From 18610299584c6f27af56fc9cabd8043930b3b6c3 Mon Sep 17 00:00:00 2001 From: "Documenter.jl" Date: Sun, 2 Feb 2025 20:20:54 +0000 Subject: [PATCH] build based on e705227 --- dev/.documenter-siteinfo.json | 2 +- dev/api/index.html | 97 ++++++++++++++++++++-------------- dev/index.html | 2 +- dev/objects.inv | Bin 697 -> 707 bytes dev/search_index.js | 2 +- 5 files changed, 59 insertions(+), 44 deletions(-) diff --git a/dev/.documenter-siteinfo.json b/dev/.documenter-siteinfo.json index 6f17d61..806e271 100644 --- a/dev/.documenter-siteinfo.json +++ b/dev/.documenter-siteinfo.json @@ -1 +1 @@ -{"documenter":{"julia_version":"1.11.3","generation_timestamp":"2025-02-02T18:14:07","documenter_version":"1.8.0"}} \ No newline at end of file +{"documenter":{"julia_version":"1.11.3","generation_timestamp":"2025-02-02T20:20:49","documenter_version":"1.8.0"}} \ No newline at end of file diff --git a/dev/api/index.html b/dev/api/index.html index 5e42fcf..1284c6b 100644 --- a/dev/api/index.html +++ b/dev/api/index.html @@ -13,7 +13,7 @@ true julia> getobs(x, [1, 3]) == Dict(:a => [1, 3], :b => x[:b][:, [1, 3]]) -truesource
MLUtils.getobs! — Function
getobs!(buffer, data, idx)

Inplace version of getobs(data, idx). If this method is defined for the type of data, then buffer should be used to store the result, instead of allocating a dedicated object.

Implementing this function is optional. In the case no such method is provided for the type of data, then buffer will be ignored and the result of getobs returned. This could be because the type of data may not lend itself to the concept of copy!. Thus, supporting a custom getobs! is optional and not required.

Custom implementations of getobs! should be consistent with getobs in terms of the output format, that is getobs!(buffer, data, idx) == getobs(data, idx).

See also getobs and numobs.

source
MLUtils.numobs — Function
numobs(data)

Return the total number of observations contained in data.

If data does not have numobs defined, then in the case of Tables.table(data) == true returns the number of rows, otherwise returns length(data).

Authors of custom data containers should implement Base.length for their type instead of numobs. numobs should only be implemented for types where there is a difference between numobs and Base.length (such as multi-dimensional arrays).

getobs supports by default nested combinations of array, tuple, named tuples, and dictionaries.

See also getobs.

Examples

julia> x = (a = [1, 2, 3], b = ones(6, 3)); # named tuples
+true
source
MLUtils.getobs! — Function
getobs!(buffer, data, idx)

Inplace version of getobs(data, idx). If this method is defined for the type of data, then buffer should be used to store the result, instead of allocating a dedicated object.

Implementing this function is optional. In the case no such method is provided for the type of data, then buffer will be ignored and the result of getobs returned. This could be because the type of data may not lend itself to the concept of copy!. Thus, supporting a custom getobs! is optional and not required.

Custom implementations of getobs! should be consistent with getobs in terms of the output format, that is getobs!(buffer, data, idx) == getobs(data, idx).

See also getobs and numobs.

source
MLUtils.numobs — Function
numobs(data)

Return the total number of observations contained in data.

If data does not have numobs defined, then in the case of Tables.table(data) == true returns the number of rows, otherwise returns length(data).

Authors of custom data containers should implement Base.length for their type instead of numobs. numobs should only be implemented for types where there is a difference between numobs and Base.length (such as multi-dimensional arrays).

getobs supports by default nested combinations of array, tuple, named tuples, and dictionaries.

See also getobs.

Examples

julia> x = (a = [1, 2, 3], b = ones(6, 3)); # named tuples
 
 julia> numobs(x)
 3
@@ -33,14 +33,14 @@
  [3] numobs(data::NamedTuple{(:a, :b), Tuple{Vector{Int64}, Matrix{Float64}}})
    @ MLUtils ~/.julia/dev/MLUtils/src/observation.jl:177
  [4] top-level scope
-   @ REPL[35]:1
source

Lazy Transforms

MLUtils.filterobs — Function
filterobs(f, data)

Return a subset of data container data including all indices i for which f(getobs(data, i)) === true.

data = 1:10
+   @ REPL[35]:1
source

Lazy Transforms

MLUtils.filterobs — Function
filterobs(f, data)

Return a subset of data container data including all indices i for which f(getobs(data, i)) === true.

data = 1:10
 numobs(data) == 10
 fdata = filterobs(>(5), data)
-numobs(fdata) == 5
source
MLUtils.groupobs — Function
groupobs(f, data)

Split data container data data into different data containers, grouping observations by f(obs).

data = -10:10
+numobs(fdata) == 5
source
MLUtils.groupobs — Function
groupobs(f, data)

Split data container data data into different data containers, grouping observations by f(obs).

data = -10:10
 datas = groupobs(>(0), data)
-length(datas) == 2
source
MLUtils.joinobs — Function
joinobs(datas...)

Concatenate data containers datas.

data1, data2 = 1:10, 11:20
+length(datas) == 2
source
MLUtils.joinobs — Function
joinobs(datas...)

Concatenate data containers datas.

data1, data2 = 1:10, 11:20
 jdata = joinumobs(data1, data2)
-getobs(jdata, 15) == 15
source
MLUtils.mapobs — Function
mapobs(f, data; batched=:auto)

Lazily map f over the observations in a data container data. Returns a new data container mdata that can be indexed and has a length. Indexing triggers the transformation f.

The batched keyword argument controls the behavior of mdata[idx] and mdata[idxs] where idx is an integer and idxs is a vector of integers:

  • batched=:auto (default). Let f handle the two cases. Calls f(getobs(data, idx)) and f(getobs(data, idxs)).
  • batched=:never. The function f is always called on a single observation. Calls f(getobs(data, idx)) and [f(getobs(data, idx)) for idx in idxs].
  • batched=:always. The function f is always called on a batch of observations. Calls getobs(f(getobs(data, [idx])), 1) and f(getobs(data, idxs)).

Examples

julia> data = (a=[1,2,3], b=[1,2,3]);
+getobs(jdata, 15) == 15
source
MLUtils.mapobs — Function
mapobs(f, data; batched=:auto)

Lazily map f over the observations in a data container data. Returns a new data container mdata that can be indexed and has a length. Indexing triggers the transformation f.

The batched keyword argument controls the behavior of mdata[idx] and mdata[idxs] where idx is an integer and idxs is a vector of integers:

  • batched=:auto (default). Let f handle the two cases. Calls f(getobs(data, idx)) and f(getobs(data, idxs)).
  • batched=:never. The function f is always called on a single observation. Calls f(getobs(data, idx)) and [f(getobs(data, idx)) for idx in idxs].
  • batched=:always. The function f is always called on a batch of observations. Calls getobs(f(getobs(data, [idx])), 1) and f(getobs(data, idxs)).

Examples

julia> data = (a=[1,2,3], b=[1,2,3]);
 
 julia> mdata = mapobs(data) do x
          (c = x.a .+ x.b,  d = x.a .- x.b)
@@ -51,10 +51,10 @@
 (c = 2, d = 0)
 
 julia> mdata[1:2]
-(c = [2, 4], d = [0, 0])
source
mapobs(fs, data)

Lazily map each function in tuple fs over the observations in data container data. Returns a tuple of transformed data containers.

source
mapobs(namedfs::NamedTuple, data)

Map a NamedTuple of functions over data, turning it into a data container of NamedTuples. Field syntax can be used to select a column of the resulting data container.

data = 1:10
+(c = [2, 4], d = [0, 0])
source
mapobs(fs, data)

Lazily map each function in tuple fs over the observations in data container data. Returns a tuple of transformed data containers.

source
mapobs(namedfs::NamedTuple, data)

Map a NamedTuple of functions over data, turning it into a data container of NamedTuples. Field syntax can be used to select a column of the resulting data container.

data = 1:10
 nameddata = mapobs((x = sqrt, y = log), data)
 getobs(nameddata, 10) == (x = sqrt(10), y = log(10))
-getobs(nameddata.x, 10) == sqrt(10)
source
mapobs(f, d::DataLoader)

Return a new dataloader based on d that applies f at each iteration.

Examples

julia> X = ones(3, 6);
+getobs(nameddata.x, 10) == sqrt(10)
source
mapobs(f, d::DataLoader)

Return a new dataloader based on d that applies f at each iteration.

Examples

julia> X = ones(3, 6);
 
 julia> function f(x)
            @show x
@@ -83,13 +83,13 @@
        end
 x = [1.0 1.0; 1.0 1.0; 1.0 1.0]
 x = [1.0 1.0; 1.0 1.0; 1.0 1.0]
-x = [1.0 1.0; 1.0 1.0; 1.0 1.0]
source
MLUtils.shuffleobs — Function
shuffleobs([rng], data)

Return a version of the dataset data that contains all the origin observations in a random reordering.

The values of data itself are not copied. Instead only the indices are shuffled. This function calls obsview to accomplish that, which means that the return value is likely of a different type than data.

Optionally, a random number generator rng can be passed as the first argument.

For this function to work, the type of data must implement numobs and getobs.

See also obsview.

Examples

# For Arrays the subset will be of type SubArray
+x = [1.0 1.0; 1.0 1.0; 1.0 1.0]
source
MLUtils.shuffleobs — Function
shuffleobs([rng], data)

Return a version of the dataset data that contains all the origin observations in a random reordering.

The values of data itself are not copied. Instead only the indices are shuffled. This function calls obsview to accomplish that, which means that the return value is likely of a different type than data.

Optionally, a random number generator rng can be passed as the first argument.

For this function to work, the type of data must implement numobs and getobs.

See also obsview.

Examples

# For Arrays the subset will be of type SubArray
 @assert typeof(shuffleobs(rand(4,10))) <: SubArray
 
 # Iterate through all observations in random order
 for x in eachobs(shuffleobs(X))
     ...
-end
source

Batching, Iteration, and Views

MLUtils.batch — Function
batch(xs)

Batch the arrays in xs into a single array with an extra dimension.

If the elements of xs are tuples, named tuples, or dicts, the output will be of the same type.

See also unbatch.

Examples

julia> batch([[1,2,3], 
+end
source

Batching, Iteration, and Views

MLUtils.batch — Function
batch(xs)

Batch the arrays in xs into a single array with an extra dimension.

If the elements of xs are tuples, named tuples, or dicts, the output will be of the same type.

See also unbatch.

Examples

julia> batch([[1,2,3], 
               [4,5,6]])
 3×2 Matrix{Int64}:
  1  4
@@ -98,15 +98,15 @@
 
 julia> batch([(a=[1,2], b=[3,4])
                (a=[5,6], b=[7,8])]) 
-(a = [1 5; 2 6], b = [3 7; 4 8])
source
MLUtils.batchsize — Function
batchsize(data::BatchView) -> Int

Return the fixed size of each batch in data.

Examples

using MLUtils
+(a = [1 5; 2 6], b = [3 7; 4 8])
source
MLUtils.batchsize — Function
batchsize(data::BatchView) -> Int

Return the fixed size of each batch in data.

Examples

using MLUtils
 X, Y = MLUtils.load_iris()
 
 A = BatchView(X, batchsize=30)
-@assert batchsize(A) == 30
source
MLUtils.batchseq — Function
batchseq(seqs, val = 0)

Take a list of N sequences, and turn them into a single sequence where each item is a batch of N. Short sequences will be padded by val.

Examples

julia> batchseq([[1, 2, 3], [4, 5]], 0)
+@assert batchsize(A) == 30
source
MLUtils.batchseq — Function
batchseq(seqs, val = 0)

Take a list of N sequences, and turn them into a single sequence where each item is a batch of N. Short sequences will be padded by val.

Examples

julia> batchseq([[1, 2, 3], [4, 5]], 0)
 3-element Vector{Vector{Int64}}:
  [1, 4]
  [2, 5]
- [3, 0]
source
MLUtils.BatchView — Type
BatchView(data, batchsize; partial=true, collate=nothing)
+ [3, 0]
source
MLUtils.BatchView — Type
BatchView(data, batchsize; partial=true, collate=nothing)
 BatchView(data; batchsize=1, partial=true, collate=nothing)

Create a view of the given data that represents it as a vector of batches. Each batch will contain an equal amount of observations in them. The batch-size can be specified using the parameter batchsize. In the case that the size of the dataset is not dividable by the specified batchsize, the remaining observations will be ignored if partial=false. If partial=true instead the last batch-size can be slightly smaller.

If used as an iterator, the object will iterate over the dataset once, effectively denoting an epoch.

Any data access is delayed until iteration or indexing is perfomed. The getobs function is called on the data object to retrieve the observations.

For BatchView to work on some data structure, the type of the given variable data must implement the data container interface. See ObsView for more info.

Arguments

  • data : The object describing the dataset. Can be of any type as long as it implements getobs and numobs (see Details for more information).

  • batchsize : The batch-size of each batch. It is the number of observations that each batch must contain (except possibly for the last one).

  • partial : If partial=false and the number of observations is not divisible by the batch-size, then the last mini-batch is dropped.

  • collate: Defines the batching behavior.

    • If nothing (default), a batch is getobs(data, indices).
    • If false, each batch is [getobs(data, i) for i in indices].
    • If true, applies MLUtils to the vector of observations in a batch, recursively collating arrays in the last dimensions. See MLUtils.batch for more information and examples.
    • If a custom function, it will be used in place of MLUtils.batch. It should take a vector of observations as input.

Se also DataLoader.

Examples

julia> using MLUtils
 
 julia> X, Y = MLUtils.load_iris();
@@ -145,7 +145,7 @@
 julia> for (x, y) in BatchView((rand(10, 4), ["a", "b", "c", "d"]), batchsize=2, collate=collate_fn)
            @assert size(x) == (10, 2)
            @assert y isa String
-       end
source
MLUtils.eachobs — Function
eachobs(data; kws...)

Return an iterator over data.

Supports the same arguments as DataLoader. The batchsize default is -1 here while it is 1 for DataLoader.

Examples

X = rand(4,100)
+       end
source
MLUtils.eachobs — Function
eachobs(data; kws...)

Return an iterator over data.

Supports the same arguments as DataLoader. The batchsize default is -1 here while it is 1 for DataLoader.

Examples

X = rand(4,100)
 
 for x in eachobs(X)
     # loop entered 100 times
@@ -163,7 +163,7 @@
 # support for tuples, named tuples, dicts
 for (x, y) in eachobs((X, Y))
     # ...
-end
source
MLUtils.DataLoader — Type
DataLoader(data; [batchsize, buffer, collate, parallel, partial, rng, shuffle])

An object that iterates over mini-batches of data, each mini-batch containing batchsize observations (except possibly the last one).

Takes as input a single data array, a tuple (or a named tuple) of arrays, or in general any data object that implements the numobs and getobs methods.

The last dimension in each array is the observation dimension, i.e. the one divided into mini-batches.

The original data is preserved in the data field of the DataLoader.

Arguments

  • data: The data to be iterated over. The data type has to be supported by numobs and getobs.
  • batchsize: If less than 0, iterates over individual observations. Otherwise, each iteration (except possibly the last) yields a mini-batch containing batchsize observations. Default 1.
  • buffer: If buffer=true and supported by the type of data, a buffer will be allocated and reused for memory efficiency. May want to set partial=false to avoid size mismatch. Finally, can pass an external buffer to be used in getobs! (depending on the collate and batchsize options, could be getobs!(buffer, data, idxs) or getobs!(buffer[i], data, idx)). Default false.
  • collate: Defines the batching behavior. Default nothing.
    • If nothing , a batch is getobs(data, indices).
    • If false, each batch is [getobs(data, i) for i in indices].
    • If true, applies MLUtils.batch to the vector of observations in a batch, recursively collating arrays in the last dimensions. See MLUtils.batch for more information and examples.
    • If a custom function, it will be used in place of MLUtils.batch. It should take a vector of observations as input.
  • parallel: Whether to use load data in parallel using worker threads. Greatly speeds up data loading by factor of available threads. Requires starting Julia with multiple threads. Check Threads.nthreads() to see the number of available threads. Passing parallel = true breaks ordering guarantees. Default false.
  • partial: This argument is used only when batchsize > 0. If partial=false and the number of observations is not divisible by the batchsize, then the last mini-batch is dropped. Default true.
  • rng: A random number generator. Default Random.default_rng().
  • `shuffle: Whether to shuffle the observations before iterating. Unlike wrapping the data container with shuffleobs(data), shuffle=true ensures that the observations are shuffled anew every time you start iterating over eachobs. Default false.

Examples

julia> Xtrain = rand(10, 100);
+end
source
MLUtils.DataLoader — Type
DataLoader(data; [batchsize, buffer, collate, parallel, partial, rng, shuffle])

An object that iterates over mini-batches of data, each mini-batch containing batchsize observations (except possibly the last one).

Takes as input a single data array, a tuple (or a named tuple) of arrays, or in general any data object that implements the numobs and getobs methods.

The last dimension in each array is the observation dimension, i.e. the one divided into mini-batches.

The original data is preserved in the data field of the DataLoader.

Arguments

  • data: The data to be iterated over. The data type has to be supported by numobs and getobs.
  • batchsize: If less than 0, iterates over individual observations. Otherwise, each iteration (except possibly the last) yields a mini-batch containing batchsize observations. Default 1.
  • buffer: If buffer=true and supported by the type of data, a buffer will be allocated and reused for memory efficiency. May want to set partial=false to avoid size mismatch. Finally, can pass an external buffer to be used in getobs! (depending on the collate and batchsize options, could be getobs!(buffer, data, idxs) or getobs!(buffer[i], data, idx)). Default false.
  • collate: Defines the batching behavior. Default nothing.
    • If nothing , a batch is getobs(data, indices).
    • If false, each batch is [getobs(data, i) for i in indices].
    • If true, applies MLUtils.batch to the vector of observations in a batch, recursively collating arrays in the last dimensions. See MLUtils.batch for more information and examples.
    • If a custom function, it will be used in place of MLUtils.batch. It should take a vector of observations as input.
  • parallel: Whether to use load data in parallel using worker threads. Greatly speeds up data loading by factor of available threads. Requires starting Julia with multiple threads. Check Threads.nthreads() to see the number of available threads. Passing parallel = true breaks ordering guarantees. Default false.
  • partial: This argument is used only when batchsize > 0. If partial=false and the number of observations is not divisible by the batchsize, then the last mini-batch is dropped. Default true.
  • rng: A random number generator. Default Random.default_rng().
  • `shuffle: Whether to shuffle the observations before iterating. Unlike wrapping the data container with shuffleobs(data), shuffle=true ensures that the observations are shuffled anew every time you start iterating over eachobs. Default false.

Examples

julia> Xtrain = rand(10, 100);
 
 julia> array_loader = DataLoader(Xtrain, batchsize=2);
 
@@ -208,7 +208,7 @@
 julia> collate_fn(batch) = join(batch);
 
 julia> first(DataLoader(["a", "b", "c", "d"], batchsize=2, collate=collate_fn))
-"ab"
source
MLUtils.obsview — Function
obsview(data, [indices])

Returns a lazy view of the observations in data that correspond to the given indices. No data will be copied except of the indices. It is similar to constructing an ObsView, but returns a SubArray if the type of data is Array or SubArray. Furthermore, this function may be extended for custom types of data that also want to provide their own subset-type.

In case data is a tuple, the constructor will be mapped over its elements. That means that the constructor returns a tuple of ObsView instead of a ObsView of tuples.

If instead you want to get the subset of observations corresponding to the given indices in their native type, use getobs.

See ObsView for more information.

source
MLUtils.ObsView — Type
ObsView(data, [indices])

Used to represent a subset of some data of arbitrary type by storing which observation-indices the subset spans. Furthermore, subsequent subsettings are accumulated without needing to access actual data.

The main purpose for the existence of ObsView is to delay data access and movement until an actual batch of data (or single observation) is needed for some computation. This is particularily useful when the data is not located in memory, but on the hard drive or some remote location. In such a scenario one wants to load the required data only when needed.

Any data access is delayed until getindex is called, and even getindex returns the result of obsview which in general avoids data movement until getobs is called. If used as an iterator, the view will iterate over the dataset once, effectively denoting an epoch. Each iteration will return a lazy subset to the current observation.

