EvoTrees

A Julia implementation of boosted trees with CPU and GPU support. Efficient histogram based algorithms with support for multiple loss functions (notably multi-target objectives such as max likelihood methods).

R binding available.

Input features are expected to be Matrix{Float64/Float32}. Tables/DataFrames format can be handled through MLJ (see below).

Supported tasks

CPU

• linear
• logistic
• Poisson
• L1 (mae regression)
• Quantile
• multiclassification (softmax)
• Gaussian (max likelihood)

Set parameter device="cpu".

GPU

• linear
• logistic
• Gaussian (max likelihood)

Set parameter device="gpu".

Installation

Latest:

julia> Pkg.add("https://github.com/Evovest/EvoTrees.jl")

From General Registry:

julia> Pkg.add("EvoTrees")

Performance

Data consists of randomly generated float32. Training is performed on 200 iterations. Code to reproduce is here.

EvoTrees: v0.8.4 XGBoost: v1.1.1

CPU: 16 threads on AMD Threadripper 3970X GPU: NVIDIA RTX 2080

Training:

Dimensions / Algo	XGBoost Hist	EvoTrees	EvoTrees GPU
100K x 100	1.10s	1.80s	3.14s
500K x 100	4.83s	4.98s	4.98s
1M x 100	9.84s	9.89s	7.37s
5M x 100	45.5s	53.8s	25.8s

Inference:

Dimensions / Algo	XGBoost Hist	EvoTrees	EvoTrees GPU
100K x 100	0.164s	0.026s	0.013s
500K x 100	0.796s	0.175s	0.055s
1M x 100	1.59s	0.396s	0.108s
5M x 100	7.96s	2.15s	0.543s

Parameters

• loss: {:linear, :logistic, :poisson, :L1, :quantile, :softmax, :gaussian}
• device: {"cpu", "gpu"}
• nrounds: integer, default=10
• λ: L2 regularization, float, default=0.0
• γ: min gain for split, default=0.0
• η: learning rate, default=0.1
• max_depth: integer, default=5
• min_weight: float >= 0 default=1.0
• rowsample: float [0,1] default=1.0
• colsample: float [0,1] default=1.0
• nbins: Int, number of bins into which features will be quantilized default=64
• α: float [0,1], set the quantile or bias in L1 default=0.5
• metric: {:mse, :rmse, :mae, :logloss, :quantile, :gini, :gaussian, :none}, default=:none
• rng: random controller, either a Random.AbstractRNG or an Int acting as a seed. Default=123.

MLJ Integration

See official project page for more info.

using StatsBase: sample
using EvoTrees
using EvoTrees: sigmoid, logit
using MLJBase

features = rand(10_000) .* 5 .- 2
X = reshape(features, (size(features)[1], 1))
Y = sin.(features) .* 0.5 .+ 0.5
Y = logit(Y) + randn(size(Y))
Y = sigmoid(Y)
y = Y
X = MLJBase.table(X)

# @load EvoTreeRegressor
# linear regression
tree_model = EvoTreeRegressor(loss=:linear, max_depth=5, η=0.05, nrounds=10)

# set machine
mach = machine(tree_model, X, y)

# partition data
train, test = partition(eachindex(y), 0.7, shuffle=true); # 70:30 split

# fit data
fit!(mach, rows=train, verbosity=1)

# continue training
mach.model.nrounds += 10
fit!(mach, rows=train, verbosity=1)

# predict on train data
pred_train = predict(mach, selectrows(X, train))
mean(abs.(pred_train - selectrows(Y, train)))

# predict on test data
pred_test = predict(mach, selectrows(X, test))
mean(abs.(pred_test - selectrows(Y, test)))

Getting started using internal API

Minimal example to fit a noisy sinus wave.

using EvoTrees
using EvoTrees: sigmoid, logit

# prepare a dataset
features = rand(10000) .* 20 .- 10
X = reshape(features, (size(features)[1], 1))
Y = sin.(features) .* 0.5 .+ 0.5
Y = logit(Y) + randn(size(Y))
Y = sigmoid(Y)
𝑖 = collect(1:size(X, 1))

