AutoMLPipeline
| Documentation | Build Status | Help |
|---|---|---|
  |
  |
  |
Star History
AutoMLPipeline (AMLP) is a package that makes it trivial to create complex ML pipeline structures using simple expressions. It leverages on the built-in macro programming features of Julia to symbolically process, manipulate pipeline expressions, and makes it easy to discover optimal structures for machine learning regression and classification.
To illustrate, here is a pipeline expression
and evaluation of a typical machine learning
workflow that extracts numerical features (numf)
for ica (Independent Component Analysis)
and pca (Principal Component Analysis)
transformations, respectively, concatenated with
the hot-bit encoding (ohe) of categorical
features (catf) of a given data for rf (Random Forest) modeling:
model = (catf |> ohe) + (numf |> pca) + (numf |> ica) |> rf
fit!(model,Xtrain,Ytrain)
prediction = transform!(model,Xtest)
score(:accuracy,prediction,Ytest)
crossvalidate(model,X,Y,"balanced_accuracy_score")Just take note that + has higher priority than |> so if you
are not sure, enclose the operations inside parentheses.
### these two expressions are the same
a |> b + c; a |> (b + c)
### these two expressions are the same
a + b |> c; (a + b) |> cPlease read this AutoMLPipeline Paper for benchmark comparisons.
- JuliaCon Proceedings:
Recorded Video/Conference Presentations:
- 2021 JuliaCon (Finding an Effective Strategy for AutoML Pipeline Optimization)
- 2020 JuliaCon (AutoMLPipeline: A ToolBox for Building ML Pipelines)
- 2021 PyData Ireland Meetup (Symbolic ML Pipeline Expression and Benchmarking)
Related Video/Conference Presentations:
- 2021 JuliaCon (Lale in Julia: A package for semi-automated data science)
- 2019 JuliaCon (TSML: Time Series Machine Learning Pipeline)
- 2021 OpenSource Guild in IBM (Overview of HPC and Data Science in Julia Programming with AutoML)
More examples can be found in the examples folder including optimizing pipelines by multi-threading or distributed computing.
Motivations
The typical workflow in machine learning classification or prediction requires some or combination of the following preprocessing steps together with modeling:
- feature extraction (e.g. ica, pca, svd)
- feature transformation (e.g. normalization, scaling, ohe)
- feature selection (anova, correlation)
- modeling (rf, adaboost, xgboost, lm, svm, mlp)
Each step has several choices of functions to use together with their corresponding parameters. Optimizing the performance of the entire pipeline is a combinatorial search of the proper order and combination of preprocessing steps, optimization of their corresponding parameters, together with searching for the optimal model and its hyper-parameters.
Because of close dependencies among various steps, we can consider the entire process to be a pipeline optimization problem (POP). POP requires simultaneous optimization of pipeline structure and parameter adaptation of its elements. As a consequence, having an elegant way to express pipeline structure can help lessen the complexity in the management and analysis of the wide-array of choices of optimization routines.
The target of future work will be the implementations of different pipeline optimization algorithms ranging from evolutionary approaches, integer programming (discrete choices of POP elements), tree/graph search, and hyper-parameter search.
Package Features
- Symbolic pipeline API for easy expression and high-level description of complex pipeline structures and processing workflow
- Common API wrappers for ML libs including Scikitlearn, DecisionTree, etc
- Easily extensible architecture by overloading just two main interfaces: fit! and transform!
- Meta-ensembles that allow composition of ensembles of ensembles (recursively if needed) for robust prediction routines
- Categorical and numerical feature selectors for specialized preprocessing routines based on types
Installation
AutoMLPipeline is in the Julia Official package registry.
The latest release can be installed at the Julia
prompt using Julia's package management which is triggered
by pressing ] at the julia prompt:
julia> ]
pkg> update
pkg> add AutoMLPipelineSample Usage
Below outlines some typical way to preprocess and model any dataset.