Arguments

  • data : The object describing the dataset. Can be of any type as long as it implements getobs and numobs (see Details for more information).

  • indices : Optional. The index or indices of the observation(s) in data that the subset should represent. Can be of type Int or some subtype of AbstractVector.

Methods

  • getindex : Returns the observation(s) of the given index/indices. No data is copied aside from the required indices.

  • numobs : Returns the total number observations in the subset.

  • getobs : Returns the underlying data that the ObsView represents at the given relative indices. Note that these indices are in "subset space", and in general will not directly correspond to the same indices in the underlying data set.

Details

For ObsView to work on some data structure, the desired type MyType must implement the following interface:

  • getobs(data::MyType, idx) : Should return the observation(s) indexed by idx. In what form is up to the user. Note that idx can be of type Int or AbstractVector.

  • numobs(data::MyType) : Should return the total number of observations in data

The following methods can also be provided and are optional:

  • getobs(data::MyType) : By default this function is the identity function. If that is not the behaviour that you want for your type, you need to provide this method as well.

  • obsview(data::MyType, idx) : If your custom type has its own kind of subset type, you can return it here. An example for such a case are SubArray for representing a subset of some AbstractArray.

  • getobs!(buffer, data::MyType, [idx]) : Inplace version of getobs(data, idx). If this method is provided for MyType, then eachobs can preallocate a buffer that is then reused every iteration. Note: buffer should be equivalent to the return value of getobs(::MyType, ...), since this is how buffer is preallocated by default.

Examples

X, Y = MLUtils.load_iris()
+"ab"
source
MLUtils.obsview — Function
obsview(data, [indices])

Returns a lazy view of the observations in data that correspond to the given indices. No data will be copied except of the indices. It is similar to constructing an ObsView, but returns a SubArray if the type of data is Array or SubArray. Furthermore, this function may be extended for custom types of data that also want to provide their own subset-type.

In case data is a tuple, the constructor will be mapped over its elements. That means that the constructor returns a tuple of ObsView instead of a ObsView of tuples.

If instead you want to get the subset of observations corresponding to the given indices in their native type, use getobs.

See ObsView for more information.

source
MLUtils.ObsView — Type
ObsView(data, [indices])

Used to represent a subset of some data of arbitrary type by storing which observation-indices the subset spans. Furthermore, subsequent subsettings are accumulated without needing to access actual data.

The main purpose for the existence of ObsView is to delay data access and movement until an actual batch of data (or single observation) is needed for some computation. This is particularily useful when the data is not located in memory, but on the hard drive or some remote location. In such a scenario one wants to load the required data only when needed.

Any data access is delayed until getindex is called, and even getindex returns the result of obsview which in general avoids data movement until getobs is called. If used as an iterator, the view will iterate over the dataset once, effectively denoting an epoch. Each iteration will return a lazy subset to the current observation.

Arguments

  • data : The object describing the dataset. Can be of any type as long as it implements getobs and numobs (see Details for more information).

  • indices : Optional. The index or indices of the observation(s) in data that the subset should represent. Can be of type Int or some subtype of AbstractVector.

Methods

  • getindex : Returns the observation(s) of the given index/indices. No data is copied aside from the required indices.

  • numobs : Returns the total number observations in the subset.

  • getobs : Returns the underlying data that the ObsView represents at the given relative indices. Note that these indices are in "subset space", and in general will not directly correspond to the same indices in the underlying data set.

Details

For ObsView to work on some data structure, the desired type MyType must implement the following interface:

  • getobs(data::MyType, idx) : Should return the observation(s) indexed by idx. In what form is up to the user. Note that idx can be of type Int or AbstractVector.

  • numobs(data::MyType) : Should return the total number of observations in data

The following methods can also be provided and are optional:

  • getobs(data::MyType) : By default this function is the identity function. If that is not the behaviour that you want for your type, you need to provide this method as well.

  • obsview(data::MyType, idx) : If your custom type has its own kind of subset type, you can return it here. An example for such a case are SubArray for representing a subset of some AbstractArray.

  • getobs!(buffer, data::MyType, [idx]) : Inplace version of getobs(data, idx). If this method is provided for MyType, then eachobs can preallocate a buffer that is then reused every iteration. Note: buffer should be equivalent to the return value of getobs(::MyType, ...), since this is how buffer is preallocated by default.

Examples

X, Y = MLUtils.load_iris()
 
 # The iris set has 150 observations and 4 features
 @assert size(X) == (4,150)
@@ -250,7 +250,22 @@
 end
 
 # Indexing: take first 10 observations
-x, y = ObsView((X, Y))[1:10]

See also

obsview, getobs, numobs, splitobs, shuffleobs, kfolds.

source
MLUtils.randobs — Function
randobs(data, [n])

Pick a random observation or a batch of n random observations from data. For this function to work, the type of data must implement numobs and getobs.

source

Partitioning

MLUtils.leavepout — Function
leavepout(n::Integer, [size = 1]) -> Tuple

Compute the train/validation assignments for k ≈ n/size repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. Each validation subset will have either size or size+1 observations assigned to it. The following code snippet generates the index-vectors for size = 2.

julia> train_idx, val_idx = leavepout(10, 2);

Each observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.

julia> train_idx
+x, y = ObsView((X, Y))[1:10]

See also

obsview, getobs, numobs, splitobs, shuffleobs, kfolds.

source
MLUtils.randobs — Function
randobs(data, [n])

Pick a random observation or a batch of n random observations from data. For this function to work, the type of data must implement numobs and getobs.

source
MLUtils.slidingwindow — Function
slidingwindow(data; size, stride=1) -> SlidingWindow

Return a vector-like view of the data for which each element is a fixed size "window" of size adjacent observations. Note that only complete windows are included in the output, which implies that it is possible for excess observations to be omitted from the view.

Note that the windows are not materialized at construction time. To actually get a copy of the data at some window use indexing or getobs.

julia> s = slidingwindow(1:20, size=6)
+slidingwindow(1:20, size=6, stride=1)
+
+julia> s[1]
+1:6
+
+julia> s[2]
+2:7

The optional parameter stride can be used to specify the distance between the start elements of each adjacent window. By default the stride is equal to 1.

julia> s = slidingwindow(1:20, size=6, stride=3)
+slidingwindow(1:20, size=6, stride=3)
+
+julia> for w in s; println(w); end
+1:6
+4:9
+7:12
+10:15
+13:18
source

Partitioning

MLUtils.leavepout — Function
leavepout(n::Integer, [size = 1]) -> Tuple

Compute the train/validation assignments for k ≈ n/size repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. Each validation subset will have either size or size+1 observations assigned to it. The following code snippet generates the index-vectors for size = 2.

julia> train_idx, val_idx = leavepout(10, 2);

Each observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.

julia> train_idx
 5-element Array{Array{Int64,1},1}:
  [3,4,5,6,7,8,9,10]
  [1,2,5,6,7,8,9,10]
@@ -264,11 +279,11 @@
  3:4
  5:6
  7:8
- 9:10
source
leavepout(data, p = 1)

Repartition a data container using a k-fold strategy, where k is chosen in such a way, that each validation subset of the resulting folds contains roughly p observations. Defaults to p = 1, which is also known as "leave-one-out" partitioning.

The resulting sequence of folds is returned as a lazy iterator. Only data subsets are created. That means no actual data is copied until getobs is invoked.

for (train, val) in leavepout(X, p=2)
-    # if nobs(X) is dividable by 2,
+ 9:10
source
leavepout(data, p = 1)

Repartition a data container using a k-fold strategy, where k is chosen in such a way, that each validation subset of the resulting folds contains roughly p observations. Defaults to p = 1, which is also known as "leave-one-out" partitioning.

The resulting sequence of folds is returned as a lazy iterator. Only data subsets are created. That means no actual data is copied until getobs is invoked.

for (train, val) in leavepout(X, p=2)
+    # if numobs(X) is dividable by 2,
     # then numobs(val) will be 2 for each iteraton,
     # otherwise it may be 3 for the first few iterations.
-end

Seekfolds for a related function.

source
MLUtils.kfolds — Function
kfolds(n::Integer, k = 5) -> Tuple

Compute the train/validation assignments for k repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. A general rule of thumb is to use either k = 5 or k = 10. The following code snippet generates the indices assignments for k = 5

julia> train_idx, val_idx = kfolds(10, 5);

Each observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.

julia> train_idx
+end

Seekfolds for a related function.

source
MLUtils.kfolds — Function
kfolds(n::Integer, k = 5) -> Tuple

Compute the train/validation assignments for k repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. A general rule of thumb is to use either k = 5 or k = 10. The following code snippet generates the indices assignments for k = 5

julia> train_idx, val_idx = kfolds(10, 5);

Each observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.

julia> train_idx
 5-element Array{Array{Int64,1},1}:
  [3,4,5,6,7,8,9,10]
  [1,2,5,6,7,8,9,10]
@@ -282,18 +297,18 @@
  3:4
  5:6
  7:8
- 9:10
source
kfolds(data, [k = 5])

Repartition a data container k times using a k folds strategy and return the sequence of folds as a lazy iterator. Only data subsets are created, which means that no actual data is copied until getobs is invoked.

Conceptually, a k-folds repartitioning strategy divides the given data into k roughly equal-sized parts. Each part will serve as validation set once, while the remaining parts are used for training. This results in k different partitions of data.

In the case that the size of the dataset is not dividable by the specified k, the remaining observations will be evenly distributed among the parts.

for (x_train, x_val) in kfolds(X, k=10)
+ 9:10
source
kfolds(data, [k = 5])

Repartition a data container k times using a k folds strategy and return the sequence of folds as a lazy iterator. Only data subsets are created, which means that no actual data is copied until getobs is invoked.

Conceptually, a k-folds repartitioning strategy divides the given data into k roughly equal-sized parts. Each part will serve as validation set once, while the remaining parts are used for training. This results in k different partitions of data.

In the case that the size of the dataset is not dividable by the specified k, the remaining observations will be evenly distributed among the parts.

for (x_train, x_val) in kfolds(X, k=10)
     # code called 10 times
-    # nobs(x_val) may differ up to ±1 over iterations
+    # numobs(x_val) may differ up to ±1 over iterations
 end

Multiple variables are supported (e.g. for labeled data)

for ((x_train, y_train), val) in kfolds((X, Y), k=10)
     # ...
 end

By default the folds are created using static splits. Use shuffleobs to randomly assign observations to the folds.

for (x_train, x_val) in kfolds(shuffleobs(X), k = 10)
     # ...
-end

See leavepout for a related function.

source
MLUtils.splitobs — Function
splitobs(n::Int; at) -> Tuple

Compute the indices for two or more disjoint subsets of the range 1:n with split sizes determined by at.

Examples

julia> splitobs(100, at=0.7)
+end

See leavepout for a related function.

source
MLUtils.splitobs — Function
splitobs(n::Int; at) -> Tuple

Compute the indices for two or more disjoint subsets of the range 1:n with split sizes determined by at.

Examples

julia> splitobs(100, at=0.7)
 (1:70, 71:100)
 
 julia> splitobs(100, at=(0.1, 0.4))
-(1:10, 11:50, 51:100)
source
splitobs([rng,] data; at, shuffle=false, stratified=nothing) -> Tuple

Partition the data into two or more subsets.

The argument at specifies how to split the data:

  • When at is a number between 0 and 1, this specifies the proportion in the first subset.
  • When at is an integer, it specifies the number of observations in the first subset.
  • When at is a tuple, entries specifies the number or proportion in each subset, except

for the last which will contain the remaning observations. The number of returned subsets is length(at)+1.

If shuffle=true, randomly permute the observations before splitting. A random number generator rng can be optionally passed as the first argument.

If stratified is not nothing, it should be an array of labels with the same length as the data. The observations will be split in such a way that the proportion of each label is preserved in each subset.

Supports any datatype implementing numobs.

It relies on obsview to create views of the data.

Examples

julia> splitobs(reshape(1:100, 1, :); at=0.7)  # simple 70%-30% split, of a matrix
+(1:10, 11:50, 51:100)
source
splitobs([rng,] data; at, shuffle=false, stratified=nothing) -> Tuple

Partition the data into two or more subsets.

The argument at specifies how to split the data:

  • When at is a number between 0 and 1, this specifies the proportion in the first subset.
  • When at is an integer, it specifies the number of observations in the first subset.
  • When at is a tuple, entries specifies the number or proportion in each subset, except

for the last which will contain the remaning observations. The number of returned subsets is length(at)+1.

If shuffle=true, randomly permute the observations before splitting. A random number generator rng can be optionally passed as the first argument.

If stratified is not nothing, it should be an array of labels with the same length as the data. The observations will be split in such a way that the proportion of each label is preserved in each subset.

Supports any datatype implementing numobs.

It relies on obsview to create views of the data.

Examples

julia> splitobs(reshape(1:100, 1, :); at=0.7)  # simple 70%-30% split, of a matrix
 ([1 2 … 69 70], [71 72 … 99 100])
 
 julia> data = (x=ones(2,10), n=1:10)  # a NamedTuple, consistent last dimension
@@ -308,7 +323,7 @@
 true
 
 julia> splitobs(1:10, at=0.5, stratified=[0,0,0,0,1,1,1,1,1,1]) # 2 zeros and 3 ones in each subset
-([1, 2, 5, 6, 7], [3, 4, 8, 9, 10])
source

Array Constructors

MLUtils.falses_like — Function
falses_like(x, [dims=size(x)])

Equivalent to fill_like(x, false, Bool, dims).

See also [fill_like] and trues_like.

source
MLUtils.fill_like — Function
fill_like(x, val, [element_type=eltype(x)], [dims=size(x)]))

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to val. The third and fourth arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

See also zeros_like and ones_like.

Examples

julia> x = rand(Float32, 2)
+([1, 2, 5, 6, 7], [3, 4, 8, 9, 10])
source

Array Constructors

MLUtils.falses_like — Function
falses_like(x, [dims=size(x)])

Equivalent to fill_like(x, false, Bool, dims).

See also [fill_like] and trues_like.

source
MLUtils.fill_like — Function
fill_like(x, val, [element_type=eltype(x)], [dims=size(x)]))

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to val. The third and fourth arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

See also zeros_like and ones_like.

Examples

julia> x = rand(Float32, 2)
 2-element Vector{Float32}:
  0.16087806
  0.89916044
@@ -329,7 +344,7 @@
 julia> fill_like(x, 1.7, Float64)
 2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:
  1.7  1.7
- 1.7  1.7
source
MLUtils.ones_like — Function
ones_like(x, [element_type=eltype(x)], [dims=size(x)]))

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 1. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

See also zeros_like and fill_like.

Examples

julia> x = rand(Float32, 2)
+ 1.7  1.7
source
MLUtils.ones_like — Function
ones_like(x, [element_type=eltype(x)], [dims=size(x)]))

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 1. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

See also zeros_like and fill_like.

Examples

julia> x = rand(Float32, 2)
 2-element Vector{Float32}:
  0.8621633
  0.5158395
@@ -350,7 +365,7 @@
 julia> ones_like(x, Float64)
 2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:
  1.0  1.0
- 1.0  1.0
source
MLUtils.rand_like — Function
rand_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

The default random number generator is used, unless a custom one is passed in explicitly as the first argument.

See also Base.rand and randn_like.

Examples

julia> x = ones(Float32, 2)
+ 1.0  1.0
source
MLUtils.rand_like — Function
rand_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

The default random number generator is used, unless a custom one is passed in explicitly as the first argument.

See also Base.rand and randn_like.

Examples

julia> x = ones(Float32, 2)
 2-element Vector{Float32}:
  1.0
  1.0
@@ -371,7 +386,7 @@
 julia> rand_like(x, Float64)
 2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:
  0.429274  0.135379
- 0.718895  0.0098756
source
MLUtils.randn_like — Function
randn_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value drawn from a normal distribution. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

The default random number generator is used, unless a custom one is passed in explicitly as the first argument.

See also Base.randn and rand_like.

Examples

julia> x = ones(Float32, 2)
+ 0.718895  0.0098756
source
MLUtils.randn_like — Function
randn_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value drawn from a normal distribution. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

The default random number generator is used, unless a custom one is passed in explicitly as the first argument.

See also Base.randn and rand_like.

Examples

julia> x = ones(Float32, 2)
 2-element Vector{Float32}:
  1.0
  1.0
@@ -392,7 +407,7 @@
 julia> randn_like(x, Float64)
 2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:
  -0.578527   0.823445
- -1.01338   -0.612053
source
MLUtils.trues_like — Function
trues_like(x, [dims=size(x)])

Equivalent to fill_like(x, true, Bool, dims).

See also [fill_like] and falses_like.

source
MLUtils.zeros_like — Function
zeros_like(x, [element_type=eltype(x)], [dims=size(x)]))

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 0. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

See also ones_like and fill_like.

Examples

julia> x = rand(Float32, 2)
+ -1.01338   -0.612053
source
MLUtils.trues_like — Function
trues_like(x, [dims=size(x)])

Equivalent to fill_like(x, true, Bool, dims).

See also [fill_like] and falses_like.

source
MLUtils.zeros_like — Function
zeros_like(x, [element_type=eltype(x)], [dims=size(x)]))

Create an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 0. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.

See also ones_like and fill_like.

Examples

julia> x = rand(Float32, 2)
 2-element Vector{Float32}:
  0.4005432
  0.36934233
@@ -413,7 +428,7 @@
 julia> zeros_like(x, Float64)
 2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:
  0.0  0.0
- 0.0  0.0
source

Resampling

MLUtils.oversample — Function
oversample([rng], data, classes; fraction=1, shuffle=true)
+ 0.0  0.0
source

Resampling

MLUtils.oversample — Function
oversample([rng], data, classes; fraction=1, shuffle=true)
 oversample([rng], data::Tuple; fraction=1, shuffle=true)

Generate a re-balanced version of data by repeatedly sampling existing observations in such a way that every class will have at least fraction times the number observations of the largest class in classes. This way, all classes will have a minimum number of observations in the resulting data set relative to what largest class has in the given (original) data.

As an example, by default (i.e. with fraction = 1) the resulting dataset will be near perfectly balanced. On the other hand, with fraction = 0.5 every class in the resulting data with have at least 50% as many observations as the largest class.

The classes input is an array with the same length as numobs(data).

The convenience parameter shuffle determines if the resulting data will be shuffled after its creation; if it is not shuffled then all the repeated samples will be together at the end, sorted by class. Defaults to true.

The random number generator rng can be optionally passed as the first argument.

The output will contain both the resampled data and classes.

# 6 observations with 3 features each
 X = rand(3, 6)
 # 2 classes, severely imbalanced
@@ -451,7 +466,7 @@
    5 │ 0.376304  0.100022   a
    6 │ 0.427064  0.0648339  a
    7 │ 0.427064  0.0648339  a
-   8 │ 0.457043  0.490688   b

See ObsView for more information on data subsets. See also undersample.

source
MLUtils.undersample — Function
undersample([rng], data, classes; shuffle=true)
+   8 │ 0.457043  0.490688   b

See ObsView for more information on data subsets. See also undersample.

source
MLUtils.undersample — Function
undersample([rng], data, classes; shuffle=true)
 undersample([rng], data::Tuple; shuffle=true)

Generate a class-balanced version of data by subsampling its observations in such a way that the resulting number of observations will be the same number for every class. This way, all classes will have as many observations in the resulting data set as the smallest class has in the given (original) data.

The convenience parameter shuffle determines if the resulting data will be shuffled after its creation; if it is not shuffled then all the observations will be in their original order. Defaults to false.

If data is a tuple and classes is not given, then it will be assumed that the last element of the tuple contains the classes.

The output will contain both the resampled data and classes.

# 6 observations with 3 features each
 X = rand(3, 6)
 # 2 classes, severely imbalanced
@@ -485,7 +500,7 @@
    1 │ 0.427064  0.0648339  a
    2 │ 0.376304  0.100022   a
    3 │ 0.467095  0.185437   b
-   4 │ 0.457043  0.490688   b

See ObsView for more information on data subsets. See also oversample.

source

Operations

MLUtils.chunk — Function
chunk(x, n; [dims])
+   4 │ 0.457043  0.490688   b

See ObsView for more information on data subsets. See also oversample.

source

Operations

MLUtils.chunk — Function
chunk(x, n; [dims])
 chunk(x; [size, dims])

Split x into n parts or alternatively, if size is an integer, into equal chunks of size size. The parts contain the same number of elements except possibly for the last one that can be smaller.

