`

is being `fit!`

and it is not the first time, then `update`

is called instead of `fit`

, unless the machine `fit!`

has been called with a new `rows`

keyword argument. However, `MLJModelInterface`

defines a fallback for `update`

which just calls `fit`

. For context, see MLJ Internals.

Learning networks wrapped as models constitute one use-case (see Composing Models): one would like each component model to be retrained only when hyperparameter changes "upstream" make this necessary. In this case MLJ provides a fallback (specifically, the fallback is for any subtype of `SupervisedNetwork = Union{DeterministicNetwork,ProbabilisticNetwork}`

). A second more generally relevant use-case is iterative models, where calls to increase the number of iterations only restarts the iterative procedure if other hyperparameters have also changed. (A useful method for inspecting model changes in such cases is `MLJModelInterface.is_same_except`

. ) For an example, see the MLJ ensemble code.

A third use-case is to avoid repeating time-consuming preprocessing of `X`

and `y`

required by some models.

In the event that the argument `fitresult`

(returned by a preceding call to `fit`

) is not sufficient for performing an update, the author can arrange for `fit`

to output in its `cache`

return value any additional information required (for example, pre-processed versions of `X`

and `y`

), as this is also passed as an argument to the `update`

method.

### Supervised models with a `transform`

method

A supervised model may optionally implement a `transform`

method, whose signature is the same as `predict`

. In that case the implementation should define a value for the `output_scitype`

trait. A declaration

`output_scitype(::Type{<:SomeSupervisedModel}) = T `

is an assurance that `scitype(transform(model, fitresult, Xnew)) <: T`

always holds, for any `model`

of type `SomeSupervisedModel`

.

A use-case for a `transform`

method for a supervised model is a neural network that learns *feature embeddings* for categorical input features as part of overall training. Such a model becomes a transformer that other supervised models can use to transform the categorical features (instead of applying the higher-dimensional one-hot encoding representations).

## Unsupervised models

Unsupervised models implement the MLJ model interface in a very similar fashion. The main differences are:

The

`fit`

method has only one training argument`X`

, as in`MLJModelInterface.fit(model, verbosity::Int, X)`

. However, it has the same return value`(fitresult, cache, report)`

. An`update`

method (e.g., for iterative models) can be optionally implemented in the same way.A

`transform`

method is compulsory and has the same signature as`predict`

, as in`MLJModelInterface.transform(model, fitresult, Xnew)`

.Instead of defining the

`target_scitype`

trait, one declares an`output_scitype`

trait (see above for the meaning).An

`inverse_transform`

can be optionally implemented. The signature is the same as`transform`

, as in`MLJModelInterface.inverse_transform(model, fitresult, Xout)`

, which:must make sense for any

`Xout`

for which`scitype(Xout) <: output_scitype(SomeSupervisedModel)`

(see below); andmust return an object

`Xin`

satisfying`scitype(Xin) <: input_scitype(SomeSupervisedModel)`

.

A

`predict`

method may be optionally implemented, and has the same signature as for supervised models, as in`MLJModelInterface.predict(model, fitresult, Xnew)`

. A use-case is clustering algorithms that`predict`

labels and`transform`

new input features into a space of lower-dimension. See Transformers that also predict for an example.

## Models that learn a probability distribution

The following API is experimental

Models that learn a probability distribution, or more generally a "sampler" object, should be regarded as `Supervised`

models that fit a distribution to the target `y`

, given a *void* input feature, `X = nothing`

. Here is a working implementation of a model to fit any distribution from the Distributions.jl package to some data `y`

, illustrating the idea (trait declarations omitted):

```
# Implmentation:
mutable struct DistributionFitter{D<:Distributions.Distribution} <: Supervised
distribution::D
end
DistributionFitter(; distribution=Distributions.Normal()) =
DistributionFitter(distribution)
function MLJModelInterface.fit(model::DistributionFitter{D},
verbosity::Int,
::Nothing,
y) where D
fitresult = Distributions.fit(D, y)
report = (params=Distributions.params(fitresult),)
cache = nothing
verbosity > 0 && @info "Fitted a $fitresult"
return fitresult, cache, report
end
MLJModelInterface.predict(model::DistributionFitter,
fitresult,
::Nothing) = fitresult
# Example use:
yhat = randn(100)
mach = machine(DistributionFitter(), nothing, y) |> fit!
yhat = predict(mach, nothing)
@assert yhat isa Distributions.Normal
```