1. Load Data, Extract Input (X) and Target (Y)
# Make sure that the input feature is a dataframe and the target output is a 1-D vector.
using AutoMLPipeline
profbdata = getprofb()
X = profbdata[:,2:end]
Y = profbdata[:,1] |> Vector;
head(x)=first(x,5)
head(profbdata)5Γ7 DataFrame. Omitted printing of 1 columns
β Row β Home.Away β Favorite_Points β Underdog_Points β Pointspread β Favorite_Name β Underdog_name β
β β String β Int64 β Int64 β Float64 β String β String β
βββββββΌββββββββββββΌββββββββββββββββββΌββββββββββββββββββΌββββββββββββββΌββββββββββββββββΌββββββββββββββββ€
β 1 β away β 27 β 24 β 4.0 β BUF β MIA β
β 2 β at_home β 17 β 14 β 3.0 β CHI β CIN β
β 3 β away β 51 β 0 β 2.5 β CLE β PIT β
β 4 β at_home β 28 β 0 β 5.5 β NO β DAL β
β 5 β at_home β 38 β 7 β 5.5 β MIN β HOU β2. Load Filters, Transformers, and Learners
using AutoMLPipeline
#### Decomposition
pca = skoperator("PCA")
fa = skoperator("FactorAnalysis")
ica = skoperator("FastICA")
#### Scaler
rb = skoperator("RobustScaler")
pt = skoperator("PowerTransformer")
norm = skoperator("Normalizer")
mx = skoperator("MinMaxScaler")
std = skoperator("StandardScaler")
#### categorical preprocessing
ohe = OneHotEncoder()
#### Column selector
catf = CatFeatureSelector()
numf = NumFeatureSelector()
disc = CatNumDiscriminator()
#### Learners
rf = skoperator("RandomForestClassifier")
gb = skoperator("GradientBoostingClassifier")
lsvc = skoperator("LinearSVC")
svc = skoperator("SVC")
mlp = skoperator("MLPClassifier")
ada = skoperator("AdaBoostClassifier")
sgd = skoperator("SGDClassifier")
skrf_reg = skoperator("RandomForestRegressor")
skgb_reg = skoperator("GradientBoostingRegressor")
jrf = RandomForest()
tree = PrunedTree()
vote = VoteEnsemble()
stack = StackEnsemble()
best = BestLearner()Note: You can get a listing of available Preprocessors and Learners by invoking the function:
skoperator()
3. Filter categories and hot-encode them
pohe = catf |> ohe
tr = fit_transform!(pohe,X,Y)
head(tr)5Γ56 DataFrame. Omitted printing of 47 columns
β Row β x1 β x2 β x3 β x4 β x5 β x6 β x7 β x8 β x9 β
β β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β
βββββββΌββββββββββΌββββββββββΌββββββββββΌββββββββββΌββββββββββΌββββββββββΌββββββββββΌββββββββββΌββββββββββ€
β 1 β 1.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β
β 2 β 0.0 β 1.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β
β 3 β 0.0 β 0.0 β 1.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β
β 4 β 0.0 β 0.0 β 0.0 β 1.0 β 0.0 β 0.0 β 0.0 β 0.0 β 0.0 β
β 5 β 0.0 β 0.0 β 0.0 β 0.0 β 1.0 β 0.0 β 0.0 β 0.0 β 0.0 β4. Numerical Feature Extraction Example
4.1 Filter numeric features, compute ica and pca features, and combine both features
pdec = (numf |> pca) + (numf |> ica)
tr = fit_transform!(pdec,X,Y)
head(tr)5Γ8 DataFrame
β Row β x1 β x2 β x3 β x4 β x1_1 β x2_1 β x3_1 β x4_1 β
β β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β
βββββββΌβββββββββββΌβββββββββββΌβββββββββββΌβββββββββββΌβββββββββββββΌβββββββββββββΌβββββββββββββΌβββββββββββββ€