In case size is a collection of integers instead, the elements of x are split into chunks of the given sizes.

If x is an array, dims can be used to specify along which dimension to split (defaults to the last dimension).

Examples

julia> chunk(1:10, 3)
 3-element Vector{UnitRange{Int64}}:
  1:4
@@ -535,7 +550,7 @@
 julia> chunk(1:6; size = [2, 4])
 2-element Vector{UnitRange{Int64}}:
  1:2
- 3:6
source
chunk(x, partition_idxs; [npartitions, dims])

Partition the array x along the dimension dims according to the indexes in partition_idxs.

partition_idxs must be sorted and contain only positive integers between 1 and the number of partitions.

If the number of partition npartitions is not provided, it is inferred from partition_idxs.

If dims is not provided, it defaults to the last dimension.

See also unbatch.

Examples

julia> x = reshape([1:10;], 2, 5)
+ 3:6
source
chunk(x, partition_idxs; [npartitions, dims])

Partition the array x along the dimension dims according to the indexes in partition_idxs.

partition_idxs must be sorted and contain only positive integers between 1 and the number of partitions.

If the number of partition npartitions is not provided, it is inferred from partition_idxs.

If dims is not provided, it defaults to the last dimension.

See also unbatch.

Examples

julia> x = reshape([1:10;], 2, 5)
 2×5 Matrix{Int64}:
  1  3  5  7   9
  2  4  6  8  10
@@ -544,17 +559,17 @@
 3-element Vector{SubArray{Int64, 2, Matrix{Int64}, Tuple{Base.Slice{Base.OneTo{Int64}}, UnitRange{Int64}}, true}}:
  [1; 2;;]
  [3 5; 4 6]
- [7 9; 8 10]
source
MLUtils.flatten — Function
flatten(x::AbstractArray)

Reshape arbitrarly-shaped input into a matrix-shaped output, preserving the size of the last dimension.

See also unsqueeze.

Examples

julia> rand(3,4,5) |> flatten |> size
-(12, 5)
source
MLUtils.group_counts — Function
group_counts(x)

Count the number of times that each element of x appears.

See also group_indices

Examples

julia> group_counts(['a', 'b', 'b'])
+ [7 9; 8 10]
source
MLUtils.flatten — Function
flatten(x::AbstractArray)

Reshape arbitrarly-shaped input into a matrix-shaped output, preserving the size of the last dimension.

See also unsqueeze.

Examples

julia> rand(3,4,5) |> flatten |> size
+(12, 5)
source
MLUtils.group_counts — Function
group_counts(x)

Count the number of times that each element of x appears.

See also group_indices

Examples

julia> group_counts(['a', 'b', 'b'])
 Dict{Char, Int64} with 2 entries:
   'a' => 1
-  'b' => 2
source
MLUtils.group_indices — Function
group_indices(x) -> Dict

Computes the indices of elements in the vector x for each distinct value contained. This information is useful for resampling strategies, such as stratified sampling.

See also group_counts.

Examples

julia> x = [:yes, :no, :maybe, :yes];
+  'b' => 2
source
MLUtils.group_indices — Function
group_indices(x) -> Dict

Computes the indices of elements in the vector x for each distinct value contained. This information is useful for resampling strategies, such as stratified sampling.

See also group_counts.

Examples

julia> x = [:yes, :no, :maybe, :yes];
 
 julia> group_indices(x)
 Dict{Symbol, Vector{Int64}} with 3 entries:
   :yes   => [1, 4]
   :maybe => [3]
-  :no    => [2]
source
MLUtils.normalise — Function
normalise(x; dims=ndims(x), ϵ=1e-5)

Normalise the array x to mean 0 and standard deviation 1 across the dimension(s) given by dims. Per default, dims is the last dimension.

ϵ is a small additive factor added to the denominator for numerical stability.

source
MLUtils.rpad_constant — Function
rpad_constant(v::AbstractArray, n::Union{Integer, Tuple}, val = 0; dims=:)

Return the given sequence padded with val along the dimensions dims up to a maximum length in each direction specified by n.

Examples

julia> rpad_constant([1, 2], 4, -1) # passing with -1 up to size 4
+  :no    => [2]
source
MLUtils.normalise — Function
normalise(x; dims=ndims(x), ϵ=1e-5)

Normalise the array x to mean 0 and standard deviation 1 across the dimension(s) given by dims. Per default, dims is the last dimension.

ϵ is a small additive factor added to the denominator for numerical stability.

source
MLUtils.rpad_constant — Function
rpad_constant(v::AbstractArray, n::Union{Integer, Tuple}, val = 0; dims=:)

Return the given sequence padded with val along the dimensions dims up to a maximum length in each direction specified by n.

Examples

julia> rpad_constant([1, 2], 4, -1) # passing with -1 up to size 4
 4-element Vector{Int64}:
   1
   2
@@ -579,13 +594,13 @@
  1  2  0  0
  3  4  0  0
  0  0  0  0
- 0  0  0  0
source
MLUtils.unbatch — Function
unbatch(x)

Reverse of the batch operation, unstacking the last dimension of the array x.

See also unstack and chunk.

Examples

julia> unbatch([1 3 5 7;
+ 0  0  0  0
source
MLUtils.unbatch — Function
unbatch(x)

Reverse of the batch operation, unstacking the last dimension of the array x.

See also unstack and chunk.

Examples

julia> unbatch([1 3 5 7;
                 2 4 6 8])
 4-element Vector{Vector{Int64}}:
  [1, 2]
  [3, 4]
  [5, 6]
- [7, 8]
source
MLUtils.unsqueeze — Function
unsqueeze(x; dims)

Return x reshaped into an array one dimensionality higher than x, where dims indicates in which dimension x is extended. dims can be an integer between 1 and ndims(x)+1.

See also flatten, stack.

Examples

julia> unsqueeze([1 2; 3 4], dims=2)
+ [7, 8]
source
MLUtils.unsqueeze — Function
unsqueeze(x; dims)

Return x reshaped into an array one dimensionality higher than x, where dims indicates in which dimension x is extended. dims can be an integer between 1 and ndims(x)+1.

See also flatten, stack.

Examples

julia> unsqueeze([1 2; 3 4], dims=2)
 2×1×2 Array{Int64, 3}:
 [:, :, 1] =
  1
@@ -604,10 +619,10 @@
 
 julia> unsqueeze(xs, dims=1)
 1×3 Matrix{Vector{Int64}}:
- [1, 2]  [3, 4]  [5, 6]
source
unsqueeze(; dims)

Returns a function which, acting on an array, inserts a dimension of size 1 at dims.

Examples

julia> rand(21, 22, 23) |> unsqueeze(dims=2) |> size
-(21, 1, 22, 23)
source
MLUtils.unstack — Function
unstack(xs; dims)

Unroll the given xs into an array of arrays along the given dimension dims.

It is the inverse operation of stack.

See also unbatch and chunk.

Examples

julia> unstack([1 3 5 7; 2 4 6 8], dims=2)
+ [1, 2]  [3, 4]  [5, 6]
source
unsqueeze(; dims)

Returns a function which, acting on an array, inserts a dimension of size 1 at dims.

Examples

julia> rand(21, 22, 23) |> unsqueeze(dims=2) |> size
+(21, 1, 22, 23)
source
MLUtils.unstack — Function
unstack(xs; dims)

Unroll the given xs into an array of arrays along the given dimension dims.

It is the inverse operation of stack.

See also unbatch and chunk.

Examples

julia> unstack([1 3 5 7; 2 4 6 8], dims=2)
 4-element Vector{Vector{Int64}}:
  [1, 2]
  [3, 4]
  [5, 6]
- [7, 8]
source

Datasets

MLUtils.Datasets.load_iris — Function
load_iris() -> X, y, names

Loads the first 150 observations from the Iris flower data set introduced by Ronald Fisher (1936). The 4 by 150 matrix X contains the numeric measurements, in which each individual column denotes an observation. The vector y contains the class labels as strings. The vector names contains the names of the features (i.e. rows of X)

[1] Fisher, Ronald A. "The use of multiple measurements in taxonomic problems." Annals of eugenics 7.2 (1936): 179-188.

source
MLUtils.Datasets.make_sin — Function
make_sin(n, start, stop; noise = 0.3, f_rand = randn) -> x, y

Generates n noisy equally spaces samples of a sinus from start to stop by adding noise .* f_rand(length(x)) to the result of sin(x).

Returns the vector x with the samples and the noisy response y.

source
MLUtils.Datasets.make_spiral — Function
make_spiral(n, a, theta, b; noise = 0.01, f_rand = randn) -> x, y

Generates n noisy responses for a spiral with two labels. Uses the radius, angle and scaling arguments to space the points in 2D space and adding noise .* f_randn(n) to the response.

Returns the 2 x n matrix x with the coordinates of the samples and the vector y with the labels.

source
MLUtils.Datasets.make_poly — Function
make_poly(coef, x; noise = 0.01, f_rand = randn) -> x, y

Generates a noisy response for a polynomial of degree length(coef) and with the coefficients given by coef. The response is generated by elmentwise computation of the polynome on the elements of x and adding noise .* f_randn(length(x)) to the result.

The vector coef contains the coefficients for the terms of the polynome. The first element of coef denotes the coefficient for the term with the highest degree, while the last element of coef denotes the intercept.

Return the input x and the noisy response y.

source
MLUtils.Datasets.make_moons — Function
make_moons(n; noise=0.0, f_rand=randn, shuffle=true) -> x, y

Generate a dataset with two interleaving half circles.

If n is an integer, the number of samples is n and the number of samples for each half circle is n ÷ 2. If n is a tuple, the first element of the tuple denotes the number of samples for the first half circle and the second element denotes the number of samples for the second half circle.

The noise level can be controlled by the noise argument.

Set shuffle=false to keep the order of the samples.

Returns a 2 x n matrix with the the samples.

source
+ [7, 8]source

Datasets

MLUtils.Datasets.load_iris — Function
load_iris() -> X, y, names

Loads the first 150 observations from the Iris flower data set introduced by Ronald Fisher (1936). The 4 by 150 matrix X contains the numeric measurements, in which each individual column denotes an observation. The vector y contains the class labels as strings. The vector names contains the names of the features (i.e. rows of X)

[1] Fisher, Ronald A. "The use of multiple measurements in taxonomic problems." Annals of eugenics 7.2 (1936): 179-188.

source
MLUtils.Datasets.make_sin — Function
make_sin(n, start, stop; noise = 0.3, f_rand = randn) -> x, y

Generates n noisy equally spaces samples of a sinus from start to stop by adding noise .* f_rand(length(x)) to the result of sin(x).

Returns the vector x with the samples and the noisy response y.

source
MLUtils.Datasets.make_spiral — Function
make_spiral(n, a, theta, b; noise = 0.01, f_rand = randn) -> x, y

Generates n noisy responses for a spiral with two labels. Uses the radius, angle and scaling arguments to space the points in 2D space and adding noise .* f_randn(n) to the response.

Returns the 2 x n matrix x with the coordinates of the samples and the vector y with the labels.

source
MLUtils.Datasets.make_poly — Function
make_poly(coef, x; noise = 0.01, f_rand = randn) -> x, y

Generates a noisy response for a polynomial of degree length(coef) and with the coefficients given by coef. The response is generated by elmentwise computation of the polynome on the elements of x and adding noise .* f_randn(length(x)) to the result.

The vector coef contains the coefficients for the terms of the polynome. The first element of coef denotes the coefficient for the term with the highest degree, while the last element of coef denotes the intercept.

Return the input x and the noisy response y.

source
MLUtils.Datasets.make_moons — Function
make_moons(n; noise=0.0, f_rand=randn, shuffle=true) -> x, y

Generate a dataset with two interleaving half circles.

If n is an integer, the number of samples is n and the number of samples for each half circle is n ÷ 2. If n is a tuple, the first element of the tuple denotes the number of samples for the first half circle and the second element denotes the number of samples for the second half circle.

The noise level can be controlled by the noise argument.

Set shuffle=false to keep the order of the samples.

Returns a 2 x n matrix with the the samples.

source
diff --git a/dev/index.html b/dev/index.html index 6381575..b9b4450 100644 --- a/dev/index.html +++ b/dev/index.html @@ -25,4 +25,4 @@ # ... train supervised model on minibatches here end end -end

In the above code snippet, the inner loop for eachobs is the only place where data other than indices is actually being copied. In fact, while x and y are materialized arrays, all the rest are data views.

Historical Notes

MLUtils.jl brings together functionalities previously found in LearnBase.jl , MLDataPattern.jl and MLLabelUtils.jl. These packages are now discontinued.

Other features were ported from the deep learning library Flux.jl, as they are of general use.

+end

In the above code snippet, the inner loop for eachobs is the only place where data other than indices is actually being copied. In fact, while x and y are materialized arrays, all the rest are data views.

Historical Notes

MLUtils.jl brings together functionalities previously found in LearnBase.jl , MLDataPattern.jl and MLLabelUtils.jl. These packages are now discontinued.

Other features were ported from the deep learning library Flux.jl, as they are of general use.