# train-eval split
𝑖_sample = sample(𝑖, size(𝑖, 1), replace = false)
train_size = 0.8
𝑖_train = 𝑖_sample[1:floor(Int, train_size * size(𝑖, 1))]
𝑖_eval = 𝑖_sample[floor(Int, train_size * size(𝑖, 1))+1:end]

X_train, X_eval = X[𝑖_train, :], X[𝑖_eval, :]
Y_train, Y_eval = Y[𝑖_train], Y[𝑖_eval]

params1 = EvoTreeRegressor(
    loss=:linear, metric=:mse,
    nrounds=100, nbins = 100,
    λ = 0.5, γ=0.1, η=0.1,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0)
model = fit_evotree(params1, X_train, Y_train, X_eval = X_eval, Y_eval = Y_eval, print_every_n = 25)
pred_eval_linear = predict(model, X_eval)

# logistic / cross-entropy
params1 = EvoTreeRegressor(
    loss=:logistic, metric = :logloss,
    nrounds=100, nbins = 100,
    λ = 0.5, γ=0.1, η=0.1,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0)
model = fit_evotree(params1, X_train, Y_train, X_eval = X_eval, Y_eval = Y_eval, print_every_n = 25)
pred_eval_logistic = predict(model, X_eval)

# Poisson
params1 = EvoTreeCount(
    loss=:poisson, metric = :poisson,
    nrounds=100, nbins = 100,
    λ = 0.5, γ=0.1, η=0.1,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0)
model = fit_evotree(params1, X_train, Y_train, X_eval = X_eval, Y_eval = Y_eval, print_every_n = 25)
@time pred_eval_poisson = predict(model, X_eval)

# L1
params1 = EvoTreeRegressor(
    loss=:L1, α=0.5, metric = :mae,
    nrounds=100, nbins=100,
    λ = 0.5, γ=0.0, η=0.1,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0)
model = fit_evotree(params1, X_train, Y_train, X_eval = X_eval, Y_eval = Y_eval, print_every_n = 25)
pred_eval_L1 = predict(model, X_eval)

Quantile Regression

# q50
params1 = EvoTreeRegressor(
    loss=:quantile, α=0.5, metric = :quantile,
    nrounds=200, nbins = 100,
    λ = 0.1, γ=0.0, η=0.05,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0)
model = fit_evotree(params1, X_train, Y_train, X_eval = X_eval, Y_eval = Y_eval, print_every_n = 25)
pred_train_q50 = predict(model, X_train)

# q20
params1 = EvoTreeRegressor(
    loss=:quantile, α=0.2, metric = :quantile,
    nrounds=200, nbins = 100,
    λ = 0.1, γ=0.0, η=0.05,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0)
model = fit_evotree(params1, X_train, Y_train, X_eval = X_eval, Y_eval = Y_eval, print_every_n = 25)
pred_train_q20 = predict(model, X_train)

# q80
params1 = EvoTreeRegressor(
    loss=:quantile, α=0.8,
    nrounds=200, nbins = 100,
    λ = 0.1, γ=0.0, η=0.05,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0)
model = fit_evotree(params1, X_train, Y_train, X_eval = X_eval, Y_eval = Y_eval, print_every_n = 25)
pred_train_q80 = predict(model, X_train)

Gaussian Max Likelihood

params1 = EvoTreeGaussian(
    loss=:gaussian, metric=:gaussian,
    nrounds=100, nbins=100,
    λ = 0.0, γ=0.0, η=0.1,
    max_depth = 6, min_weight = 1.0,
    rowsample=0.5, colsample=1.0, seed=123)

Feature importance

Returns the normalized gain by feature.

features_gain = importance(model, var_names)

Plot

Plot a given tree of the model:

plot(model, 2)

Note that 1st tree is used to set the bias so the first real tree is #2.

Save/Load

EvoTrees.save(model, "data/model.bson")
model = EvoTrees.load("data/model.bson");