β 1 β 2.47477 β 7.87074 β -1.10495 β 0.902431 β 0.0168432 β 0.00319873 β -0.0467633 β 0.026742 β
β 2 β -5.47113 β -3.82946 β -2.08342 β 1.00524 β -0.0327947 β -0.0217808 β -0.0451314 β 0.00702006 β
β 3 β 30.4068 β -10.8073 β -6.12339 β 0.883938 β -0.0734292 β 0.115776 β -0.0425357 β 0.0497831 β
β 4 β 8.18372 β -15.507 β -1.43203 β 1.08255 β -0.0656664 β 0.0368666 β -0.0457154 β -0.0192752 β
β 5 β 16.6176 β -6.68636 β -1.66597 β 0.978243 β -0.0338749 β 0.0643065 β -0.0461703 β 0.00671696 β4.2 Filter numeric features, transform to robust and power transform scaling, perform ica and pca, respectively, and combine both
ppt = (numf |> rb |> ica) + (numf |> pt |> pca)
tr = fit_transform!(ppt,X,Y)
head(tr)5Γ8 DataFrame
β Row β x1 β x2 β x3 β x4 β x1_1 β x2_1 β x3_1 β x4_1 β
β β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β Float64 β
βββββββΌββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββββββΌββββββββββββΌβββββββββββΌβββββββββββββΌββββββββββββ€
β 1 β -0.00308891 β -0.0269009 β -0.0166298 β 0.0467559 β -0.64552 β 1.40289 β -0.0284468 β 0.111773 β
β 2 β 0.0217799 β -0.00699717 β 0.0329868 β 0.0449952 β -0.832404 β 0.475629 β -1.14881 β -0.01702 β
β 3 β -0.115577 β -0.0503802 β 0.0736173 β 0.0420466 β 1.54491 β 1.65258 β -1.35967 β -2.57866 β
β 4 β -0.0370057 β 0.0190459 β 0.065814 β 0.0454864 β 1.32065 β 0.563565 β -2.05839 β -0.74898 β
β 5 β -0.0643088 β -0.00711682 β 0.0340452 β 0.0459816 β 1.1223 β 1.45555 β -0.88864 β -0.776195 β5. A Pipeline for the Voting Ensemble Classification
# take all categorical columns and hot-bit encode each,
# concatenate them to the numerical features,
# and feed them to the voting ensemble
using AutoMLPipeline.Utils
pvote = (catf |> ohe) + (numf) |> vote
pred = fit_transform!(pvote,X,Y)
sc=score(:accuracy,pred,Y)
println(sc)
crossvalidate(pvote,X,Y,"accuracy_score")fold: 1, 0.5373134328358209
fold: 2, 0.7014925373134329
fold: 3, 0.5294117647058824
fold: 4, 0.6716417910447762
fold: 5, 0.6716417910447762
fold: 6, 0.6119402985074627
fold: 7, 0.5074626865671642
fold: 8, 0.6323529411764706
fold: 9, 0.6268656716417911
fold: 10, 0.5671641791044776
errors: 0
(mean = 0.6057287093942055, std = 0.06724940684190235, folds = 10, errors = 0)Note: crossvalidate() supports the following sklearn's performance metric
classification:
accuracy_score,balanced_accuracy_score,cohen_kappa_scorejaccard_score,matthews_corrcoef,hamming_loss,zero_one_lossf1_score,precision_score,recall_score,
regression:
mean_squared_error,mean_squared_log_errormean_absolute_error,median_absolute_errorr2_score,max_error,mean_poisson_deviancemean_gamma_deviance,mean_tweedie_deviance,explained_variance_score
6. Use @pipelinex instead of @pipeline to print the corresponding function calls in 6
julia> @pipelinex (catf |> ohe) + (numf) |> vote
:(Pipeline(ComboPipeline(Pipeline(catf, ohe), numf), vote))
# another way is to use @macroexpand with @pipeline
julia> @macroexpand @pipeline (catf |> ohe) + (numf) |> vote
:(Pipeline(ComboPipeline(Pipeline(catf, ohe), numf), vote))7. A Pipeline for the Random Forest (RF) Classification