diff --git a/dev/objects.inv b/dev/objects.inv index 4222012937e07fd71ef4602e18bca439c99bfbdc..f7297e540a9eb9e6efbdc05b3e1fc96e3c1b093b 100644 GIT binary patch delta 580 zcmV-K0=xaW1;Yi9iho5Zl4EYJ(;@-f#xUAkjTO1FX;Y-ahaLOtOVZB-bjcockT^TL zJ3G{J$?%OB%kG(?q`oEOQFT~B&(A;l*RNzSeyaRaPq0Oel43*%Xiyw5pT7hYIU=eN zY&Z394brqqcP1GTFyw@(o_o|_xsqvkT@97SI^;$VpC|#2=G*T#Z3n9zJ@=g|;SWpzFnDz(8x* z)I%?f)s9-F(*i{pIfqpbV8NihBg6bIlGBgu$}3yBXnYzJFh1y-9O0-2{WSnAtG+cz~8um(Tvi}6Zncom=ud7;T3vSHHZ=W0tfzAsiA2-~~p6`;4 zKKxjEu08!?$Xd#KGw!~uvPXL14L#+WmpIFBmxA`587LHHL0zD#LFMIH43 delta 570 zcmV-A0>%Bq1-S)~ihqSEl4EYJ+adwn#xUAkjWxBhX;Y-aha3CrOVST10bQ~O9VE`q z?#>RiT-NwXjIHl$#R$D6h5+B##T;lQb9xR#usE!=3t9pf+CRNd&~iiH$$j5hdWL8riBg8TNQ53e}x#H?QE^M zEGWgvA_?}Wm6;2^$!Z{7TaflJ*X$5M7%WSBm=V6@`+qgon=}X09n^CcZX`&X1N&QB zSmmZsqAr^CL_v)@AnmThq=Z_zu4uyM#UbY18wuxJ)|_E}G+c x = (a = [1, 2, 3], b = rand(6, 3));\n\njulia> getobs(x, 2) == (a = 2, b = x.b[:, 2])\ntrue\n\njulia> getobs(x, [1, 3]) == (a = [1, 3], b = x.b[:, [1, 3]])\ntrue\n\njulia> x = Dict(:a => [1, 2, 3], :b => rand(6, 3));\n\njulia> getobs(x, 2) == Dict(:a => 2, :b => x[:b][:, 2])\ntrue\n\njulia> getobs(x, [1, 3]) == Dict(:a => [1, 3], :b => x[:b][:, [1, 3]])\ntrue\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.getobs!","page":"API","title":"MLUtils.getobs!","text":"getobs!(buffer, data, idx)\n\nInplace version of getobs(data, idx). If this method is defined for the type of data, then buffer should be used to store the result, instead of allocating a dedicated object.\n\nImplementing this function is optional. In the case no such method is provided for the type of data, then buffer will be ignored and the result of getobs returned. This could be because the type of data may not lend itself to the concept of copy!. Thus, supporting a custom getobs! is optional and not required.\n\nCustom implementations of getobs! should be consistent with getobs in terms of the output format, that is getobs!(buffer, data, idx) == getobs(data, idx).\n\nSee also getobs and numobs. \n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.numobs","page":"API","title":"MLUtils.numobs","text":"numobs(data)\n\nReturn the total number of observations contained in data.\n\nIf data does not have numobs defined, then in the case of Tables.table(data) == true returns the number of rows, otherwise returns length(data).\n\nAuthors of custom data containers should implement Base.length for their type instead of numobs. numobs should only be implemented for types where there is a difference between numobs and Base.length (such as multi-dimensional arrays).\n\ngetobs supports by default nested combinations of array, tuple, named tuples, and dictionaries. \n\nSee also getobs.\n\nExamples\n\njulia> x = (a = [1, 2, 3], b = ones(6, 3)); # named tuples\n\njulia> numobs(x)\n3\n\njulia> x = Dict(:a => [1, 2, 3], :b => ones(6, 3)); # dictionaries\n\njulia> numobs(x) \n3\n\nAll internal containers must have the same number of observations:\n\njulia> x = (a = [1, 2, 3, 4], b = ones(6, 3));\n\njulia> numobs(x)\nERROR: DimensionMismatch: All data containers must have the same number of observations.\nStacktrace:\n [1] _check_numobs_error()\n @ MLUtils ~/.julia/dev/MLUtils/src/observation.jl:163\n [2] _check_numobs\n @ ~/.julia/dev/MLUtils/src/observation.jl:130 [inlined]\n [3] numobs(data::NamedTuple{(:a, :b), Tuple{Vector{Int64}, Matrix{Float64}}})\n @ MLUtils ~/.julia/dev/MLUtils/src/observation.jl:177\n [4] top-level scope\n @ REPL[35]:1\n\n\n\n\n\n","category":"function"},{"location":"api/#Lazy-Transforms","page":"API","title":"Lazy Transforms","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"filterobs\ngroupobs\njoinobs\nmapobs\nshuffleobs","category":"page"},{"location":"api/#MLUtils.filterobs","page":"API","title":"MLUtils.filterobs","text":"filterobs(f, data)\n\nReturn a subset of data container data including all indices i for which f(getobs(data, i)) === true.\n\ndata = 1:10\nnumobs(data) == 10\nfdata = filterobs(>(5), data)\nnumobs(fdata) == 5\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.groupobs","page":"API","title":"MLUtils.groupobs","text":"groupobs(f, data)\n\nSplit data container data data into different data containers, grouping observations by f(obs).\n\ndata = -10:10\ndatas = groupobs(>(0), data)\nlength(datas) == 2\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.joinobs","page":"API","title":"MLUtils.joinobs","text":"joinobs(datas...)\n\nConcatenate data containers datas.\n\ndata1, data2 = 1:10, 11:20\njdata = joinumobs(data1, data2)\ngetobs(jdata, 15) == 15\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.mapobs","page":"API","title":"MLUtils.mapobs","text":"mapobs(f, data; batched=:auto)\n\nLazily map f over the observations in a data container data. Returns a new data container mdata that can be indexed and has a length. Indexing triggers the transformation f.\n\nThe batched keyword argument controls the behavior of mdata[idx] and mdata[idxs] where idx is an integer and idxs is a vector of integers:\n\nbatched=:auto (default). Let f handle the two cases. Calls f(getobs(data, idx)) and f(getobs(data, idxs)).\nbatched=:never. The function f is always called on a single observation. Calls f(getobs(data, idx)) and [f(getobs(data, idx)) for idx in idxs].\nbatched=:always. The function f is always called on a batch of observations. Calls getobs(f(getobs(data, [idx])), 1) and f(getobs(data, idxs)).\n\nExamples\n\njulia> data = (a=[1,2,3], b=[1,2,3]);\n\njulia> mdata = mapobs(data) do x\n (c = x.a .+ x.b, d = x.a .- x.b)\n end\nmapobs(#25, (a = [1, 2, 3], b = [1, 2, 3]); batched=:auto))\n\njulia> mdata[1]\n(c = 2, d = 0)\n\njulia> mdata[1:2]\n(c = [2, 4], d = [0, 0])\n\n\n\n\n\nmapobs(fs, data)\n\nLazily map each function in tuple fs over the observations in data container data. Returns a tuple of transformed data containers.\n\n\n\n\n\nmapobs(namedfs::NamedTuple, data)\n\nMap a NamedTuple of functions over data, turning it into a data container of NamedTuples. Field syntax can be used to select a column of the resulting data container.\n\ndata = 1:10\nnameddata = mapobs((x = sqrt, y = log), data)\ngetobs(nameddata, 10) == (x = sqrt(10), y = log(10))\ngetobs(nameddata.x, 10) == sqrt(10)\n\n\n\n\n\nmapobs(f, d::DataLoader)\n\nReturn a new dataloader based on d that applies f at each iteration. \n\nExamples\n\njulia> X = ones(3, 6);\n\njulia> function f(x)\n @show x\n return x\n end\nf (generic function with 1 method)\n\njulia> d = DataLoader(X, batchsize=2, collate=false);\n\njulia> d = mapobs(f, d);\n\njulia> for x in d\n @assert size(x) == (2,)\n @assert size(x[1]) == (3,)\n end\nx = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]\nx = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]\nx = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]\n\njulia> d2 = DataLoader(X, batchsize=2, collate=true);\n\njulia> d2 = mapobs(f, d2);\n\njulia> for x in d2\n @assert size(x) == (3, 2)\n end\nx = [1.0 1.0; 1.0 1.0; 1.0 1.0]\nx = [1.0 1.0; 1.0 1.0; 1.0 1.0]\nx = [1.0 1.0; 1.0 1.0; 1.0 1.0]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.shuffleobs","page":"API","title":"MLUtils.shuffleobs","text":"shuffleobs([rng], data)\n\nReturn a version of the dataset data that contains all the origin observations in a random reordering.\n\nThe values of data itself are not copied. Instead only the indices are shuffled. This function calls obsview to accomplish that, which means that the return value is likely of a different type than data.\n\nOptionally, a random number generator rng can be passed as the first argument. \n\nFor this function to work, the type of data must implement numobs and getobs. \n\nSee also obsview.\n\nExamples\n\n# For Arrays the subset will be of type SubArray\n@assert typeof(shuffleobs(rand(4,10))) <: SubArray\n\n# Iterate through all observations in random order\nfor x in eachobs(shuffleobs(X))\n ...\nend\n\n\n\n\n\n","category":"function"},{"location":"api/#Batching,-Iteration,-and-Views","page":"API","title":"Batching, Iteration, and Views","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"batch\nbatchsize\nbatchseq\nBatchView\neachobs\nDataLoader\nobsview\nObsView\nrandobs","category":"page"},{"location":"api/#MLUtils.batch","page":"API","title":"MLUtils.batch","text":"batch(xs)\n\nBatch the arrays in xs into a single array with an extra dimension.\n\nIf the elements of xs are tuples, named tuples, or dicts, the output will be of the same type. \n\nSee also unbatch.\n\nExamples\n\njulia> batch([[1,2,3], \n [4,5,6]])\n3×2 Matrix{Int64}:\n 1 4\n 2 5\n 3 6\n\njulia> batch([(a=[1,2], b=[3,4])\n (a=[5,6], b=[7,8])]) \n(a = [1 5; 2 6], b = [3 7; 4 8])\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.batchsize","page":"API","title":"MLUtils.batchsize","text":"batchsize(data::BatchView) -> Int\n\nReturn the fixed size of each batch in data.\n\nExamples\n\nusing MLUtils\nX, Y = MLUtils.load_iris()\n\nA = BatchView(X, batchsize=30)\n@assert batchsize(A) == 30\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.batchseq","page":"API","title":"MLUtils.batchseq","text":"batchseq(seqs, val = 0)\n\nTake a list of N sequences, and turn them into a single sequence where each item is a batch of N. Short sequences will be padded by val.\n\nExamples\n\njulia> batchseq([[1, 2, 3], [4, 5]], 0)\n3-element Vector{Vector{Int64}}:\n [1, 4]\n [2, 5]\n [3, 0]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.BatchView","page":"API","title":"MLUtils.BatchView","text":"BatchView(data, batchsize; partial=true, collate=nothing)\nBatchView(data; batchsize=1, partial=true, collate=nothing)\n\nCreate a view of the given data that represents it as a vector of batches. Each batch will contain an equal amount of observations in them. The batch-size can be specified using the parameter batchsize. In the case that the size of the dataset is not dividable by the specified batchsize, the remaining observations will be ignored if partial=false. If partial=true instead the last batch-size can be slightly smaller.\n\nIf used as an iterator, the object will iterate over the dataset once, effectively denoting an epoch. \n\nAny data access is delayed until iteration or indexing is perfomed. The getobs function is called on the data object to retrieve the observations.\n\nFor BatchView to work on some data structure, the type of the given variable data must implement the data container interface. See ObsView for more info.\n\nArguments\n\ndata : The object describing the dataset. Can be of any type as long as it implements getobs and numobs (see Details for more information).\nbatchsize : The batch-size of each batch. It is the number of observations that each batch must contain (except possibly for the last one).\npartial : If partial=false and the number of observations is not divisible by the batch-size, then the last mini-batch is dropped.\ncollate: Defines the batching behavior. \nIf nothing (default), a batch is getobs(data, indices). \nIf false, each batch is [getobs(data, i) for i in indices]. \nIf true, applies MLUtils to the vector of observations in a batch, recursively collating arrays in the last dimensions. See MLUtils.batch for more information and examples.\nIf a custom function, it will be used in place of MLUtils.batch. It should take a vector of observations as input.\n\nSe also DataLoader.\n\nExamples\n\njulia> using MLUtils\n\njulia> X, Y = MLUtils.load_iris();\n\njulia> A = BatchView(X, batchsize=30);\n\njulia> @assert eltype(A) <: Matrix{Float64}\n\njulia> @assert length(A) == 5 # Iris has 150 observations\n\njulia> @assert size(A[1]) == (4,30) # Iris has 4 features\n\njulia> for x in BatchView(X, batchsize=30)\n # 5 batches of size 30 observations\n @assert size(x) == (4, 30)\n @assert numobs(x) === 30\n end\n\njulia> for (x, y) in BatchView((X, Y), batchsize=20, partial=true)\n # 7 batches of size 20 observations + 1 batch of 10 observations\n @assert typeof(x) <: Matrix{Float64}\n @assert typeof(y) <: Vector{String}\n end\n\njulia> for batch in BatchView((X, Y), batchsize=20, partial=false, collate=false)\n # 7 batches of size 20 observations\n @assert length(batch) == 20\n x1, y1 = batch[1]\n end\n\njulia> function collate_fn(batch)\n # collate observations into a custom batch\n return hcat([x[1] for x in batch]...), join([x[2] for x in batch])\n end;\n\njulia> for (x, y) in BatchView((rand(10, 4), [\"a\", \"b\", \"c\", \"d\"]), batchsize=2, collate=collate_fn)\n @assert size(x) == (10, 2)\n @assert y isa String\n end\n\n\n\n\n\n","category":"type"},{"location":"api/#MLUtils.eachobs","page":"API","title":"MLUtils.eachobs","text":"eachobs(data; kws...)\n\nReturn an iterator over data.\n\nSupports the same arguments as DataLoader. The batchsize default is -1 here while it is 1 for DataLoader.\n\nExamples\n\nX = rand(4,100)\n\nfor x in eachobs(X)\n # loop entered 100 times\n @assert typeof(x) <: Vector{Float64}\n @assert size(x) == (4,)\nend\n\n# mini-batch iterations\nfor x in eachobs(X, batchsize=10)\n # loop entered 10 times\n @assert typeof(x) <: Matrix{Float64}\n @assert size(x) == (4,10)\nend\n\n# support for tuples, named tuples, dicts\nfor (x, y) in eachobs((X, Y))\n # ...\nend\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.DataLoader","page":"API","title":"MLUtils.DataLoader","text":"DataLoader(data; [batchsize, buffer, collate, parallel, partial, rng, shuffle])\n\nAn object that iterates over mini-batches of data, each mini-batch containing batchsize observations (except possibly the last one).\n\nTakes as input a single data array, a tuple (or a named tuple) of arrays, or in general any data object that implements the numobs and getobs methods.\n\nThe last dimension in each array is the observation dimension, i.e. the one divided into mini-batches.\n\nThe original data is preserved in the data field of the DataLoader.\n\nArguments\n\ndata: The data to be iterated over. The data type has to be supported by numobs and getobs.\nbatchsize: If less than 0, iterates over individual observations. Otherwise, each iteration (except possibly the last) yields a mini-batch containing batchsize observations. Default 1.\nbuffer: If buffer=true and supported by the type of data, a buffer will be allocated and reused for memory efficiency. May want to set partial=false to avoid size mismatch. Finally, can pass an external buffer to be used in getobs! (depending on the collate and batchsize options, could be getobs!(buffer, data, idxs) or getobs!(buffer[i], data, idx)). Default false. \ncollate: Defines the batching behavior. Default nothing. \nIf nothing , a batch is getobs(data, indices). \nIf false, each batch is [getobs(data, i) for i in indices]. \nIf true, applies MLUtils.batch to the vector of observations in a batch, recursively collating arrays in the last dimensions. See MLUtils.batch for more information and examples.\nIf a custom function, it will be used in place of MLUtils.batch. It should take a vector of observations as input.\nparallel: Whether to use load data in parallel using worker threads. Greatly speeds up data loading by factor of available threads. Requires starting Julia with multiple threads. Check Threads.nthreads() to see the number of available threads. Passing parallel = true breaks ordering guarantees. Default false.\npartial: This argument is used only when batchsize > 0. If partial=false and the number of observations is not divisible by the batchsize, then the last mini-batch is dropped. Default true.\nrng: A random number generator. Default Random.default_rng().\n`shuffle: Whether to shuffle the observations before iterating. Unlike wrapping the data container with shuffleobs(data), shuffle=true ensures that the observations are shuffled anew every time you start iterating over eachobs. Default false.\n\nExamples\n\njulia> Xtrain = rand(10, 100);\n\njulia> array_loader = DataLoader(Xtrain, batchsize=2);\n\njulia> for x in array_loader\n @assert size(x) == (10, 2)\n # do something with x, 50 times\n end\n\njulia> array_loader.data === Xtrain\ntrue\n\njulia> tuple_loader = DataLoader((Xtrain,), batchsize=2); # similar, but yielding 1-element tuples\n\njulia> for x in tuple_loader\n @assert x isa Tuple{Matrix}\n @assert size(x[1]) == (10, 2)\n end\n\njulia> Ytrain = rand('a':'z', 100); # now make a DataLoader yielding 2-element named tuples\n\njulia> train_loader = DataLoader((data=Xtrain, label=Ytrain), batchsize=5, shuffle=true);\n\njulia> for epoch in 1:100\n for (x, y) in train_loader # access via tuple destructuring\n @assert size(x) == (10, 5)\n @assert size(y) == (5,)\n # loss += f(x, y) # etc, runs 100 * 20 times\n end\n end\n\njulia> first(train_loader).label isa Vector{Char} # access via property name\ntrue\n\njulia> first(train_loader).label == Ytrain[1:5] # because of shuffle=true\nfalse\n\njulia> foreach(println∘summary, DataLoader(rand(Int8, 10, 64), batchsize=30)) # partial=false would omit last\n10×30 Matrix{Int8}\n10×30 Matrix{Int8}\n10×4 Matrix{Int8}\n\njulia> collate_fn(batch) = join(batch);\n\njulia> first(DataLoader([\"a\", \"b\", \"c\", \"d\"], batchsize=2, collate=collate_fn))\n\"ab\"\n\n\n\n\n\n","category":"type"},{"location":"api/#MLUtils.obsview","page":"API","title":"MLUtils.obsview","text":"obsview(data, [indices])\n\nReturns a lazy view of the observations in data that correspond to the given indices. No data will be copied except of the indices. It is similar to constructing an ObsView, but returns a SubArray if the type of data is Array or SubArray. Furthermore, this function may be extended for custom types of data that also want to provide their own subset-type.\n\nIn case data is a tuple, the constructor will be mapped over its elements. That means that the constructor returns a tuple of ObsView instead of a ObsView of tuples.\n\nIf instead you want to get the subset of observations corresponding to the given indices in their native type, use getobs.\n\nSee ObsView for more information.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.ObsView","page":"API","title":"MLUtils.ObsView","text":"ObsView(data, [indices])\n\nUsed to represent a subset of some data of arbitrary type by storing which observation-indices the subset spans. Furthermore, subsequent subsettings are accumulated without needing to access actual data.\n\nThe main purpose for the existence of ObsView is to delay data access and movement until an actual batch of data (or single observation) is needed for some computation. This is particularily useful when the data is not located in memory, but on the hard drive or some remote location. In such a scenario one wants to load the required data only when needed.\n\nAny data access is delayed until getindex is called, and even getindex returns the result of obsview which in general avoids data movement until getobs is called. If used as an iterator, the view will iterate over the dataset once, effectively denoting an epoch. Each iteration will return a lazy subset to the current observation.\n\nArguments\n\ndata : The object describing the dataset. Can be of any type as long as it implements getobs and numobs (see Details for more information).\nindices : Optional. The index or indices of the observation(s) in data that the subset should represent. Can be of type Int or some subtype of AbstractVector.\n\nMethods\n\ngetindex : Returns the observation(s) of the given index/indices. No data is copied aside from the required indices.\nnumobs : Returns the total number observations in the subset.\ngetobs : Returns the underlying data that the ObsView represents at the given relative indices. Note that these indices are in \"subset space\", and in general will not directly correspond to the same indices in the underlying data set.\n\nDetails\n\nFor ObsView to work on some data structure, the desired type MyType must implement the following interface:\n\ngetobs(data::MyType, idx) : Should return the observation(s) indexed by idx. In what form is up to the user. Note that idx can be of type Int or AbstractVector.\nnumobs(data::MyType) : Should return the total number of observations in data\n\nThe following methods can also be provided and are optional:\n\ngetobs(data::MyType) : By default this function is the identity function. If that is not the behaviour that you want for your type, you need to provide this method as well.\nobsview(data::MyType, idx) : If your custom type has its own kind of subset type, you can return it here. An example for such a case are SubArray for representing a subset of some AbstractArray.\ngetobs!(buffer, data::MyType, [idx]) : Inplace version of getobs(data, idx). If this method is provided for MyType, then eachobs can preallocate a buffer that is then reused every iteration. Note: buffer should be equivalent to the return value of getobs(::MyType, ...), since this is how buffer is preallocated by default.