## Convenience methods

`MLJModelInterface.int`

— Functionint(x; type=nothing)

The positional integer of the `CategoricalString`

or `CategoricalValue`

`x`

, in the ordering defined by the pool of `x`

. The type of `int(x)`

is the reference type of `x`

.

Not to be confused with `x.ref`

, which is unchanged by reordering of the pool of `x`

, but has the same type.

```
int(X::CategoricalArray)
int(W::Array{<:CategoricalString})
int(W::Array{<:CategoricalValue})
```

Broadcasted versions of `int`

.

```
julia> v = categorical([:c, :b, :c, :a])
julia> levels(v)
3-element Array{Symbol,1}:
:a
:b
:c
julia> int(v)
4-element Array{UInt32,1}:
0x00000003
0x00000002
0x00000003
0x00000001
```

See also: `decoder`

.

`MLJModelInterface.classes`

— Function`classes(x)`

All the categorical elements with the same pool as `x`

(including `x`

), returned as a list, with an ordering consistent with the pool. Here `x`

has `CategoricalValue`

or `CategoricalString`

type, and `classes(x)`

is a vector of the same eltype. Note that `x in classes(x)`

is always true.

Not to be confused with `levels(x.pool)`

. See the example below.

```
julia> v = categorical([:c, :b, :c, :a])
4-element CategoricalArrays.CategoricalArray{Symbol,1,UInt32}:
:c
:b
:c
:a
julia> levels(v)
3-element Array{Symbol,1}:
:a
:b
:c
julia> x = v[4]
CategoricalArrays.CategoricalValue{Symbol,UInt32} :a
julia> classes(x)
3-element CategoricalArrays.CategoricalArray{Symbol,1,UInt32}:
:a
:b
:c
julia> levels(x.pool)
3-element Array{Symbol,1}:
:a
:b
:c
```

```
classes(d::UnivariateFinite)
classes(d::UnivariateFiniteArray)
```

A list of categorial elements in the common pool of classes used to construct `d`

.

```
v = categorical(["yes", "maybe", "no", "yes"])
d = UnivariateFinite(v[1:2], [0.3, 0.7])
classes(d) # CategoricalArray{String,1,UInt32}["maybe", "no", "yes"]
```

`MLJModelInterface.decoder`

— Function`d = decoder(x)`

A callable object for decoding the integer representation of a `CategoricalString`

or `CategoricalValue`

sharing the same pool as `x`

. (Here `x`

is of one of these two types.) Specifically, one has `d(int(y)) == y`

for all `y in classes(x)`

. One can also call `d`

on integer arrays, in which case `d`

is broadcast over all elements.

```
julia> v = categorical([:c, :b, :c, :a])
julia> int(v)
4-element Array{UInt32,1}:
0x00000003
0x00000002
0x00000003
0x00000001
julia> d = decoder(v[3])
julia> d(int(v)) == v
true
```

*Warning:* It is *not* true that `int(d(u)) == u`

always holds.

`MLJModelInterface.matrix`

— Function`matrix(X; transpose=false)`

If `X <: AbstractMatrix`

, return `X`

or `permutedims(X)`

if `transpose=true`

. If `X`

is a Tables.jl compatible table source, convert `X`

into a `Matrix`

.

`MLJModelInterface.table`

— Function`table(columntable; prototype=nothing)`

Convert a named tuple of vectors or tuples `columntable`

, into a table of the "preferred sink type" of `prototype`

. This is often the type of `prototype`

itself, when `prototype`

is a sink; see the Tables.jl documentation. If `prototype`

is not specified, then a named tuple of vectors is returned.