# compute the pca, ica, fa of the numerical columns,
# combine them with the hot-bit encoded categorical features
# and feed all to the random forest classifier
prf = (numf |> rb |> pca) + (numf |> rb |> ica) + (numf |> rb |> fa) + (catf |> ohe) |> rf
pred = fit_transform!(prf,X,Y)
score(:accuracy,pred,Y) |> println
crossvalidate(prf,X,Y,"accuracy_score")fold: 1, 0.6119402985074627
fold: 2, 0.7611940298507462
fold: 3, 0.6764705882352942
fold: 4, 0.6716417910447762
fold: 5, 0.6716417910447762
fold: 6, 0.6567164179104478
fold: 7, 0.6268656716417911
fold: 8, 0.7058823529411765
fold: 9, 0.6417910447761194
fold: 10, 0.6865671641791045
errors: 0
(mean = 0.6710711150131694, std = 0.04231869797446545, folds = 10, errors = 0)8. A Pipeline for the Linear Support Vector for Classification (LSVC)
plsvc = ((numf |> rb |> pca)+(numf |> rb |> fa)+(numf |> rb |> ica)+(catf |> ohe )) |> lsvc
pred = fit_transform!(plsvc,X,Y)
score(:accuracy,pred,Y) |> println
crossvalidate(plsvc,X,Y,"accuracy_score")fold: 1, 0.6567164179104478
fold: 2, 0.7164179104477612
fold: 3, 0.8235294117647058
fold: 4, 0.7164179104477612
fold: 5, 0.7313432835820896
fold: 6, 0.6567164179104478
fold: 7, 0.7164179104477612
fold: 8, 0.7352941176470589
fold: 9, 0.746268656716418
fold: 10, 0.6865671641791045
errors: 0
(mean = 0.7185689201053556, std = 0.04820829087095355, folds = 10, errors = 0)9. A Pipeline for Random Forest Regression
iris = getiris()
Xreg = iris[:,1:3]
Yreg = iris[:,4] |> Vector
pskrfreg = (catf |> ohe) + (numf) |> skrf_reg
res=crossvalidate(pskrfreg,Xreg,Yreg,"mean_absolute_error",10)fold: 1, 0.1827433333333334
fold: 2, 0.18350888888888886
fold: 3, 0.11627222222222248
fold: 4, 0.1254152380952376
fold: 5, 0.16502333333333377
fold: 6, 0.10900222222222226
fold: 7, 0.12561111111111076
fold: 8, 0.14243000000000025
fold: 9, 0.12130555555555576
fold: 10, 0.18811111111111098
errors: 0
(mean = 0.1459423015873016, std = 0.030924217263958102, folds = 10, errors = 0)Note: More examples can be found in the test directory of the package. Since the code is written in Julia, you are highly encouraged to read the source code and feel free to extend or adapt the package to your problem. Please feel free to submit PRs to improve the package features.
10. Performance Comparison of Several Learners
10.1 Sequential Processing
using Random
using DataFrames
Random.seed!(1)
jrf = RandomForest()
tree = PrunedTree()
disc = CatNumDiscriminator()
ada = skoperator("AdaBoostClassifier")
sgd = skoperator("SGDClassifier")
std = skoperator("StandardScaler")
lsvc = skoperator("LinearSVC")
learners = DataFrame()
for learner in [jrf,ada,sgd,tree,lsvc]
pcmc = @pipeline disc |> ((catf |> ohe) + (numf |> std)) |> learner
println(learner.name[1:end-4])
mean,sd,_ = crossvalidate(pcmc,X,Y,"accuracy_score",10)
global learners = vcat(learners,DataFrame(name=learner.name[1:end-4],mean=mean,sd=sd))
end;
@show learners;learners = 5Γ3 DataFrame
β Row β name β mean β sd β
β β String β Float64 β Float64 β
βββββββΌβββββββββββββββββββββββββΌβββββββββββΌββββββββββββ€