\n\nExamples\n\nX, Y = MLUtils.load_iris()\n\n# The iris set has 150 observations and 4 features\n@assert size(X) == (4,150)\n\n# Represents the 80 observations as a ObsView\nv = ObsView(X, 21:100)\n@assert numobs(v) == 80\n@assert typeof(v) <: ObsView\n# getobs indexes into v\n@assert getobs(v, 1:10) == X[:, 21:30]\n\n# Use `obsview` to avoid boxing into ObsView\n# for types that provide a custom \"subset\", such as arrays.\n# Here it instead creates a native SubArray.\nv = obsview(X, 1:100)\n@assert numobs(v) == 100\n@assert typeof(v) <: SubArray\n\n# Also works for tuples of arbitrary length\nsubset = obsview((X, Y), 1:100)\n@assert numobs(subset) == 100\n@assert typeof(subset) <: Tuple # tuple of SubArray\n\n# Use as iterator\nfor x in ObsView(X)\n @assert typeof(x) <: SubArray{Float64,1}\nend\n\n# iterate over each individual labeled observation\nfor (x, y) in ObsView((X, Y))\n @assert typeof(x) <: SubArray{Float64,1}\n @assert typeof(y) <: String\nend\n\n# same but in random order\nfor (x, y) in ObsView(shuffleobs((X, Y)))\n @assert typeof(x) <: SubArray{Float64,1}\n @assert typeof(y) <: String\nend\n\n# Indexing: take first 10 observations\nx, y = ObsView((X, Y))[1:10]\n\nSee also\n\nobsview, getobs, numobs, splitobs, shuffleobs, kfolds.\n\n\n\n\n\n","category":"type"},{"location":"api/#MLUtils.randobs","page":"API","title":"MLUtils.randobs","text":"randobs(data, [n])\n\nPick a random observation or a batch of n random observations from data. For this function to work, the type of data must implement numobs and getobs.\n\n\n\n\n\n","category":"function"},{"location":"api/#Partitioning","page":"API","title":"Partitioning","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"leavepout\nkfolds\nsplitobs","category":"page"},{"location":"api/#MLUtils.leavepout","page":"API","title":"MLUtils.leavepout","text":"leavepout(n::Integer, [size = 1]) -> Tuple\n\nCompute the train/validation assignments for k ≈ n/size repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. Each validation subset will have either size or size+1 observations assigned to it. The following code snippet generates the index-vectors for size = 2.\n\njulia> train_idx, val_idx = leavepout(10, 2);\n\nEach observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.\n\njulia> train_idx\n5-element Array{Array{Int64,1},1}:\n [3,4,5,6,7,8,9,10]\n [1,2,5,6,7,8,9,10]\n [1,2,3,4,7,8,9,10]\n [1,2,3,4,5,6,9,10]\n [1,2,3,4,5,6,7,8]\n\njulia> val_idx\n5-element Array{UnitRange{Int64},1}:\n 1:2\n 3:4\n 5:6\n 7:8\n 9:10\n\n\n\n\n\nleavepout(data, p = 1)\n\nRepartition a data container using a k-fold strategy, where k is chosen in such a way, that each validation subset of the resulting folds contains roughly p observations. Defaults to p = 1, which is also known as \"leave-one-out\" partitioning.\n\nThe resulting sequence of folds is returned as a lazy iterator. Only data subsets are created. That means no actual data is copied until getobs is invoked.\n\nfor (train, val) in leavepout(X, p=2)\n # if nobs(X) is dividable by 2,\n # then numobs(val) will be 2 for each iteraton,\n # otherwise it may be 3 for the first few iterations.\nend\n\nSeekfolds for a related function.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.kfolds","page":"API","title":"MLUtils.kfolds","text":"kfolds(n::Integer, k = 5) -> Tuple\n\nCompute the train/validation assignments for k repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. A general rule of thumb is to use either k = 5 or k = 10. The following code snippet generates the indices assignments for k = 5\n\njulia> train_idx, val_idx = kfolds(10, 5);\n\nEach observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.\n\njulia> train_idx\n5-element Array{Array{Int64,1},1}:\n [3,4,5,6,7,8,9,10]\n [1,2,5,6,7,8,9,10]\n [1,2,3,4,7,8,9,10]\n [1,2,3,4,5,6,9,10]\n [1,2,3,4,5,6,7,8]\n\njulia> val_idx\n5-element Array{UnitRange{Int64},1}:\n 1:2\n 3:4\n 5:6\n 7:8\n 9:10\n\n\n\n\n\nkfolds(data, [k = 5])\n\nRepartition a data container k times using a k folds strategy and return the sequence of folds as a lazy iterator. Only data subsets are created, which means that no actual data is copied until getobs is invoked.\n\nConceptually, a k-folds repartitioning strategy divides the given data into k roughly equal-sized parts. Each part will serve as validation set once, while the remaining parts are used for training. This results in k different partitions of data.\n\nIn the case that the size of the dataset is not dividable by the specified k, the remaining observations will be evenly distributed among the parts.\n\nfor (x_train, x_val) in kfolds(X, k=10)\n # code called 10 times\n # nobs(x_val) may differ up to ±1 over iterations\nend\n\nMultiple variables are supported (e.g. for labeled data)\n\nfor ((x_train, y_train), val) in kfolds((X, Y), k=10)\n # ...\nend\n\nBy default the folds are created using static splits. Use shuffleobs to randomly assign observations to the folds.\n\nfor (x_train, x_val) in kfolds(shuffleobs(X), k = 10)\n # ...\nend\n\nSee leavepout for a related function.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.splitobs","page":"API","title":"MLUtils.splitobs","text":"splitobs(n::Int; at) -> Tuple\n\nCompute the indices for two or more disjoint subsets of the range 1:n with split sizes determined by at.\n\nExamples\n\njulia> splitobs(100, at=0.7)\n(1:70, 71:100)\n\njulia> splitobs(100, at=(0.1, 0.4))\n(1:10, 11:50, 51:100)\n\n\n\n\n\nsplitobs([rng,] data; at, shuffle=false, stratified=nothing) -> Tuple\n\nPartition the data into two or more subsets.\n\nThe argument at specifies how to split the data:\n\nWhen at is a number between 0 and 1, this specifies the proportion in the first subset.\nWhen at is an integer, it specifies the number of observations in the first subset.\nWhen at is a tuple, entries specifies the number or proportion in each subset, except\n\nfor the last which will contain the remaning observations. The number of returned subsets is length(at)+1.\n\nIf shuffle=true, randomly permute the observations before splitting. A random number generator rng can be optionally passed as the first argument.\n\nIf stratified is not nothing, it should be an array of labels with the same length as the data. The observations will be split in such a way that the proportion of each label is preserved in each subset.\n\nSupports any datatype implementing numobs. \n\nIt relies on obsview to create views of the data.\n\nExamples\n\njulia> splitobs(reshape(1:100, 1, :); at=0.7) # simple 70%-30% split, of a matrix\n([1 2 … 69 70], [71 72 … 99 100])\n\njulia> data = (x=ones(2,10), n=1:10) # a NamedTuple, consistent last dimension\n(x = [1.0 1.0 … 1.0 1.0; 1.0 1.0 … 1.0 1.0], n = 1:10)\n\njulia> splitobs(data, at=(0.5, 0.3)) # a 50%-30%-20% split, e.g. train/test/validation\n((x = [1.0 1.0 … 1.0 1.0; 1.0 1.0 … 1.0 1.0], n = 1:5), (x = [1.0 1.0 1.0; 1.0 1.0 1.0], n = 6:8), (x = [1.0 1.0; 1.0 1.0], n = 9:10))\n\njulia> train, test = splitobs((reshape(1.0:100.0, 1, :), 101:200), at=0.7, shuffle=true); # split a Tuple\n\njulia> vec(test[1]) .+ 100 == test[2]\ntrue\n\njulia> splitobs(1:10, at=0.5, stratified=[0,0,0,0,1,1,1,1,1,1]) # 2 zeros and 3 ones in each subset\n([1, 2, 5, 6, 7], [3, 4, 8, 9, 10])\n\n\n\n\n\n","category":"function"},{"location":"api/#Array-Constructors","page":"API","title":"Array Constructors","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"falses_like\nfill_like\nones_like\nrand_like\nrandn_like\ntrues_like\nzeros_like","category":"page"},{"location":"api/#MLUtils.falses_like","page":"API","title":"MLUtils.falses_like","text":"falses_like(x, [dims=size(x)])\n\nEquivalent to fill_like(x, false, Bool, dims).\n\nSee also [fill_like] and trues_like.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.fill_like","page":"API","title":"MLUtils.fill_like","text":"fill_like(x, val, [element_type=eltype(x)], [dims=size(x)]))\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to val. The third and fourth arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nSee also zeros_like and ones_like.\n\nExamples\n\njulia> x = rand(Float32, 2)\n2-element Vector{Float32}:\n 0.16087806\n 0.89916044\n\njulia> fill_like(x, 1.7, (3, 3))\n3×3 Matrix{Float32}:\n 1.7 1.7 1.7\n 1.7 1.7 1.7\n 1.7 1.7 1.7\n\njulia> using CUDA\n\njulia> x = CUDA.rand(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 0.803167 0.476101\n 0.303041 0.317581\n\njulia> fill_like(x, 1.7, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 1.7 1.7\n 1.7 1.7\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.ones_like","page":"API","title":"MLUtils.ones_like","text":"ones_like(x, [element_type=eltype(x)], [dims=size(x)]))\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 1. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nSee also zeros_like and fill_like.\n\nExamples\n\njulia> x = rand(Float32, 2)\n2-element Vector{Float32}:\n 0.8621633\n 0.5158395\n\njulia> ones_like(x, (3, 3))\n3×3 Matrix{Float32}:\n 1.0 1.0 1.0\n 1.0 1.0 1.0\n 1.0 1.0 1.0\n\njulia> using CUDA\n\njulia> x = CUDA.rand(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 0.82297 0.656143\n 0.701828 0.391335\n\njulia> ones_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 1.0 1.0\n 1.0 1.0\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.rand_like","page":"API","title":"MLUtils.rand_like","text":"rand_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nThe default random number generator is used, unless a custom one is passed in explicitly as the first argument.\n\nSee also Base.rand and randn_like.\n\nExamples\n\njulia> x = ones(Float32, 2)\n2-element Vector{Float32}:\n 1.0\n 1.0\n\njulia> rand_like(x, (3, 3))\n3×3 Matrix{Float32}:\n 0.780032 0.920552 0.53689\n 0.121451 0.741334 0.5449\n 0.55348 0.138136 0.556404\n\njulia> using CUDA\n\njulia> CUDA.ones(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 1.0 1.0\n 1.0 1.0\n\njulia> rand_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 0.429274 0.135379\n 0.718895 0.0098756\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.randn_like","page":"API","title":"MLUtils.randn_like","text":"randn_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value drawn from a normal distribution. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nThe default random number generator is used, unless a custom one is passed in explicitly as the first argument.\n\nSee also Base.randn and rand_like.\n\nExamples\n\njulia> x = ones(Float32, 2)\n2-element Vector{Float32}:\n 1.0\n 1.0\n\njulia> randn_like(x, (3, 3))\n3×3 Matrix{Float32}:\n -0.385331 0.956231 0.0745102\n 1.43756 -0.967328 2.06311\n 0.0482372 1.78728 -0.902547\n\njulia> using CUDA\n\njulia> CUDA.ones(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 1.0 1.0\n 1.0 1.0\n\njulia> randn_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n -0.578527 0.823445\n -1.01338 -0.612053\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.trues_like","page":"API","title":"MLUtils.trues_like","text":"trues_like(x, [dims=size(x)])\n\nEquivalent to fill_like(x, true, Bool, dims).\n\nSee also [fill_like] and falses_like.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.zeros_like","page":"API","title":"MLUtils.zeros_like","text":"zeros_like(x, [element_type=eltype(x)], [dims=size(x)]))\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 0. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nSee also ones_like and fill_like.\n\nExamples\n\njulia> x = rand(Float32, 2)\n2-element Vector{Float32}:\n 0.4005432\n 0.36934233\n\njulia> zeros_like(x, (3, 3))\n3×3 Matrix{Float32}:\n 0.0 0.0 0.0\n 0.0 0.0 0.0\n 0.0 0.0 0.0\n\njulia> using CUDA\n\njulia> x = CUDA.rand(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 0.0695155 0.667979\n 0.558468 0.59903\n\njulia> zeros_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 0.0 0.0\n 0.0 0.0\n\n\n\n\n\n","category":"function"},{"location":"api/#Resampling","page":"API","title":"Resampling","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"oversample\nundersample","category":"page"},{"location":"api/#MLUtils.oversample","page":"API","title":"MLUtils.oversample","text":"oversample([rng], data, classes; fraction=1, shuffle=true)\noversample([rng], data::Tuple; fraction=1, shuffle=true)\n\nGenerate a re-balanced version of data by repeatedly sampling existing observations in such a way that every class will have at least fraction times the number observations of the largest class in classes. This way, all classes will have a minimum number of observations in the resulting data set relative to what largest class has in the given (original) data.\n\nAs an example, by default (i.e. with fraction = 1) the resulting dataset will be near perfectly balanced. On the other hand, with fraction = 0.5 every class in the resulting data with have at least 50% as many observations as the largest class.\n\nThe classes input is an array with the same length as numobs(data). \n\nThe convenience parameter shuffle determines if the resulting data will be shuffled after its creation; if it is not shuffled then all the repeated samples will be together at the end, sorted by class. Defaults to true.\n\nThe random number generator rng can be optionally passed as the first argument. \n\nThe output will contain both the resampled data and classes.\n\n# 6 observations with 3 features each\nX = rand(3, 6)\n# 2 classes, severely imbalanced\nY = [\"a\", \"b\", \"b\", \"b\", \"b\", \"a\"]\n\n# oversample the class \"a\" to match \"b\"\nX_bal, Y_bal = oversample(X, Y)\n\n# this results in a bigger dataset with repeated data\n@assert size(X_bal) == (3,8)\n@assert length(Y_bal) == 8\n\n# now both \"a\", and \"b\" have 4 observations each\n@assert sum(Y_bal .== \"a\") == 4\n@assert sum(Y_bal .== \"b\") == 4\n\nFor this function to work, the type of data must implement numobs and getobs. \n\nIf data is a tuple and classes is not given, then it will be assumed that the last element of the tuple contains the classes.\n\njulia> data = DataFrame(X1=rand(6), X2=rand(6), Y=[:a,:b,:b,:b,:b,:a])\n6×3 DataFrames.DataFrame\n│ Row │ X1 │ X2 │ Y │\n├─────┼───────────┼─────────────┼───┤\n│ 1 │ 0.226582 │ 0.0443222 │ a │\n│ 2 │ 0.504629 │ 0.722906 │ b │\n│ 3 │ 0.933372 │ 0.812814 │ b │\n│ 4 │ 0.522172 │ 0.245457 │ b │\n│ 5 │ 0.505208 │ 0.11202 │ b │\n│ 6 │ 0.0997825 │ 0.000341996 │ a │\n\njulia> getobs(oversample(data, data.Y))\n8×3 DataFrame\n Row │ X1 X2 Y \n │ Float64 Float64 Symbol \n─────┼─────────────────────────────\n 1 │ 0.376304 0.100022 a\n 2 │ 0.467095 0.185437 b\n 3 │ 0.481957 0.319906 b\n 4 │ 0.336762 0.390811 b\n 5 │ 0.376304 0.100022 a\n 6 │ 0.427064 0.0648339 a\n 7 │ 0.427064 0.0648339 a\n 8 │ 0.457043 0.490688 b\n\nSee ObsView for more information on data subsets. See also undersample.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.undersample","page":"API","title":"MLUtils.undersample","text":"undersample([rng], data, classes; shuffle=true)\nundersample([rng], data::Tuple; shuffle=true)\n\nGenerate a class-balanced version of data by subsampling its observations in such a way that the resulting number of observations will be the same number for every class. This way, all classes will have as many observations in the resulting data set as the smallest class has in the given (original) data.\n\nThe convenience parameter shuffle determines if the resulting data will be shuffled after its creation; if it is not shuffled then all the observations will be in their original order. Defaults to false.\n\nIf data is a tuple and classes is not given, then it will be assumed that the last element of the tuple contains the classes.\n\nThe output will contain both the resampled data and classes.\n\n# 6 observations with 3 features each\nX = rand(3, 6)\n# 2 classes, severely imbalanced\nY = [\"a\", \"b\", \"b\", \"b\", \"b\", \"a\"]\n\n# subsample the class \"b\" to match \"a\"\nX_bal, Y_bal = undersample(X, Y)\n\n# this results in a smaller dataset\n@assert size(X_bal) == (3,4)\n@assert length(Y_bal) == 4\n\n# now both \"a\", and \"b\" have 2 observations each\n@assert sum(Y_bal .== \"a\") == 2\n@assert sum(Y_bal .== \"b\") == 2\n\nFor this function to work, the type of data must implement numobs and getobs. \n\nNote that if data is a tuple, then it will be assumed that the last element of the tuple contains the targets.\n\njulia> data = DataFrame(X1=rand(6), X2=rand(6), Y=[:a,:b,:b,:b,:b,:a])\n6×3 DataFrames.DataFrame\n│ Row │ X1 │ X2 │ Y │\n├─────┼───────────┼─────────────┼───┤\n│ 1 │ 0.226582 │ 0.0443222 │ a │\n│ 2 │ 0.504629 │ 0.722906 │ b │\n│ 3 │ 0.933372 │ 0.812814 │ b │\n│ 4 │ 0.522172 │ 0.245457 │ b │\n│ 5 │ 0.505208 │ 0.11202 │ b │\n│ 6 │ 0.0997825 │ 0.000341996 │ a │\n\njulia> getobs(undersample(data, data.Y))\n4×3 DataFrame\n Row │ X1 X2 Y \n │ Float64 Float64 Symbol \n─────┼─────────────────────────────\n 1 │ 0.427064 0.0648339 a\n 2 │ 0.376304 0.100022 a\n 3 │ 0.467095 0.185437 b\n 4 │ 0.457043 0.490688 b\n\nSee ObsView for more information on data subsets. See also oversample.\n\n\n\n\n\n","category":"function"},{"location":"api/#Operations","page":"API","title":"Operations","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"chunk\nflatten\ngroup_counts\ngroup_indices\nnormalise\nrpad_constant\nunbatch\nunsqueeze\nunstack","category":"page"},{"location":"api/#MLUtils.chunk","page":"API","title":"MLUtils.chunk","text":"chunk(x, n; [dims])\nchunk(x; [size, dims])\n\nSplit x into n parts or alternatively, if size is an integer, into equal chunks of size size. The parts contain the same number of elements except possibly for the last one that can be smaller.\n\nIn case size is a collection of integers instead, the elements of x are split into chunks of the given sizes.\n\nIf x is an array, dims can be used to specify along which dimension to split (defaults to the last dimension).\n\nExamples\n\njulia> chunk(1:10, 3)\n3-element Vector{UnitRange{Int64}}:\n 1:4\n 5:8\n 9:10\n\njulia> chunk(1:10; size = 2)\n5-element Vector{UnitRange{Int64}}:\n 1:2\n 3:4\n 5:6\n 7:8\n 9:10\n\njulia> x = reshape(collect(1:20), (5, 4))\n5×4 Matrix{Int64}:\n 1 6 11 16\n 2 7 12 17\n 3 8 13 18\n 4 9 14 19\n 5 10 15 20\n\njulia> xs = chunk(x, 2, dims=1)\n2-element Vector{SubArray{Int64, 2, Matrix{Int64}, Tuple{UnitRange{Int64}, Base.Slice{Base.OneTo{Int64}}}, false}}:\n [1 6 11 16; 2 7 12 17; 3 8 13 18]\n [4 9 14 19; 5 10 15 20]\n\njulia> xs[1]\n3×4 view(::Matrix{Int64}, 1:3, :) with eltype Int64:\n 1 6 11 16\n 2 7 12 17\n 3 8 13 18\n\njulia> xes = chunk(x; size = 2, dims = 2)\n2-element Vector{SubArray{Int64, 2, Matrix{Int64}, Tuple{Base.Slice{Base.OneTo{Int64}}, UnitRange{Int64}}, true}}:\n [1 6; 2 7; … ; 4 9; 5 10]\n [11 16; 12 17; … ; 14 19; 15 20]\n\njulia> xes[2]\n5×2 view(::Matrix{Int64}, :, 3:4) with eltype Int64:\n 11 16\n 12 17\n 13 18\n 14 19\n 15 20\n\njulia> chunk(1:6; size = [2, 4])\n2-element Vector{UnitRange{Int64}}:\n 1:2\n 3:6\n\n\n\n\n\nchunk(x, partition_idxs; [npartitions, dims])\n\nPartition the array x along the dimension dims according to the indexes in partition_idxs.\n\npartition_idxs must be sorted and contain only positive integers between 1 and the number of partitions. \n\nIf the number of partition npartitions is not provided, it is inferred from partition_idxs.\n\nIf dims is not provided, it defaults to the last dimension.\n\nSee also unbatch.\n\nExamples\n\njulia> x = reshape([1:10;], 2, 5)\n2×5 Matrix{Int64}:\n 1 3 5 7 9\n 2 4 6 8 10\n\njulia> chunk(x, [1, 2, 2, 3, 3])\n3-element Vector{SubArray{Int64, 2, Matrix{Int64}, Tuple{Base.Slice{Base.OneTo{Int64}}, UnitRange{Int64}}, true}}:\n [1; 2;;]\n [3 5; 4 6]\n [7 9; 8 10]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.flatten","page":"API","title":"MLUtils.flatten","text":"flatten(x::AbstractArray)\n\nReshape arbitrarly-shaped input into a matrix-shaped output, preserving the size of the last dimension.\n\nSee also unsqueeze.\n\nExamples\n\njulia> rand(3,4,5) |> flatten |> size\n(12, 5)\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.group_counts","page":"API","title":"MLUtils.group_counts","text":"group_counts(x)\n\nCount the number of times that each element of x appears.\n\nSee also group_indices\n\nExamples\n\njulia> group_counts(['a', 'b', 'b'])\nDict{Char, Int64} with 2 entries:\n 'a' => 1\n 'b' => 2\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.group_indices","page":"API","title":"MLUtils.group_indices","text":"group_indices(x) -> Dict\n\nComputes the indices of elements in the vector x for each distinct value contained. This information is useful for resampling strategies, such as stratified sampling.\n\nSee also group_counts.\n\nExamples\n\njulia> x = [:yes, :no, :maybe, :yes];\n\njulia> group_indices(x)\nDict{Symbol, Vector{Int64}} with 3 entries:\n :yes => [1, 4]\n :maybe => [3]\n :no => [2]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.normalise","page":"API","title":"MLUtils.normalise","text":"normalise(x; dims=ndims(x), ϵ=1e-5)\n\nNormalise the array x to mean 0 and standard deviation 1 across the dimension(s) given by dims. Per default, dims is the last dimension. \n\nϵ is a small additive factor added to the denominator for numerical stability.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.rpad_constant","page":"API","title":"MLUtils.rpad_constant","text":"rpad_constant(v::AbstractArray, n::Union{Integer, Tuple}, val = 0; dims=:)\n\nReturn the given sequence padded with val along the dimensions dims up to a maximum length in each direction specified by n.\n\nExamples\n\njulia> rpad_constant([1, 2], 4, -1) # passing with -1 up to size 4\n4-element Vector{Int64}:\n 1\n 2\n -1\n -1\n\njulia> rpad_constant([1, 2, 3], 2) # no padding if length is already greater than n\n3-element Vector{Int64}:\n 1\n 2\n 3\n\njulia> rpad_constant([1 2; 3 4], 4; dims=1) # padding along the first dimension\n4×2 Matrix{Int64}:\n 1 2\n 3 4\n 0 0\n 0 0\n\njulia> rpad_constant([1 2; 3 4], 4) # padding along all dimensions by default\n4×4 Matrix{Int64}:\n 1 2 0 0\n 3 4 0 0\n 0 0 0 0\n 0 0 0 0\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.unbatch","page":"API","title":"MLUtils.unbatch","text":"unbatch(x)\n\nReverse of the batch operation, unstacking the last dimension of the array x.\n\nSee also unstack and chunk.\n\nExamples\n\njulia> unbatch([1 3 5 7;\n 2 4 6 8])\n4-element Vector{Vector{Int64}}:\n [1, 2]\n [3, 4]\n [5, 6]\n [7, 8]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.unsqueeze","page":"API","title":"MLUtils.unsqueeze","text":"unsqueeze(x; dims)\n\nReturn x reshaped into an array one dimensionality higher than x, where dims indicates in which dimension x is extended. dims can be an integer between 1 and ndims(x)+1.\n\nSee also flatten, stack.\n\nExamples\n\njulia> unsqueeze([1 2; 3 4], dims=2)\n2×1×2 Array{Int64, 3}:\n[:, :, 1] =\n 1\n 3\n\n[:, :, 2] =\n 2\n 4\n\n\njulia> xs = [[1, 2], [3, 4], [5, 6]]\n3-element Vector{Vector{Int64}}:\n [1, 2]\n [3, 4]\n [5, 6]\n\njulia> unsqueeze(xs, dims=1)\n1×3 Matrix{Vector{Int64}}:\n [1, 2] [3, 4] [5, 6]\n\n\n\n\n\nunsqueeze(; dims)\n\nReturns a function which, acting on an array, inserts a dimension of size 1 at dims.\n\nExamples\n\njulia> rand(21, 22, 23) |> unsqueeze(dims=2) |> size\n(21, 1, 22, 23)\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.unstack","page":"API","title":"MLUtils.unstack","text":"unstack(xs; dims)\n\nUnroll the given xs into an array of arrays along the given dimension dims.\n\nIt is the inverse operation of stack.\n\nSee also unbatch and chunk.\n\nExamples\n\njulia> unstack([1 3 5 7; 2 4 6 8], dims=2)\n4-element Vector{Vector{Int64}}:\n [1, 2]\n [3, 4]\n [5, 6]\n [7, 8]\n\n\n\n\n\n","category":"function"},{"location":"api/#Datasets","page":"API","title":"Datasets","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"Datasets.load_iris\nDatasets.make_sin\nDatasets.make_spiral\nDatasets.make_poly\nDatasets.make_moons","category":"page"},{"location":"api/#MLUtils.Datasets.load_iris","page":"API","title":"MLUtils.Datasets.load_iris","text":"load_iris() -> X, y, names\n\nLoads the first 150 observations from the Iris flower data set introduced by Ronald Fisher (1936). The 4 by 150 matrix X contains the numeric measurements, in which each individual column denotes an observation. The vector y contains the class labels as strings. The vector names contains the names of the features (i.e. rows of X)\n\n[1] Fisher, Ronald A. \"The use of multiple measurements in taxonomic problems.\" Annals of eugenics 7.2 (1936): 179-188.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_sin","page":"API","title":"MLUtils.Datasets.make_sin","text":"make_sin(n, start, stop; noise = 0.3, f_rand = randn) -> x, y\n\nGenerates n noisy equally spaces samples of a sinus from start to stop by adding noise .* f_rand(length(x)) to the result of sin(x).\n\nReturns the vector x with the samples and the noisy response y.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_spiral","page":"API","title":"MLUtils.Datasets.make_spiral","text":"make_spiral(n, a, theta, b; noise = 0.01, f_rand = randn) -> x, y\n\nGenerates n noisy responses for a spiral with two labels. Uses the radius, angle and scaling arguments to space the points in 2D space and adding noise .* f_randn(n) to the response.\n\nReturns the 2 x n matrix x with the coordinates of the samples and the vector y with the labels.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_poly","page":"API","title":"MLUtils.Datasets.make_poly","text":"make_poly(coef, x; noise = 0.01, f_rand = randn) -> x, y\n\nGenerates a noisy response for a polynomial of degree length(coef) and with the coefficients given by coef. The response is generated by elmentwise computation of the polynome on the elements of x and adding noise .* f_randn(length(x)) to the result.\n\nThe vector coef contains the coefficients for the terms of the polynome. The first element of coef denotes the coefficient for the term with the highest degree, while the last element of coef denotes the intercept.\n\nReturn the input x and the noisy response y.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_moons","page":"API","title":"MLUtils.Datasets.make_moons","text":"make_moons(n; noise=0.0, f_rand=randn, shuffle=true) -> x, y\n\nGenerate a dataset with two interleaving half circles. \n\nIf n is an integer, the number of samples is n and the number of samples for each half circle is n ÷ 2. If n is a tuple, the first element of the tuple denotes the number of samples for the first half circle and the second element denotes the number of samples for the second half circle.\n\nThe noise level can be controlled by the noise argument.\n\nSet shuffle=false to keep the order of the samples.\n\nReturns a 2 x n matrix with the the samples. \n\n\n\n\n\n","category":"function"},{"location":"#MLUtils.jl","page":"Home","title":"MLUtils.jl","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"(Image: ) (Image: ) (Image: ) (Image: )","category":"page"},{"location":"","page":"Home","title":"Home","text":"MLUtils.jl defines interfaces and implements common utilities for Machine Learning pipelines.","category":"page"},{"location":"#Features","page":"Home","title":"Features","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"An extensible dataset interface (numobs and getobs).\nData iteration and dataloaders (eachobs and DataLoader).\nLazy data views (obsview). \nResampling procedures (undersample and oversample).\nTrain/test splits (splitobs) \nData partitioning and aggregation tools (batch, unbatch, chunk, group_counts, group_indices).\nFolds for cross-validation (kfolds, leavepout).\nDatasets lazy tranformations (mapobs, filterobs, groupobs, joinobs, shuffleobs).\nToy datasets for demonstration purpose. \nOther data handling utilities (flatten, normalise, unsqueeze, stack, unstack).","category":"page"},{"location":"#Examples","page":"Home","title":"Examples","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"Let us take a look at a hello world example to get a feeling for how to use this package in a typical ML scenario. ","category":"page"},{"location":"","page":"Home","title":"Home","text":"using MLUtils\n\n# X is a matrix of floats\n# Y is a vector of strings\nX, Y = load_iris()\n\n# The iris dataset is ordered according to their labels,\n# which means that we should shuffle the dataset before\n# partitioning it into training- and test-set.\nXs, Ys = shuffleobs((X, Y))\n\n# We leave out 15 % of the data for testing\ncv_data, test_data = splitobs((Xs, Ys); at=0.85)\n\n# Next we partition the data using a 10-fold scheme.\nfor (train_data, val_data) in kfolds(cv_data; k=10)\n\n # We apply a lazy transform for data augmentation\n train_data = mapobs(xy -> (xy[1] .+ 0.1 .* randn.(), xy[2]), train_data)\n\n for epoch = 1:10\n # Iterate over the data using mini-batches of 5 observations each\n for (x, y) in eachobs(train_data, batchsize=5)\n # ... train supervised model on minibatches here\n end\n end\nend","category":"page"},{"location":"","page":"Home","title":"Home","text":"In the above code snippet, the inner loop for eachobs is the only place where data other than indices is actually being copied. In fact, while x and y are materialized arrays, all the rest are data views. ","category":"page"},{"location":"#Historical-Notes","page":"Home","title":"Historical Notes","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"MLUtils.jl brings together functionalities previously found in LearnBase.jl , MLDataPattern.jl and MLLabelUtils.jl. These packages are now discontinued. ","category":"page"},{"location":"","page":"Home","title":"Home","text":"Other features were ported from the deep learning library Flux.jl, as they are of general use. ","category":"page"},{"location":"#Alternatives-and-Related-Packages","page":"Home","title":"Alternatives and Related Packages","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"MLJ.jl is a more complete package for managing the whole machine learning pipeline if you are looking for a sklearn replacement.\nNNlib.jl provides utility functions for neural networks.\nTableTransforms.jl contains transformations for tabular datasets.\nDataAugmentation.jl. Efficient, composable data augmentation for machine and deep learning with support for n-dimensional images, keypoints and categorical masks.","category":"page"}] +[{"location":"api/","page":"API","title":"API","text":"CollapsedDocStrings = true","category":"page"},{"location":"api/#API-Reference","page":"API","title":"API Reference","text":"","category":"section"},{"location":"api/#Core-API","page":"API","title":"Core API","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"getobs\ngetobs!\nnumobs","category":"page"},{"location":"api/#MLUtils.getobs","page":"API","title":"MLUtils.getobs","text":"getobs(data, [idx])\n\nReturn the observations corresponding to the observation index idx. Note that idx can be any type as long as data has defined getobs for that type. If idx is not provided, then materialize all observations in data.\n\nIf data does not have getobs defined, then in the case of Tables.table(data) == true returns the row(s) in position idx, otherwise returns data[idx].\n\nAuthors of custom data containers should implement Base.getindex for their type instead of getobs. getobs should only be implemented for types where there is a difference between getobs and Base.getindex (such as multi-dimensional arrays).\n\nThe returned observation(s) should be in the form intended to be passed as-is to some learning algorithm. There is no strict interface requirement on how this \"actual data\" must look like. Every author behind some custom data container can make this decision themselves. The output should be consistent when idx is a scalar vs vector.\n\ngetobs supports by default nested combinations of array, tuple, named tuples, and dictionaries. \n\nSee also getobs! and numobs.\n\nExamples\n\njulia> x = (a = [1, 2, 3], b = rand(6, 3));\n\njulia> getobs(x, 2) == (a = 2, b = x.b[:, 2])\ntrue\n\njulia> getobs(x, [1, 3]) == (a = [1, 3], b = x.b[:, [1, 3]])\ntrue\n\njulia> x = Dict(:a => [1, 2, 3], :b => rand(6, 3));\n\njulia> getobs(x, 2) == Dict(:a => 2, :b => x[:b][:, 2])\ntrue\n\njulia> getobs(x, [1, 3]) == Dict(:a => [1, 3], :b => x[:b][:, [1, 3]])\ntrue\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.getobs!","page":"API","title":"MLUtils.getobs!","text":"getobs!(buffer, data, idx)\n\nInplace version of getobs(data, idx). If this method is defined for the type of data, then buffer should be used to store the result, instead of allocating a dedicated object.\n\nImplementing this function is optional. In the case no such method is provided for the type of data, then buffer will be ignored and the result of getobs returned. This could be because the type of data may not lend itself to the concept of copy!. Thus, supporting a custom getobs! is optional and not required.\n\nCustom implementations of getobs! should be consistent with getobs in terms of the output format, that is getobs!(buffer, data, idx) == getobs(data, idx).\n\nSee also getobs and numobs. \n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.numobs","page":"API","title":"MLUtils.numobs","text":"numobs(data)\n\nReturn the total number of observations contained in data.\n\nIf data does not have numobs defined, then in the case of Tables.table(data) == true returns the number of rows, otherwise returns length(data).\n\nAuthors of custom data containers should implement Base.length for their type instead of numobs. numobs should only be implemented for types where there is a difference between numobs and Base.length (such as multi-dimensional arrays).\n\ngetobs supports by default nested combinations of array, tuple, named tuples, and dictionaries. \n\nSee also getobs.\n\nExamples\n\njulia> x = (a = [1, 2, 3], b = ones(6, 3)); # named tuples\n\njulia> numobs(x)\n3\n\njulia> x = Dict(:a => [1, 2, 3], :b => ones(6, 3)); # dictionaries\n\njulia> numobs(x) \n3\n\nAll internal containers must have the same number of observations:\n\njulia> x = (a = [1, 2, 3, 4], b = ones(6, 3));\n\njulia> numobs(x)\nERROR: DimensionMismatch: All data containers must have the same number of observations.\nStacktrace:\n [1] _check_numobs_error()\n @ MLUtils ~/.julia/dev/MLUtils/src/observation.jl:163\n [2] _check_numobs\n @ ~/.julia/dev/MLUtils/src/observation.jl:130 [inlined]\n [3] numobs(data::NamedTuple{(:a, :b), Tuple{Vector{Int64}, Matrix{Float64}}})\n @ MLUtils ~/.julia/dev/MLUtils/src/observation.jl:177\n [4] top-level scope\n @ REPL[35]:1\n\n\n\n\n\n","category":"function"},{"location":"api/#Lazy-Transforms","page":"API","title":"Lazy Transforms","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"filterobs\ngroupobs\njoinobs\nmapobs\nshuffleobs","category":"page"},{"location":"api/#MLUtils.filterobs","page":"API","title":"MLUtils.filterobs","text":"filterobs(f, data)\n\nReturn a subset of data container data including all indices i for which f(getobs(data, i)) === true.\n\ndata = 1:10\nnumobs(data) == 10\nfdata = filterobs(>(5), data)\nnumobs(fdata) == 5\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.groupobs","page":"API","title":"MLUtils.groupobs","text":"groupobs(f, data)\n\nSplit data container data data into different data containers, grouping observations by f(obs).\n\ndata = -10:10\ndatas = groupobs(>(0), data)\nlength(datas) == 2\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.joinobs","page":"API","title":"MLUtils.joinobs","text":"joinobs(datas...)\n\nConcatenate data containers datas.\n\ndata1, data2 = 1:10, 11:20\njdata = joinumobs(data1, data2)\ngetobs(jdata, 15) == 15\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.mapobs","page":"API","title":"MLUtils.mapobs","text":"mapobs(f, data; batched=:auto)\n\nLazily map f over the observations in a data container data. Returns a new data container mdata that can be indexed and has a length. Indexing triggers the transformation f.\n\nThe batched keyword argument controls the behavior of mdata[idx] and mdata[idxs] where idx is an integer and idxs is a vector of integers:\n\nbatched=:auto (default). Let f handle the two cases. Calls f(getobs(data, idx)) and f(getobs(data, idxs)).\nbatched=:never. The function f is always called on a single observation. Calls f(getobs(data, idx)) and [f(getobs(data, idx)) for idx in idxs].\nbatched=:always. The function f is always called on a batch of observations. Calls getobs(f(getobs(data, [idx])), 1) and f(getobs(data, idxs)).\n\nExamples\n\njulia> data = (a=[1,2,3], b=[1,2,3]);\n\njulia> mdata = mapobs(data) do x\n (c = x.a .+ x.b, d = x.a .- x.b)\n end\nmapobs(#25, (a = [1, 2, 3], b = [1, 2, 3]); batched=:auto))\n\njulia> mdata[1]\n(c = 2, d = 0)\n\njulia> mdata[1:2]\n(c = [2, 4], d = [0, 0])\n\n\n\n\n\nmapobs(fs, data)\n\nLazily map each function in tuple fs over the observations in data container data. Returns a tuple of transformed data containers.\n\n\n\n\n\nmapobs(namedfs::NamedTuple, data)\n\nMap a NamedTuple of functions over data, turning it into a data container of NamedTuples. Field syntax can be used to select a column of the resulting data container.\n\ndata = 1:10\nnameddata = mapobs((x = sqrt, y = log), data)\ngetobs(nameddata, 10) == (x = sqrt(10), y = log(10))\ngetobs(nameddata.x, 10) == sqrt(10)\n\n\n\n\n\nmapobs(f, d::DataLoader)\n\nReturn a new dataloader based on d that applies f at each iteration. \n\nExamples\n\njulia> X = ones(3, 6);\n\njulia> function f(x)\n @show x\n return x\n end\nf (generic function with 1 method)\n\njulia> d = DataLoader(X, batchsize=2, collate=false);\n\njulia> d = mapobs(f, d);\n\njulia> for x in d\n @assert size(x) == (2,)\n @assert size(x[1]) == (3,)\n end\nx = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]\nx = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]\nx = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]\n\njulia> d2 = DataLoader(X, batchsize=2, collate=true);\n\njulia> d2 = mapobs(f, d2);\n\njulia> for x in d2\n @assert size(x) == (3, 2)\n end\nx = [1.0 1.0; 1.0 1.0; 1.0 1.0]\nx = [1.0 1.0; 1.0 1.0; 1.0 1.0]\nx = [1.0 1.0; 1.0 1.0; 1.0 1.0]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.shuffleobs","page":"API","title":"MLUtils.shuffleobs","text":"shuffleobs([rng], data)\n\nReturn a version of the dataset data that contains all the origin observations in a random reordering.\n\nThe values of data itself are not copied. Instead only the indices are shuffled. This function calls obsview to accomplish that, which means that the return value is likely of a different type than data.\n\nOptionally, a random number generator rng can be passed as the first argument. \n\nFor this function to work, the type of data must implement numobs and getobs. \n\nSee also obsview.\n\nExamples\n\n# For Arrays the subset will be of type SubArray\n@assert typeof(shuffleobs(rand(4,10))) <: SubArray\n\n# Iterate through all observations in random order\nfor x in eachobs(shuffleobs(X))\n ...\nend\n\n\n\n\n\n","category":"function"},{"location":"api/#Batching,-Iteration,-and-Views","page":"API","title":"Batching, Iteration, and Views","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"batch\nbatchsize\nbatchseq\nBatchView\neachobs\nDataLoader\nobsview\nObsView\nrandobs\nslidingwindow","category":"page"},{"location":"api/#MLUtils.batch","page":"API","title":"MLUtils.batch","text":"batch(xs)\n\nBatch the arrays in xs into a single array with an extra dimension.\n\nIf the elements of xs are tuples, named tuples, or dicts, the output will be of the same type. \n\nSee also unbatch.\n\nExamples\n\njulia> batch([[1,2,3], \n [4,5,6]])\n3×2 Matrix{Int64}:\n 1 4\n 2 5\n 3 6\n\njulia> batch([(a=[1,2], b=[3,4])\n (a=[5,6], b=[7,8])]) \n(a = [1 5; 2 6], b = [3 7; 4 8])\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.batchsize","page":"API","title":"MLUtils.batchsize","text":"batchsize(data::BatchView) -> Int\n\nReturn the fixed size of each batch in data.\n\nExamples\n\nusing MLUtils\nX, Y = MLUtils.load_iris()\n\nA = BatchView(X, batchsize=30)\n@assert batchsize(A) == 30\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.batchseq","page":"API","title":"MLUtils.batchseq","text":"batchseq(seqs, val = 0)\n\nTake a list of N sequences, and turn them into a single sequence where each item is a batch of N. Short sequences will be padded by val.\n\nExamples\n\njulia> batchseq([[1, 2, 3], [4, 5]], 0)\n3-element Vector{Vector{Int64}}:\n [1, 4]\n [2, 5]\n [3, 0]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.BatchView","page":"API","title":"MLUtils.BatchView","text":"BatchView(data, batchsize; partial=true, collate=nothing)\nBatchView(data; batchsize=1, partial=true, collate=nothing)\n\nCreate a view of the given data that represents it as a vector of batches. Each batch will contain an equal amount of observations in them. The batch-size can be specified using the parameter batchsize. In the case that the size of the dataset is not dividable by the specified batchsize, the remaining observations will be ignored if partial=false. If partial=true instead the last batch-size can be slightly smaller.\n\nIf used as an iterator, the object will iterate over the dataset once, effectively denoting an epoch. \n\nAny data access is delayed until iteration or indexing is perfomed. The getobs function is called on the data object to retrieve the observations.\n\nFor BatchView to work on some data structure, the type of the given variable data must implement the data container interface. See ObsView for more info.\n\nArguments\n\ndata : The object describing the dataset. Can be of any type as long as it implements getobs and numobs (see Details for more information).\nbatchsize : The batch-size of each batch. It is the number of observations that each batch must contain (except possibly for the last one).\npartial : If partial=false and the number of observations is not divisible by the batch-size, then the last mini-batch is dropped.\ncollate: Defines the batching behavior. \nIf nothing (default), a batch is getobs(data, indices). \nIf false, each batch is [getobs(data, i) for i in indices]. \nIf true, applies MLUtils to the vector of observations in a batch, recursively collating arrays in the last dimensions. See MLUtils.batch for more information and examples.