`table(A::AbstractMatrix; names=nothing, prototype=nothing)`

Wrap an abstract matrix `A`

as a Tables.jl compatible table with the specified column `names`

(a tuple of symbols). If `names`

are not specified, `names=(:x1, :x2, ..., :xn)`

is used, where `n=size(A, 2)`

.

If a `prototype`

is specified, then the matrix is materialized as a table of the preferred sink type of `prototype`

, rather than wrapped. Note that if `prototype`

is *not* specified, then `matrix(table(A))`

is essentially a no-op.

`MLJModelInterface.select`

— Function`select(X, r, c)`

Select element(s) of a table or matrix at row(s) r and column(s) c. An object of the sink type of X (or a matrix) is returned unless c is a single integer or symbol. In that case a vector is returned, unless r is a single integer, in which case a single element is returned.

See also: `selectrows`

, `selectcols`

.

`MLJModelInterface.selectrows`

— Function`selectrows(X, r)`

Select single or multiple rows from a table, abstract vector or matrix `X`

. If `X`

is tabular, the object returned is a table of the preferred sink type of `typeof(X)`

, even if only a single row is selected.

`selectrows(X::AbstractNode, r)`

Returns a `Node`

object `N`

such that `N() = selectrows(X(), r)`

(and `N(rows=s) = selectrows(X(rows=s), r)`

).

`MLJModelInterface.selectcols`

— Function`selectcols(X, c)`

Select single or multiple columns from a matrix or table `X`

. If `c`

is an abstract vector of integers or symbols, then the object returned is a table of the preferred sink type of `typeof(X)`

. If `c`

is a *single* integer or column, then an `AbstractVector`

is returned.

`selectcols(X::AbstractNode, c)`

Returns `Node`

object `N`

such that `N() = selectcols(X(), c)`

.

`MLJBase.UnivariateFinite`

— Type```
UnivariateFinite(support,
probs;
pool=nothing,
augmented=false,
ordered=false)
```

Construct a discrete univariate distribution whose finite support is the elements of the vector `support`

, and whose corresponding probabilities are elements of the vector `probs`

. Alternatively, construct an abstract *array* of `UnivariateFinite`

distributions by choosing `probs`

to be an array of one higher dimension than the array generated.

Unless `pool`

is specified, `support`

should have type `AbstractVector{<:CategoricalValue}`

and all elements are assumed to share the same categorical pool, which may be larger than `support`

.

*Important.* All levels of the common pool have associated probabilities, not just those in the specified `support`

. However, these probabilities are always zero (see example below).

If `probs`

is a matrix, it should have a column for each class in `support`

(or one less, if `augment=true`

). More generally, `probs`

will be an array whose size is of the form `(n1, n2, ..., nk, c)`

, where `c = length(support)`

(or one less, if `augment=true`

) and the constructor then returns an array of size `(n1, n2, ..., nk)`

.

```
using CategoricalArrays
v = categorical([:x, :x, :y, :x, :z])
julia> UnivariateFinite(classes(v), [0.2, 0.3, 0.5])
UnivariateFinite{Multiclass{3}}(x=>0.2, y=>0.3, z=>0.5)
julia> d = UnivariateFinite([v[1], v[end]], [0.1, 0.9])
UnivariateFinite{Multiclass{3}(x=>0.1, z=>0.9)
julia> rand(d, 3)
3-element Array{Any,1}:
CategoricalArrays.CategoricalValue{Symbol,UInt32} :z
CategoricalArrays.CategoricalValue{Symbol,UInt32} :z
CategoricalArrays.CategoricalValue{Symbol,UInt32} :z
julia> levels(d)
3-element Array{Symbol,1}:
:x
:y
:z
julia> pdf(d, :y)
0.0
```

**Specifying a pool**

Alternatively, `support`

may be a list of raw (non-categorical) elements if `pool`

is:

some

`CategoricalArray`

,`CategoricalValue`

or`CategoricalPool`

, such that`support`

is a subset of`levels(pool)`

`missing`

, in which case a new categorical pool is created which has`support`

as its only levels.

In the last case, specify `ordered=true`

if the pool is to be considered ordered.