β 1 β rf β 0.653424 β 0.0754433 β
β 2 β AdaBoostClassifier β 0.69504 β 0.0514792 β
β 3 β SGDClassifier β 0.694908 β 0.0641564 β
β 4 β prunetree β 0.621927 β 0.0578242 β
β 5 β LinearSVC β 0.726097 β 0.0498317 β10.2 Parallel Processing
using Random
using DataFrames
using Distributed
nprocs() == 1 && addprocs()
@everywhere using DataFrames
@everywhere using AutoMLPipeline
@everywhere profbdata = getprofb()
@everywhere X = profbdata[:,2:end]
@everywhere Y = profbdata[:,1] |> Vector;
@everywhere jrf = RandomForest()
@everywhere ohe = OneHotEncoder()
@everywhere catf = CatFeatureSelector()
@everywhere numf = NumFeatureSelector()
@everywhere tree = PrunedTree()
@everywhere disc = CatNumDiscriminator()
@everywhere ada = skoperator("AdaBoostClassifier")
@everywhere sgd = skoperator("SGDClassifier")
@everywhere std = skoperator("StandardScaler")
@everywhere lsvc = skoperator("LinearSVC")
learners = @sync @distributed (vcat) for learner in [jrf,ada,sgd,tree,lsvc]
pcmc = disc |> ((catf |> ohe) + (numf |> std)) |> learner
println(learner.name[1:end-4])
mean,sd,_ = crossvalidate(pcmc,X,Y,"accuracy_score",10)
DataFrame(name=learner.name[1:end-4],mean=mean,sd=sd)
end
@show learners; From worker 3: AdaBoostClassifier
From worker 4: SGDClassifier
From worker 5: prunetree
From worker 2: rf
From worker 6: LinearSVC
From worker 4: fold: 1, 0.6716417910447762
From worker 5: fold: 1, 0.6567164179104478
From worker 6: fold: 1, 0.6865671641791045
From worker 2: fold: 1, 0.7164179104477612
From worker 4: fold: 2, 0.7164179104477612
From worker 5: fold: 2, 0.6119402985074627
From worker 6: fold: 2, 0.8059701492537313
From worker 2: fold: 2, 0.6716417910447762
From worker 4: fold: 3, 0.6764705882352942
....
learners = 5Γ3 DataFrame
β Row β name β mean β sd β
β β String β Float64 β Float64 β
βββββββΌβββββββββββββββββββββββββΌβββββββββββΌββββββββββββ€
β 1 β rf β 0.647388 β 0.0764844 β
β 2 β AdaBoostClassifier β 0.712862 β 0.0471003 β
β 3 β SGDClassifier β 0.710009 β 0.05173 β
β 4 β prunetree β 0.60428 β 0.0403121 β
β 5 β LinearSVC β 0.726383 β 0.0467506 β11. Automatic Selection of Best Learner
You can use * operation as a selector function which outputs the result of the best learner.
If we use the same pre-processing pipeline in 10, we expect that the average performance of
best learner which is lsvc will be around 73.0.
Random.seed!(1)
pcmc = disc |> ((catf |> ohe) + (numf |> std)) |> (jrf * ada * sgd * tree * lsvc)
crossvalidate(pcmc,X,Y,"accuracy_score",10)fold: 1, 0.7164179104477612
fold: 2, 0.7910447761194029
fold: 3, 0.6911764705882353
fold: 4, 0.7761194029850746
fold: 5, 0.6567164179104478
fold: 6, 0.7014925373134329
fold: 7, 0.6417910447761194
fold: 8, 0.7058823529411765
fold: 9, 0.746268656716418
fold: 10, 0.835820895522388
errors: 0
(mean = 0.7262730465320456, std = 0.060932268798867976, folds = 10, errors = 0)12. Learners as Transformers
It is also possible to use learners in the middle of expression to serve as transformers and their outputs become inputs to the final learner as illustrated below.