\nIf a custom function, it will be used in place of MLUtils.batch. It should take a vector of observations as input.\n\nSe also DataLoader.\n\nExamples\n\njulia> using MLUtils\n\njulia> X, Y = MLUtils.load_iris();\n\njulia> A = BatchView(X, batchsize=30);\n\njulia> @assert eltype(A) <: Matrix{Float64}\n\njulia> @assert length(A) == 5 # Iris has 150 observations\n\njulia> @assert size(A[1]) == (4,30) # Iris has 4 features\n\njulia> for x in BatchView(X, batchsize=30)\n # 5 batches of size 30 observations\n @assert size(x) == (4, 30)\n @assert numobs(x) === 30\n end\n\njulia> for (x, y) in BatchView((X, Y), batchsize=20, partial=true)\n # 7 batches of size 20 observations + 1 batch of 10 observations\n @assert typeof(x) <: Matrix{Float64}\n @assert typeof(y) <: Vector{String}\n end\n\njulia> for batch in BatchView((X, Y), batchsize=20, partial=false, collate=false)\n # 7 batches of size 20 observations\n @assert length(batch) == 20\n x1, y1 = batch[1]\n end\n\njulia> function collate_fn(batch)\n # collate observations into a custom batch\n return hcat([x[1] for x in batch]...), join([x[2] for x in batch])\n end;\n\njulia> for (x, y) in BatchView((rand(10, 4), [\"a\", \"b\", \"c\", \"d\"]), batchsize=2, collate=collate_fn)\n @assert size(x) == (10, 2)\n @assert y isa String\n end\n\n\n\n\n\n","category":"type"},{"location":"api/#MLUtils.eachobs","page":"API","title":"MLUtils.eachobs","text":"eachobs(data; kws...)\n\nReturn an iterator over data.\n\nSupports the same arguments as DataLoader. The batchsize default is -1 here while it is 1 for DataLoader.\n\nExamples\n\nX = rand(4,100)\n\nfor x in eachobs(X)\n # loop entered 100 times\n @assert typeof(x) <: Vector{Float64}\n @assert size(x) == (4,)\nend\n\n# mini-batch iterations\nfor x in eachobs(X, batchsize=10)\n # loop entered 10 times\n @assert typeof(x) <: Matrix{Float64}\n @assert size(x) == (4,10)\nend\n\n# support for tuples, named tuples, dicts\nfor (x, y) in eachobs((X, Y))\n # ...\nend\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.DataLoader","page":"API","title":"MLUtils.DataLoader","text":"DataLoader(data; [batchsize, buffer, collate, parallel, partial, rng, shuffle])\n\nAn object that iterates over mini-batches of data, each mini-batch containing batchsize observations (except possibly the last one).\n\nTakes as input a single data array, a tuple (or a named tuple) of arrays, or in general any data object that implements the numobs and getobs methods.\n\nThe last dimension in each array is the observation dimension, i.e. the one divided into mini-batches.\n\nThe original data is preserved in the data field of the DataLoader.\n\nArguments\n\ndata: The data to be iterated over. The data type has to be supported by numobs and getobs.\nbatchsize: If less than 0, iterates over individual observations. Otherwise, each iteration (except possibly the last) yields a mini-batch containing batchsize observations. Default 1.\nbuffer: If buffer=true and supported by the type of data, a buffer will be allocated and reused for memory efficiency. May want to set partial=false to avoid size mismatch. Finally, can pass an external buffer to be used in getobs! (depending on the collate and batchsize options, could be getobs!(buffer, data, idxs) or getobs!(buffer[i], data, idx)). Default false. \ncollate: Defines the batching behavior. Default nothing. \nIf nothing , a batch is getobs(data, indices). \nIf false, each batch is [getobs(data, i) for i in indices]. \nIf true, applies MLUtils.batch to the vector of observations in a batch, recursively collating arrays in the last dimensions. See MLUtils.batch for more information and examples.\nIf a custom function, it will be used in place of MLUtils.batch. It should take a vector of observations as input.\nparallel: Whether to use load data in parallel using worker threads. Greatly speeds up data loading by factor of available threads. Requires starting Julia with multiple threads. Check Threads.nthreads() to see the number of available threads. Passing parallel = true breaks ordering guarantees. Default false.\npartial: This argument is used only when batchsize > 0. If partial=false and the number of observations is not divisible by the batchsize, then the last mini-batch is dropped. Default true.\nrng: A random number generator. Default Random.default_rng().\n`shuffle: Whether to shuffle the observations before iterating. Unlike wrapping the data container with shuffleobs(data), shuffle=true ensures that the observations are shuffled anew every time you start iterating over eachobs. Default false.\n\nExamples\n\njulia> Xtrain = rand(10, 100);\n\njulia> array_loader = DataLoader(Xtrain, batchsize=2);\n\njulia> for x in array_loader\n @assert size(x) == (10, 2)\n # do something with x, 50 times\n end\n\njulia> array_loader.data === Xtrain\ntrue\n\njulia> tuple_loader = DataLoader((Xtrain,), batchsize=2); # similar, but yielding 1-element tuples\n\njulia> for x in tuple_loader\n @assert x isa Tuple{Matrix}\n @assert size(x[1]) == (10, 2)\n end\n\njulia> Ytrain = rand('a':'z', 100); # now make a DataLoader yielding 2-element named tuples\n\njulia> train_loader = DataLoader((data=Xtrain, label=Ytrain), batchsize=5, shuffle=true);\n\njulia> for epoch in 1:100\n for (x, y) in train_loader # access via tuple destructuring\n @assert size(x) == (10, 5)\n @assert size(y) == (5,)\n # loss += f(x, y) # etc, runs 100 * 20 times\n end\n end\n\njulia> first(train_loader).label isa Vector{Char} # access via property name\ntrue\n\njulia> first(train_loader).label == Ytrain[1:5] # because of shuffle=true\nfalse\n\njulia> foreach(println∘summary, DataLoader(rand(Int8, 10, 64), batchsize=30)) # partial=false would omit last\n10×30 Matrix{Int8}\n10×30 Matrix{Int8}\n10×4 Matrix{Int8}\n\njulia> collate_fn(batch) = join(batch);\n\njulia> first(DataLoader([\"a\", \"b\", \"c\", \"d\"], batchsize=2, collate=collate_fn))\n\"ab\"\n\n\n\n\n\n","category":"type"},{"location":"api/#MLUtils.obsview","page":"API","title":"MLUtils.obsview","text":"obsview(data, [indices])\n\nReturns a lazy view of the observations in data that correspond to the given indices. No data will be copied except of the indices. It is similar to constructing an ObsView, but returns a SubArray if the type of data is Array or SubArray. Furthermore, this function may be extended for custom types of data that also want to provide their own subset-type.\n\nIn case data is a tuple, the constructor will be mapped over its elements. That means that the constructor returns a tuple of ObsView instead of a ObsView of tuples.\n\nIf instead you want to get the subset of observations corresponding to the given indices in their native type, use getobs.\n\nSee ObsView for more information.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.ObsView","page":"API","title":"MLUtils.ObsView","text":"ObsView(data, [indices])\n\nUsed to represent a subset of some data of arbitrary type by storing which observation-indices the subset spans. Furthermore, subsequent subsettings are accumulated without needing to access actual data.\n\nThe main purpose for the existence of ObsView is to delay data access and movement until an actual batch of data (or single observation) is needed for some computation. This is particularily useful when the data is not located in memory, but on the hard drive or some remote location. In such a scenario one wants to load the required data only when needed.\n\nAny data access is delayed until getindex is called, and even getindex returns the result of obsview which in general avoids data movement until getobs is called. If used as an iterator, the view will iterate over the dataset once, effectively denoting an epoch. Each iteration will return a lazy subset to the current observation.\n\nArguments\n\ndata : The object describing the dataset. Can be of any type as long as it implements getobs and numobs (see Details for more information).\nindices : Optional. The index or indices of the observation(s) in data that the subset should represent. Can be of type Int or some subtype of AbstractVector.\n\nMethods\n\ngetindex : Returns the observation(s) of the given index/indices. No data is copied aside from the required indices.\nnumobs : Returns the total number observations in the subset.\ngetobs : Returns the underlying data that the ObsView represents at the given relative indices. Note that these indices are in \"subset space\", and in general will not directly correspond to the same indices in the underlying data set.\n\nDetails\n\nFor ObsView to work on some data structure, the desired type MyType must implement the following interface:\n\ngetobs(data::MyType, idx) : Should return the observation(s) indexed by idx. In what form is up to the user. Note that idx can be of type Int or AbstractVector.\nnumobs(data::MyType) : Should return the total number of observations in data\n\nThe following methods can also be provided and are optional:\n\ngetobs(data::MyType) : By default this function is the identity function. If that is not the behaviour that you want for your type, you need to provide this method as well.\nobsview(data::MyType, idx) : If your custom type has its own kind of subset type, you can return it here. An example for such a case are SubArray for representing a subset of some AbstractArray.\ngetobs!(buffer, data::MyType, [idx]) : Inplace version of getobs(data, idx). If this method is provided for MyType, then eachobs can preallocate a buffer that is then reused every iteration. Note: buffer should be equivalent to the return value of getobs(::MyType, ...), since this is how buffer is preallocated by default.\n\nExamples\n\nX, Y = MLUtils.load_iris()\n\n# The iris set has 150 observations and 4 features\n@assert size(X) == (4,150)\n\n# Represents the 80 observations as a ObsView\nv = ObsView(X, 21:100)\n@assert numobs(v) == 80\n@assert typeof(v) <: ObsView\n# getobs indexes into v\n@assert getobs(v, 1:10) == X[:, 21:30]\n\n# Use `obsview` to avoid boxing into ObsView\n# for types that provide a custom \"subset\", such as arrays.\n# Here it instead creates a native SubArray.\nv = obsview(X, 1:100)\n@assert numobs(v) == 100\n@assert typeof(v) <: SubArray\n\n# Also works for tuples of arbitrary length\nsubset = obsview((X, Y), 1:100)\n@assert numobs(subset) == 100\n@assert typeof(subset) <: Tuple # tuple of SubArray\n\n# Use as iterator\nfor x in ObsView(X)\n @assert typeof(x) <: SubArray{Float64,1}\nend\n\n# iterate over each individual labeled observation\nfor (x, y) in ObsView((X, Y))\n @assert typeof(x) <: SubArray{Float64,1}\n @assert typeof(y) <: String\nend\n\n# same but in random order\nfor (x, y) in ObsView(shuffleobs((X, Y)))\n @assert typeof(x) <: SubArray{Float64,1}\n @assert typeof(y) <: String\nend\n\n# Indexing: take first 10 observations\nx, y = ObsView((X, Y))[1:10]\n\nSee also\n\nobsview, getobs, numobs, splitobs, shuffleobs, kfolds.\n\n\n\n\n\n","category":"type"},{"location":"api/#MLUtils.randobs","page":"API","title":"MLUtils.randobs","text":"randobs(data, [n])\n\nPick a random observation or a batch of n random observations from data. For this function to work, the type of data must implement numobs and getobs.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.slidingwindow","page":"API","title":"MLUtils.slidingwindow","text":"slidingwindow(data; size, stride=1) -> SlidingWindow\n\nReturn a vector-like view of the data for which each element is a fixed size \"window\" of size adjacent observations. Note that only complete windows are included in the output, which implies that it is possible for excess observations to be omitted from the view.\n\nNote that the windows are not materialized at construction time. To actually get a copy of the data at some window use indexing or getobs.\n\njulia> s = slidingwindow(1:20, size=6)\nslidingwindow(1:20, size=6, stride=1)\n\njulia> s[1]\n1:6\n\njulia> s[2]\n2:7\n\nThe optional parameter stride can be used to specify the distance between the start elements of each adjacent window. By default the stride is equal to 1.\n\njulia> s = slidingwindow(1:20, size=6, stride=3)\nslidingwindow(1:20, size=6, stride=3)\n\njulia> for w in s; println(w); end\n1:6\n4:9\n7:12\n10:15\n13:18\n\n\n\n\n\n","category":"function"},{"location":"api/#Partitioning","page":"API","title":"Partitioning","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"leavepout\nkfolds\nsplitobs","category":"page"},{"location":"api/#MLUtils.leavepout","page":"API","title":"MLUtils.leavepout","text":"leavepout(n::Integer, [size = 1]) -> Tuple\n\nCompute the train/validation assignments for k ≈ n/size repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. Each validation subset will have either size or size+1 observations assigned to it. The following code snippet generates the index-vectors for size = 2.\n\njulia> train_idx, val_idx = leavepout(10, 2);\n\nEach observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.\n\njulia> train_idx\n5-element Array{Array{Int64,1},1}:\n [3,4,5,6,7,8,9,10]\n [1,2,5,6,7,8,9,10]\n [1,2,3,4,7,8,9,10]\n [1,2,3,4,5,6,9,10]\n [1,2,3,4,5,6,7,8]\n\njulia> val_idx\n5-element Array{UnitRange{Int64},1}:\n 1:2\n 3:4\n 5:6\n 7:8\n 9:10\n\n\n\n\n\nleavepout(data, p = 1)\n\nRepartition a data container using a k-fold strategy, where k is chosen in such a way, that each validation subset of the resulting folds contains roughly p observations. Defaults to p = 1, which is also known as \"leave-one-out\" partitioning.\n\nThe resulting sequence of folds is returned as a lazy iterator. Only data subsets are created. That means no actual data is copied until getobs is invoked.\n\nfor (train, val) in leavepout(X, p=2)\n # if numobs(X) is dividable by 2,\n # then numobs(val) will be 2 for each iteraton,\n # otherwise it may be 3 for the first few iterations.\nend\n\nSeekfolds for a related function.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.kfolds","page":"API","title":"MLUtils.kfolds","text":"kfolds(n::Integer, k = 5) -> Tuple\n\nCompute the train/validation assignments for k repartitions of n observations, and return them in the form of two vectors. The first vector contains the index-vectors for the training subsets, and the second vector the index-vectors for the validation subsets respectively. A general rule of thumb is to use either k = 5 or k = 10. The following code snippet generates the indices assignments for k = 5\n\njulia> train_idx, val_idx = kfolds(10, 5);\n\nEach observation is assigned to the validation subset once (and only once). Thus, a union over all validation index-vectors reproduces the full range 1:n. Note that there is no random assignment of observations to subsets, which means that adjacent observations are likely to be part of the same validation subset.\n\njulia> train_idx\n5-element Array{Array{Int64,1},1}:\n [3,4,5,6,7,8,9,10]\n [1,2,5,6,7,8,9,10]\n [1,2,3,4,7,8,9,10]\n [1,2,3,4,5,6,9,10]\n [1,2,3,4,5,6,7,8]\n\njulia> val_idx\n5-element Array{UnitRange{Int64},1}:\n 1:2\n 3:4\n 5:6\n 7:8\n 9:10\n\n\n\n\n\nkfolds(data, [k = 5])\n\nRepartition a data container k times using a k folds strategy and return the sequence of folds as a lazy iterator. Only data subsets are created, which means that no actual data is copied until getobs is invoked.\n\nConceptually, a k-folds repartitioning strategy divides the given data into k roughly equal-sized parts. Each part will serve as validation set once, while the remaining parts are used for training. This results in k different partitions of data.\n\nIn the case that the size of the dataset is not dividable by the specified k, the remaining observations will be evenly distributed among the parts.\n\nfor (x_train, x_val) in kfolds(X, k=10)\n # code called 10 times\n # numobs(x_val) may differ up to ±1 over iterations\nend\n\nMultiple variables are supported (e.g. for labeled data)\n\nfor ((x_train, y_train), val) in kfolds((X, Y), k=10)\n # ...\nend\n\nBy default the folds are created using static splits. Use shuffleobs to randomly assign observations to the folds.\n\nfor (x_train, x_val) in kfolds(shuffleobs(X), k = 10)\n # ...\nend\n\nSee leavepout for a related function.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.splitobs","page":"API","title":"MLUtils.splitobs","text":"splitobs(n::Int; at) -> Tuple\n\nCompute the indices for two or more disjoint subsets of the range 1:n with split sizes determined by at.\n\nExamples\n\njulia> splitobs(100, at=0.7)\n(1:70, 71:100)\n\njulia> splitobs(100, at=(0.1, 0.4))\n(1:10, 11:50, 51:100)\n\n\n\n\n\nsplitobs([rng,] data; at, shuffle=false, stratified=nothing) -> Tuple\n\nPartition the data into two or more subsets.\n\nThe argument at specifies how to split the data:\n\nWhen at is a number between 0 and 1, this specifies the proportion in the first subset.\nWhen at is an integer, it specifies the number of observations in the first subset.\nWhen at is a tuple, entries specifies the number or proportion in each subset, except\n\nfor the last which will contain the remaning observations. The number of returned subsets is length(at)+1.\n\nIf shuffle=true, randomly permute the observations before splitting. A random number generator rng can be optionally passed as the first argument.\n\nIf stratified is not nothing, it should be an array of labels with the same length as the data. The observations will be split in such a way that the proportion of each label is preserved in each subset.\n\nSupports any datatype implementing numobs. \n\nIt relies on obsview to create views of the data.\n\nExamples\n\njulia> splitobs(reshape(1:100, 1, :); at=0.7) # simple 70%-30% split, of a matrix\n([1 2 … 69 70], [71 72 … 99 100])\n\njulia> data = (x=ones(2,10), n=1:10) # a NamedTuple, consistent last dimension\n(x = [1.0 1.0 … 1.0 1.0; 1.0 1.0 … 1.0 1.0], n = 1:10)\n\njulia> splitobs(data, at=(0.5, 0.3)) # a 50%-30%-20% split, e.g. train/test/validation\n((x = [1.0 1.0 … 1.0 1.0; 1.0 1.0 … 1.0 1.0], n = 1:5), (x = [1.0 1.0 1.0; 1.0 1.0 1.0], n = 6:8), (x = [1.0 1.0; 1.0 1.0], n = 9:10))\n\njulia> train, test = splitobs((reshape(1.0:100.0, 1, :), 101:200), at=0.7, shuffle=true); # split a Tuple\n\njulia> vec(test[1]) .+ 100 == test[2]\ntrue\n\njulia> splitobs(1:10, at=0.5, stratified=[0,0,0,0,1,1,1,1,1,1]) # 2 zeros and 3 ones in each subset\n([1, 2, 5, 6, 7], [3, 4, 8, 9, 10])\n\n\n\n\n\n","category":"function"},{"location":"api/#Array-Constructors","page":"API","title":"Array Constructors","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"falses_like\nfill_like\nones_like\nrand_like\nrandn_like\ntrues_like\nzeros_like","category":"page"},{"location":"api/#MLUtils.falses_like","page":"API","title":"MLUtils.falses_like","text":"falses_like(x, [dims=size(x)])\n\nEquivalent to fill_like(x, false, Bool, dims).\n\nSee also [fill_like] and trues_like.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.fill_like","page":"API","title":"MLUtils.fill_like","text":"fill_like(x, val, [element_type=eltype(x)], [dims=size(x)]))\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to val. The third and fourth arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nSee also zeros_like and ones_like.\n\nExamples\n\njulia> x = rand(Float32, 2)\n2-element Vector{Float32}:\n 0.16087806\n 0.89916044\n\njulia> fill_like(x, 1.7, (3, 3))\n3×3 Matrix{Float32}:\n 1.7 1.7 1.7\n 1.7 1.7 1.7\n 1.7 1.7 1.7\n\njulia> using CUDA\n\njulia> x = CUDA.rand(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 0.803167 0.476101\n 0.303041 0.317581\n\njulia> fill_like(x, 1.7, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 1.7 1.7\n 1.7 1.7\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.ones_like","page":"API","title":"MLUtils.ones_like","text":"ones_like(x, [element_type=eltype(x)], [dims=size(x)]))\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 1. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nSee also zeros_like and fill_like.\n\nExamples\n\njulia> x = rand(Float32, 2)\n2-element Vector{Float32}:\n 0.8621633\n 0.5158395\n\njulia> ones_like(x, (3, 3))\n3×3 Matrix{Float32}:\n 1.0 1.0 1.0\n 1.0 1.0 1.0\n 1.0 1.0 1.0\n\njulia> using CUDA\n\njulia> x = CUDA.rand(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 0.82297 0.656143\n 0.701828 0.391335\n\njulia> ones_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 1.0 1.0\n 1.0 1.0\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.rand_like","page":"API","title":"MLUtils.rand_like","text":"rand_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nThe default random number generator is used, unless a custom one is passed in explicitly as the first argument.\n\nSee also Base.rand and randn_like.\n\nExamples\n\njulia> x = ones(Float32, 2)\n2-element Vector{Float32}:\n 1.0\n 1.0\n\njulia> rand_like(x, (3, 3))\n3×3 Matrix{Float32}:\n 0.780032 0.920552 0.53689\n 0.121451 0.741334 0.5449\n 0.55348 0.138136 0.556404\n\njulia> using CUDA\n\njulia> CUDA.ones(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 1.0 1.0\n 1.0 1.0\n\njulia> rand_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 0.429274 0.135379\n 0.718895 0.0098756\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.randn_like","page":"API","title":"MLUtils.randn_like","text":"randn_like([rng=default_rng()], x, [element_type=eltype(x)], [dims=size(x)])\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to a random value drawn from a normal distribution. The last two arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nThe default random number generator is used, unless a custom one is passed in explicitly as the first argument.\n\nSee also Base.randn and rand_like.\n\nExamples\n\njulia> x = ones(Float32, 2)\n2-element Vector{Float32}:\n 1.0\n 1.0\n\njulia> randn_like(x, (3, 3))\n3×3 Matrix{Float32}:\n -0.385331 0.956231 0.0745102\n 1.43756 -0.967328 2.06311\n 0.0482372 1.78728 -0.902547\n\njulia> using CUDA\n\njulia> CUDA.ones(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 1.0 1.0\n 1.0 1.0\n\njulia> randn_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n -0.578527 0.823445\n -1.01338 -0.612053\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.trues_like","page":"API","title":"MLUtils.trues_like","text":"trues_like(x, [dims=size(x)])\n\nEquivalent to fill_like(x, true, Bool, dims).\n\nSee also [fill_like] and falses_like.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.zeros_like","page":"API","title":"MLUtils.zeros_like","text":"zeros_like(x, [element_type=eltype(x)], [dims=size(x)]))\n\nCreate an array with the given element type and size, based upon the given source array x. All element of the new array will be set to 0. The second and third arguments are both optional, defaulting to the given array's eltype and size. The dimensions may be specified as an integer or as a tuple argument.\n\nSee also ones_like and fill_like.\n\nExamples\n\njulia> x = rand(Float32, 2)\n2-element Vector{Float32}:\n 0.4005432\n 0.36934233\n\njulia> zeros_like(x, (3, 3))\n3×3 Matrix{Float32}:\n 0.0 0.0 0.0\n 0.0 0.0 0.0\n 0.0 0.0 0.0\n\njulia> using CUDA\n\njulia> x = CUDA.rand(2, 2)\n2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:\n 0.0695155 0.667979\n 0.558468 0.59903\n\njulia> zeros_like(x, Float64)\n2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:\n 0.0 0.0\n 0.0 0.0\n\n\n\n\n\n","category":"function"},{"location":"api/#Resampling","page":"API","title":"Resampling","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"oversample\nundersample","category":"page"},{"location":"api/#MLUtils.oversample","page":"API","title":"MLUtils.oversample","text":"oversample([rng], data, classes; fraction=1, shuffle=true)\noversample([rng], data::Tuple; fraction=1, shuffle=true)\n\nGenerate a re-balanced version of data by repeatedly sampling existing observations in such a way that every class will have at least fraction times the number observations of the largest class in classes. This way, all classes will have a minimum number of observations in the resulting data set relative to what largest class has in the given (original) data.\n\nAs an example, by default (i.e. with fraction = 1) the resulting dataset will be near perfectly balanced. On the other hand, with fraction = 0.5 every class in the resulting data with have at least 50% as many observations as the largest class.\n\nThe classes input is an array with the same length as numobs(data). \n\nThe convenience parameter shuffle determines if the resulting data will be shuffled after its creation; if it is not shuffled then all the repeated samples will be together at the end, sorted by class. Defaults to true.\n\nThe random number generator rng can be optionally passed as the first argument. \n\nThe output will contain both the resampled data and classes.\n\n# 6 observations with 3 features each\nX = rand(3, 6)\n# 2 classes, severely imbalanced\nY = [\"a\", \"b\", \"b\", \"b\", \"b\", \"a\"]\n\n# oversample the class \"a\" to match \"b\"\nX_bal, Y_bal = oversample(X, Y)\n\n# this results in a bigger dataset with repeated data\n@assert size(X_bal) == (3,8)\n@assert length(Y_bal) == 8\n\n# now both \"a\", and \"b\" have 4 observations each\n@assert sum(Y_bal .== \"a\") == 4\n@assert sum(Y_bal .== \"b\") == 4\n\nFor this function to work, the type of data must implement numobs and getobs. \n\nIf data is a tuple and classes is not given, then it will be assumed that the last element of the tuple contains the classes.\n\njulia> data = DataFrame(X1=rand(6), X2=rand(6), Y=[:a,:b,:b,:b,:b,:a])\n6×3 DataFrames.DataFrame\n│ Row │ X1 │ X2 │ Y │\n├─────┼───────────┼─────────────┼───┤\n│ 1 │ 0.226582 │ 0.0443222 │ a │\n│ 2 │ 0.504629 │ 0.722906 │ b │\n│ 3 │ 0.933372 │ 0.812814 │ b │\n│ 4 │ 0.522172 │ 0.245457 │ b │\n│ 5 │ 0.505208 │ 0.11202 │ b │\n│ 6 │ 0.0997825 │ 0.000341996 │ a │\n\njulia> getobs(oversample(data, data.Y))\n8×3 DataFrame\n Row │ X1 X2 Y \n │ Float64 Float64 Symbol \n─────┼─────────────────────────────\n 1 │ 0.376304 0.100022 a\n 2 │ 0.467095 0.185437 b\n 3 │ 0.481957 0.319906 b\n 4 │ 0.336762 0.390811 b\n 5 │ 0.376304 0.100022 a\n 6 │ 0.427064 0.0648339 a\n 7 │ 0.427064 0.0648339 a\n 8 │ 0.457043 0.490688 b\n\nSee ObsView for more information on data subsets. See also undersample.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.undersample","page":"API","title":"MLUtils.undersample","text":"undersample([rng], data, classes; shuffle=true)\nundersample([rng], data::Tuple; shuffle=true)\n\nGenerate a class-balanced version of data by subsampling its observations in such a way that the resulting number of observations will be the same number for every class. This way, all classes will have as many observations in the resulting data set as the smallest class has in the given (original) data.\n\nThe convenience parameter shuffle determines if the resulting data will be shuffled after its creation; if it is not shuffled then all the observations will be in their original order. Defaults to false.\n\nIf data is a tuple and classes is not given, then it will be assumed that the last element of the tuple contains the classes.\n\nThe output will contain both the resampled data and classes.\n\n# 6 observations with 3 features each\nX = rand(3, 6)\n# 2 classes, severely imbalanced\nY = [\"a\", \"b\", \"b\", \"b\", \"b\", \"a\"]\n\n# subsample the class \"b\" to match \"a\"\nX_bal, Y_bal = undersample(X, Y)\n\n# this results in a smaller dataset\n@assert size(X_bal) == (3,4)\n@assert length(Y_bal) == 4\n\n# now both \"a\", and \"b\" have 2 observations each\n@assert sum(Y_bal .== \"a\") == 2\n@assert sum(Y_bal .== \"b\") == 2\n\nFor this function to work, the type of data must implement numobs and getobs. \n\nNote that if data is a tuple, then it will be assumed that the last element of the tuple contains the targets.\n\njulia> data = DataFrame(X1=rand(6), X2=rand(6), Y=[:a,:b,:b,:b,:b,:a])\n6×3 DataFrames.DataFrame\n│ Row │ X1 │ X2 │ Y │\n├─────┼───────────┼─────────────┼───┤\n│ 1 │ 0.226582 │ 0.0443222 │ a │\n│ 2 │ 0.504629 │ 0.722906 │ b │\n│ 3 │ 0.933372 │ 0.812814 │ b │\n│ 4 │ 0.522172 │ 0.245457 │ b │\n│ 5 │ 0.505208 │ 0.11202 │ b │\n│ 6 │ 0.0997825 │ 0.000341996 │ a │\n\njulia> getobs(undersample(data, data.Y))\n4×3 DataFrame\n Row │ X1 X2 Y \n │ Float64 Float64 Symbol \n─────┼─────────────────────────────\n 1 │ 0.427064 0.0648339 a\n 2 │ 0.376304 0.100022 a\n 3 │ 0.467095 0.185437 b\n 4 │ 0.457043 0.490688 b\n\nSee ObsView for more information on data subsets. See also oversample.\n\n\n\n\n\n","category":"function"},{"location":"api/#Operations","page":"API","title":"Operations","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"chunk\nflatten\ngroup_counts\ngroup_indices\nnormalise\nrpad_constant\nunbatch\nunsqueeze\nunstack","category":"page"},{"location":"api/#MLUtils.chunk","page":"API","title":"MLUtils.chunk","text":"chunk(x, n; [dims])\nchunk(x; [size, dims])\n\nSplit x into n parts or alternatively, if size is an integer, into equal chunks of size size. The parts contain the same number of elements except possibly for the last one that can be smaller.\n\nIn case size is a collection of integers instead, the elements of x are split into chunks of the given sizes.\n\nIf x is an array, dims can be used to specify along which dimension to split (defaults to the last dimension).\n\nExamples\n\njulia> chunk(1:10, 3)\n3-element Vector{UnitRange{Int64}}:\n 1:4\n 5:8\n 9:10\n\njulia> chunk(1:10; size = 2)\n5-element Vector{UnitRange{Int64}}:\n 1:2\n 3:4\n 5:6\n 7:8\n 9:10\n\njulia> x = reshape(collect(1:20), (5, 4))\n5×4 Matrix{Int64}:\n 1 6 11 16\n 2 7 12 17\n 3 8 13 18\n 4 9 14 19\n 5 10 15 20\n\njulia> xs = chunk(x, 2, dims=1)\n2-element Vector{SubArray{Int64, 2, Matrix{Int64}, Tuple{UnitRange{Int64}, Base.Slice{Base.OneTo{Int64}}}, false}}:\n [1 6 11 16; 2 7 12 17; 3 8 13 18]\n [4 9 14 19; 5 10 15 20]\n\njulia> xs[1]\n3×4 view(::Matrix{Int64}, 1:3, :) with eltype Int64:\n 1 6 11 16\n 2 7 12 17\n 3 8 13 18\n\njulia> xes = chunk(x; size = 2, dims = 2)\n2-element Vector{SubArray{Int64, 2, Matrix{Int64}, Tuple{Base.Slice{Base.OneTo{Int64}}, UnitRange{Int64}}, true}}:\n [1 6; 2 7; … ; 4 9; 5 10]\n [11 16; 12 17; … ; 14 19; 15 20]\n\njulia> xes[2]\n5×2 view(::Matrix{Int64}, :, 3:4) with eltype Int64:\n 11 16\n 12 17\n 13 18\n 14 19\n 15 20\n\njulia> chunk(1:6; size = [2, 4])\n2-element Vector{UnitRange{Int64}}:\n 1:2\n 3:6\n\n\n\n\n\nchunk(x, partition_idxs; [npartitions, dims])\n\nPartition the array x along the dimension dims according to the indexes in partition_idxs.\n\npartition_idxs must be sorted and contain only positive integers between 1 and the number of partitions. \n\nIf the number of partition npartitions is not provided, it is inferred from partition_idxs.\n\nIf dims is not provided, it defaults to the last dimension.\n\nSee also unbatch.\n\nExamples\n\njulia> x = reshape([1:10;], 2, 5)\n2×5 Matrix{Int64}:\n 1 3 5 7 9\n 2 4 6 8 10\n\njulia> chunk(x, [1, 2, 2, 3, 3])\n3-element Vector{SubArray{Int64, 2, Matrix{Int64}, Tuple{Base.Slice{Base.OneTo{Int64}}, UnitRange{Int64}}, true}}:\n [1; 2;;]\n [3 5; 4 6]\n [7 9; 8 10]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.flatten","page":"API","title":"MLUtils.flatten","text":"flatten(x::AbstractArray)\n\nReshape arbitrarly-shaped input into a matrix-shaped output, preserving the size of the last dimension.\n\nSee also unsqueeze.\n\nExamples\n\njulia> rand(3,4,5) |> flatten |> size\n(12, 5)\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.group_counts","page":"API","title":"MLUtils.group_counts","text":"group_counts(x)\n\nCount the number of times that each element of x appears.\n\nSee also group_indices\n\nExamples\n\njulia> group_counts(['a', 'b', 'b'])\nDict{Char, Int64} with 2 entries:\n 'a' => 1\n 'b' => 2\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.group_indices","page":"API","title":"MLUtils.group_indices","text":"group_indices(x) -> Dict\n\nComputes the indices of elements in the vector x for each distinct value contained. This information is useful for resampling strategies, such as stratified sampling.\n\nSee also group_counts.\n\nExamples\n\njulia> x = [:yes, :no, :maybe, :yes];\n\njulia> group_indices(x)\nDict{Symbol, Vector{Int64}} with 3 entries:\n :yes => [1, 4]\n :maybe => [3]\n :no => [2]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.normalise","page":"API","title":"MLUtils.normalise","text":"normalise(x; dims=ndims(x), ϵ=1e-5)\n\nNormalise the array x to mean 0 and standard deviation 1 across the dimension(s) given by dims. Per default, dims is the last dimension. \n\nϵ is a small additive factor added to the denominator for numerical stability.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.rpad_constant","page":"API","title":"MLUtils.rpad_constant","text":"rpad_constant(v::AbstractArray, n::Union{Integer, Tuple}, val = 0; dims=:)\n\nReturn the given sequence padded with val along the dimensions dims up to a maximum length in each direction specified by n.\n\nExamples\n\njulia> rpad_constant([1, 2], 4, -1) # passing with -1 up to size 4\n4-element Vector{Int64}:\n 1\n 2\n -1\n -1\n\njulia> rpad_constant([1, 2, 3], 2) # no padding if length is already greater than n\n3-element Vector{Int64}:\n 1\n 2\n 3\n\njulia> rpad_constant([1 2; 3 4], 4; dims=1) # padding along the first dimension\n4×2 Matrix{Int64}:\n 1 2\n 3 4\n 0 0\n 0 0\n\njulia> rpad_constant([1 2; 3 4], 4) # padding along all dimensions by default\n4×4 Matrix{Int64}:\n 1 2 0 0\n 3 4 0 0\n 0 0 0 0\n 0 0 0 0\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.unbatch","page":"API","title":"MLUtils.unbatch","text":"unbatch(x)\n\nReverse of the batch operation, unstacking the last dimension of the array x.\n\nSee also unstack and chunk.\n\nExamples\n\njulia> unbatch([1 3 5 7;\n 2 4 6 8])\n4-element Vector{Vector{Int64}}:\n [1, 2]\n [3, 4]\n [5, 6]\n [7, 8]\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.unsqueeze","page":"API","title":"MLUtils.unsqueeze","text":"unsqueeze(x; dims)\n\nReturn x reshaped into an array one dimensionality higher than x, where dims indicates in which dimension x is extended. dims can be an integer between 1 and ndims(x)+1.\n\nSee also flatten, stack.\n\nExamples\n\njulia> unsqueeze([1 2; 3 4], dims=2)\n2×1×2 Array{Int64, 3}:\n[:, :, 1] =\n 1\n 3\n\n[:, :, 2] =\n 2\n 4\n\n\njulia> xs = [[1, 2], [3, 4], [5, 6]]\n3-element Vector{Vector{Int64}}:\n [1, 2]\n [3, 4]\n [5, 6]\n\njulia> unsqueeze(xs, dims=1)\n1×3 Matrix{Vector{Int64}}:\n [1, 2] [3, 4] [5, 6]\n\n\n\n\n\nunsqueeze(; dims)\n\nReturns a function which, acting on an array, inserts a dimension of size 1 at dims.\n\nExamples\n\njulia> rand(21, 22, 23) |> unsqueeze(dims=2) |> size\n(21, 1, 22, 23)\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.unstack","page":"API","title":"MLUtils.unstack","text":"unstack(xs; dims)\n\nUnroll the given xs into an array of arrays along the given dimension dims.\n\nIt is the inverse operation of stack.\n\nSee also unbatch and chunk.\n\nExamples\n\njulia> unstack([1 3 5 7; 2 4 6 8], dims=2)\n4-element Vector{Vector{Int64}}:\n [1, 2]\n [3, 4]\n [5, 6]\n [7, 8]\n\n\n\n\n\n","category":"function"},{"location":"api/#Datasets","page":"API","title":"Datasets","text":"","category":"section"},{"location":"api/","page":"API","title":"API","text":"Datasets.load_iris\nDatasets.make_sin\nDatasets.make_spiral\nDatasets.make_poly\nDatasets.make_moons","category":"page"},{"location":"api/#MLUtils.Datasets.load_iris","page":"API","title":"MLUtils.Datasets.load_iris","text":"load_iris() -> X, y, names\n\nLoads the first 150 observations from the Iris flower data set introduced by Ronald Fisher (1936). The 4 by 150 matrix X contains the numeric measurements, in which each individual column denotes an observation. The vector y contains the class labels as strings. The vector names contains the names of the features (i.e. rows of X)\n\n[1] Fisher, Ronald A. \"The use of multiple measurements in taxonomic problems.\" Annals of eugenics 7.2 (1936): 179-188.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_sin","page":"API","title":"MLUtils.Datasets.make_sin","text":"make_sin(n, start, stop; noise = 0.3, f_rand = randn) -> x, y\n\nGenerates n noisy equally spaces samples of a sinus from start to stop by adding noise .* f_rand(length(x)) to the result of sin(x).\n\nReturns the vector x with the samples and the noisy response y.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_spiral","page":"API","title":"MLUtils.Datasets.make_spiral","text":"make_spiral(n, a, theta, b; noise = 0.01, f_rand = randn) -> x, y\n\nGenerates n noisy responses for a spiral with two labels. Uses the radius, angle and scaling arguments to space the points in 2D space and adding noise .* f_randn(n) to the response.\n\nReturns the 2 x n matrix x with the coordinates of the samples and the vector y with the labels.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_poly","page":"API","title":"MLUtils.Datasets.make_poly","text":"make_poly(coef, x; noise = 0.01, f_rand = randn) -> x, y\n\nGenerates a noisy response for a polynomial of degree length(coef) and with the coefficients given by coef. The response is generated by elmentwise computation of the polynome on the elements of x and adding noise .* f_randn(length(x)) to the result.\n\nThe vector coef contains the coefficients for the terms of the polynome. The first element of coef denotes the coefficient for the term with the highest degree, while the last element of coef denotes the intercept.\n\nReturn the input x and the noisy response y.\n\n\n\n\n\n","category":"function"},{"location":"api/#MLUtils.Datasets.make_moons","page":"API","title":"MLUtils.Datasets.make_moons","text":"make_moons(n; noise=0.0, f_rand=randn, shuffle=true) -> x, y\n\nGenerate a dataset with two interleaving half circles. \n\nIf n is an integer, the number of samples is n and the number of samples for each half circle is n ÷ 2. If n is a tuple, the first element of the tuple denotes the number of samples for the first half circle and the second element denotes the number of samples for the second half circle.\n\nThe noise level can be controlled by the noise argument.\n\nSet shuffle=false to keep the order of the samples.\n\nReturns a 2 x n matrix with the the samples. \n\n\n\n\n\n","category":"function"},{"location":"#MLUtils.jl","page":"Home","title":"MLUtils.jl","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"(Image: ) (Image: ) (Image: ) (Image: )","category":"page"},{"location":"","page":"Home","title":"Home","text":"MLUtils.jl defines interfaces and implements common utilities for Machine Learning pipelines.","category":"page"},{"location":"#Features","page":"Home","title":"Features","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"An extensible dataset interface (numobs and getobs).\nData iteration and dataloaders (eachobs and DataLoader).\nLazy data views (obsview). \nResampling procedures (undersample and oversample).\nTrain/test splits (splitobs) \nData partitioning and aggregation tools (batch, unbatch, chunk, group_counts, group_indices).\nFolds for cross-validation (kfolds, leavepout).\nDatasets lazy tranformations (mapobs, filterobs, groupobs, joinobs, shuffleobs).\nToy datasets for demonstration purpose. \nOther data handling utilities (flatten, normalise, unsqueeze, stack, unstack).","category":"page"},{"location":"#Examples","page":"Home","title":"Examples","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"Let us take a look at a hello world example to get a feeling for how to use this package in a typical ML scenario. ","category":"page"},{"location":"","page":"Home","title":"Home","text":"using MLUtils\n\n# X is a matrix of floats\n# Y is a vector of strings\nX, Y = load_iris()\n\n# The iris dataset is ordered according to their labels,\n# which means that we should shuffle the dataset before\n# partitioning it into training- and test-set.\nXs, Ys = shuffleobs((X, Y))\n\n# We leave out 15 % of the data for testing\ncv_data, test_data = splitobs((Xs, Ys); at=0.85)\n\n# Next we partition the data using a 10-fold scheme.\nfor (train_data, val_data) in kfolds(cv_data; k=10)\n\n # We apply a lazy transform for data augmentation\n train_data = mapobs(xy -> (xy[1] .+ 0.1 .* randn.(), xy[2]), train_data)\n\n for epoch = 1:10\n # Iterate over the data using mini-batches of 5 observations each\n for (x, y) in eachobs(train_data, batchsize=5)\n # ... train supervised model on minibatches here\n end\n end\nend","category":"page"},{"location":"","page":"Home","title":"Home","text":"In the above code snippet, the inner loop for eachobs is the only place where data other than indices is actually being copied. In fact, while x and y are materialized arrays, all the rest are data views. ","category":"page"},{"location":"#Historical-Notes","page":"Home","title":"Historical Notes","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"MLUtils.jl brings together functionalities previously found in LearnBase.jl , MLDataPattern.jl and MLLabelUtils.jl. These packages are now discontinued. ","category":"page"},{"location":"","page":"Home","title":"Home","text":"Other features were ported from the deep learning library Flux.jl, as they are of general use. ","category":"page"},{"location":"#Alternatives-and-Related-Packages","page":"Home","title":"Alternatives and Related Packages","text":"","category":"section"},{"location":"","page":"Home","title":"Home","text":"MLJ.jl is a more complete package for managing the whole machine learning pipeline if you are looking for a sklearn replacement.\nNNlib.jl provides utility functions for neural networks.\nTableTransforms.jl contains transformations for tabular datasets.\nDataAugmentation.jl. Efficient, composable data augmentation for machine and deep learning with support for n-dimensional images, keypoints and categorical masks.","category":"page"}] }