```
julia> UnivariateFinite([:x, :z], [0.1, 0.9], pool=missing, ordered=true)
UnivariateFinite{OrderedFactor{2}}(x=>0.1, z=>0.9)
julia> d = UnivariateFinite([:x, :z], [0.1, 0.9], pool=v) # v defined above
UnivariateFinite(x=>0.1, z=>0.9) (Multiclass{3} samples)
julia> pdf(d, :y) # allowed as `:y in levels(v)`
0.0
v = categorical([:x, :x, :y, :x, :z, :w])
probs = rand(100, 3)
probs = probs ./ sum(probs, dims=2)
julia> UnivariateFinite([:x, :y, :z], probs, pool=v)
100-element UnivariateFiniteVector{Multiclass{4},Symbol,UInt32,Float64}:
UnivariateFinite{Multiclass{4}}(x=>0.194, y=>0.3, z=>0.505)
UnivariateFinite{Multiclass{4}}(x=>0.727, y=>0.234, z=>0.0391)
UnivariateFinite{Multiclass{4}}(x=>0.674, y=>0.00535, z=>0.321)
⋮
UnivariateFinite{Multiclass{4}}(x=>0.292, y=>0.339, z=>0.369)
```

**Probability augmentation**

Unless `augment=true`

, sums of elements along the last axis (row-sums in the case of a matrix) must be equal to one, and otherwise such an array is created by inserting appropriate elements *ahead* of those provided. This means the provided probabilities are associated with the the classes `c2, c3, ..., cn`

.

`UnivariateFinite(prob_given_class; pool=nothing, ordered=false)`

Construct a discrete univariate distribution whose finite support is the set of keys of the provided dictionary, `prob_given_class`

, and whose values specify the corresponding probabilities.

The type requirements on the keys of the dictionary are the same as the elements of `support`

given above with this exception: if non-categorical elements (raw labels) are used as keys, then `pool=...`

must be specified and cannot be `missing`

.

If the values (probabilities) are arrays instead of scalars, then an abstract array of `UnivariateFinite`

elements is created, with the same size as the array.

### Where to place code implementing new models

Note that different packages can implement models having the same name without causing conflicts, although an MLJ user cannot simultaneously *load* two such models.

There are two options for making a new model implementation available to all MLJ users:

**Native implementations**(preferred option). The implementation code lives in the same package that contains the learning algorithms implementing the interface. In this case, it is sufficient to open an issue at MLJ requesting the package to be registered with MLJ. Registering a package allows the MLJ user to access its models' metadata and to selectively load them.**External implementations**(short-term alternative). The model implementation code is necessarily separate from the package`SomePkg`

defining the learning algorithm being wrapped. In this case, the recommended procedure is to include the implementation code at MLJModels/src via a pull-request, and test code at MLJModels/test. Assuming`SomePkg`

is the only package imported by the implementation code, one needs to: (i) register`SomePkg`

with MLJ as explained above; and (ii) add a corresponding`@require`

line in the PR to MLJModels/src/MLJModels.jl to enable lazy-loading of that package by MLJ (following the pattern of existing additions). If other packages must be imported, add them to the MLJModels project file after checking they are not already there. If it is really necessary, packages can be also added to Project.toml for testing purposes.

Additionally, one needs to ensure that the implementation code defines the `package_name`

and `load_path`

model traits appropriately, so that `MLJ`

's `@load`

macro can find the necessary code (see MLJModels/src for examples). The `@load`

command can only be tested after registration. If changes are made, lodge an new issue at MLJ requesting your changes to be updated.

### How to add models to the MLJ model registry?

The MLJ model registry is located in the MLJModels.jl repository. To add a model, you need to follow these steps

Ensure your model conforms to the interface defined above

Raise an issue at MLJModels.jl and point out where the MLJ-interface implementation is, e.g. by providing a link to the code.

An administrator will then review your implementation and work with you to add the model to the registry