expr = (
((numf |> rb)+(catf |> ohe) |> gb) +
((numf |> rb)+(catf |> ohe) |> rf)
) |> ohe |> ada;
crossvalidate(expr,X,Y,"accuracy_score")fold: 1, 0.6567164179104478
fold: 2, 0.5522388059701493
fold: 3, 0.7205882352941176
fold: 4, 0.7313432835820896
fold: 5, 0.6567164179104478
fold: 6, 0.6119402985074627
fold: 7, 0.6119402985074627
fold: 8, 0.6470588235294118
fold: 9, 0.6716417910447762
fold: 10, 0.6119402985074627
errors: 0
(mean = 0.6472124670763829, std = 0.053739947087648336, folds = 10, errors = 0)One can even include selector function as part of transformer preprocessing routine:
pjrf = disc |> ((catf |> ohe) + (numf |> std)) |>
((jrf * ada ) + (sgd * tree * lsvc)) |> ohe |> ada
crossvalidate(pjrf,X,Y,"accuracy_score")fold: 1, 0.7164179104477612
fold: 2, 0.7164179104477612
fold: 3, 0.7941176470588235
fold: 4, 0.7761194029850746
fold: 5, 0.6268656716417911
fold: 6, 0.6716417910447762
fold: 7, 0.7611940298507462
fold: 8, 0.7352941176470589
fold: 9, 0.7761194029850746
fold: 10, 0.6865671641791045
errors: 0
(mean = 0.7260755048287972, std = 0.0532393731318768, folds = 10, errors = 0)Note: The ohe is necessary in both examples
because the outputs of the learners and selector function are categorical
values that need to be hot-bit encoded before feeding to the final ada learner.
13. Tree Visualization of the Pipeline Structure
You can visualize the pipeline by using AbstractTrees Julia package.
# package installation
using Pkg
Pkg.update()
Pkg.add("AbstractTrees")
# load the packages
using AbstractTrees
using AutoMLPipeline
expr = @pipelinex (catf |> ohe) + (numf |> pca) + (numf |> ica) |> rf
:(Pipeline(ComboPipeline(Pipeline(catf, ohe), Pipeline(numf, pca), Pipeline(numf, ica)), rf))
print_tree(stdout, expr):(Pipeline(ComboPipeline(Pipeline(catf, ohe), Pipeline(numf, pca), Pipeline(numf, ica)), rf))
ββ :Pipeline
ββ :(ComboPipeline(Pipeline(catf, ohe), Pipeline(numf, pca), Pipeline(numf, ica)))
β ββ :ComboPipeline
β ββ :(Pipeline(catf, ohe))
β β ββ :Pipeline
β β ββ :catf
β β ββ :ohe
β ββ :(Pipeline(numf, pca))
β β ββ :Pipeline
β β ββ :numf
β β ββ :pca
β ββ :(Pipeline(numf, ica))
β ββ :Pipeline
β ββ :numf
β ββ :ica
ββ :rfExtending AutoMLPipeline
If you want to add your own filter or transformer or learner,
take note that filters and transformers process the
input features but ignores the output argument. On the other hand,
learners process both their input and output arguments during fit!
while transform! expects one input argument in all cases.
First step is to import the abstract types and define your own mutable structure
as subtype of either Learner or Transformer. Next is to import the fit! and
transform! functions so that you can overload them. Also, you must
load the DataFrames package because it is the main format for data processing.
Finally, implement your own fit and transform and export them.
using DataFrames
using AutoMLPipeline.AbsTypes
# import functions for overloading
import AutoMLPipeline.AbsTypes: fit!, transform!
# export the new definitions for dynamic dispatch
export fit!, transform!, MyFilter
# define your filter structure
mutable struct MyFilter <: Transformer
name::String
model::Dict
args::Dict
function MyFilter(args::Dict())
....
end
end
# define your fit! function.
function fit!(fl::MyFilter, inputfeatures::DataFrame, target::Vector=Vector())
....
end
#define your transform! function
function transform!(fl::MyFilter, inputfeatures::DataFrame)::DataFrame
....
endNote that the main format to exchange data is dataframe which requires transform!
output to return a dataframe. The features as input for fit! and transform! shall
be in dataframe format too. This is necessary so that
the pipeline passes the dataframe format consistently to
its corresponding filters/transformers/learners. Once you have
this transformer, you can use it as part of the pipeline
together with the other learners and transformers.
Feature Requests and Contributions
We welcome contributions, feature requests, and suggestions. Here is the link to open an issue for any problems you encounter. If you want to contribute, please follow the guidelines in contributors page.
Help usage
Usage questions can be posted in: