Merge pull request #431 from JuliaTrustworthyAI/430-move-to-explicit-ad

JuliaTrustworthyAI · Apr 17, 2024 · fd40d9f · fd40d9f · pat-alt · Apr 17, 2024
2 parents 871d0ae + ad8a815
commit fd40d9f
Show file tree

Hide file tree

Showing 36 changed files with 173 additions and 72 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -4,21 +4,31 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-*Note*: We try to adhere to these practices as of version `1.1.1`.
+*Note*: We try to adhere to these practices as of version [v1.1.1].
 
-## [Unreleased]
+## Version [v1.1.3] - 2024-04-17
 
 ### Added
 
 - Adds a section on `Convergence` to the documentation, `Changelog.jl` functionality and a few doc tests. [#429]
 
-## [1.1.2] - 2024-04-16
+### Changed
+
+- Changes style of taking gradients for the counterfactual search from implicit to explicit. [#430]
+- Removed all implicit imports. [#430]
+
+### Removed 
+
+- Removed CUDA.jl dependency, because redundant. [#430]
+- Removed Parameters.jl dependency, because redundant. [#430]
+
+## Version [v1.1.2] - 2024-04-16
 
 ### Changed
 
 - Replaces the GIF in the README and introduction of docs for a static image. 
 
-## [1.1.1] - 2024-04-15
+## Version [v1.1.1] - 2024-04-15
 
 ### Added
 

diff --git a/Project.toml b/Project.toml
@@ -4,7 +4,6 @@ authors = ["Patrick Altmeyer <[email protected]>"]
 version = "1.1.3"
 
 [deps]
-CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 CategoricalArrays = "324d7699-5711-5eae-9e2f-1d82baa6b597"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 DataFrames = "a93c6f00-e57d-5684-b7b6-d8193f3e46c0"
@@ -19,7 +18,6 @@ MLJDecisionTreeInterface = "c6f25543-311c-4c74-83dc-3ea6d1015661"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
 MultivariateStats = "6f286f6a-111f-5878-ab1e-185364afe411"
 PackageExtensionCompat = "65ce6f38-6b18-4e1d-a461-8949797d7930"
-Parameters = "d96e819e-fc66-5662-9728-84c9c7592b0a"
 ProgressMeter = "92933f4c-e287-5a05-a399-4b506db050ca"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Serialization = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
@@ -41,7 +39,6 @@ NeuroTreeExt = "NeuroTreeModels"
 
 [compat]
 Aqua = "0.8"
-CUDA = "3, 4, 5"
 CategoricalArrays = "0.10"
 ChainRulesCore = "1.15"
 DataFrames = "1"
@@ -59,7 +56,6 @@ MLUtils = "0.2, 0.3, 0.4"
 MultivariateStats = "0.9, 0.10"
 NeuroTreeModels = "1.1.0"
 PackageExtensionCompat = "1"
-Parameters = "0.12"
 ProgressMeter = "1"
 Random = "1.6, 1.7, 1.8, 1.9, 1.10"
 Serialization = "1.6, 1.7, 1.8, 1.9, 1.10"

diff --git a/src/CounterfactualExplanations.jl b/src/CounterfactualExplanations.jl
@@ -1,14 +1,14 @@
 module CounterfactualExplanations
 
 # Package extensions:
-using PackageExtensionCompat
+using PackageExtensionCompat: PackageExtensionCompat, @require_extensions
 function __init__()
     @require_extensions
 end
 
 # Dependencies:
 using Flux
-using TaijaBase
+using TaijaBase: TaijaBase
 
 # Setup:
 include("artifacts_setup.jl")

diff --git a/src/artifacts_setup.jl b/src/artifacts_setup.jl
@@ -1,4 +1,4 @@
-using LazyArtifacts
+using LazyArtifacts: LazyArtifacts, @artifact_str
 
 function generate_artifact_dir(name::String)
     _artifacts_julia_version = "$(Int(VERSION.major)).$(Int(VERSION.minor))"

diff --git a/src/convergence/invalidation_rate.jl b/src/convergence/invalidation_rate.jl
@@ -1,3 +1,7 @@
+using Distributions: Distributions
+using Flux: Flux
+using LinearAlgebra: LinearAlgebra
+
 Base.@kwdef struct InvalidationRateConvergence <: AbstractConvergence
     invalidation_rate::AbstractFloat = 0.1
     max_iter::Int = 100
@@ -44,7 +48,9 @@ function invalidation_rate(ce::AbstractCounterfactualExplanation)
     end
     gradᵀ = LinearAlgebra.transpose(grad)
 
-    identity_matrix = LinearAlgebra.Matrix{Float32}(I, length(grad), length(grad))
+    identity_matrix = LinearAlgebra.Matrix{Float32}(
+        LinearAlgebra.I, length(grad), length(grad)
+    )
     denominator = sqrt(gradᵀ * ce.convergence.variance * identity_matrix * grad)[1]
 
     normalized_gradient = f_loss / denominator

diff --git a/src/counterfactuals/Counterfactuals.jl b/src/counterfactuals/Counterfactuals.jl
@@ -3,11 +3,11 @@ using .DataPreprocessing
 using .GenerativeModels
 using .Generators
 using .Models
-using ChainRulesCore
+using ChainRulesCore: ChainRulesCore
 using Flux
-using MLUtils
+using MLUtils: MLUtils
 using MultivariateStats
-using Statistics
+using Statistics: Statistics
 using StatsBase
 
 include("core_struct.jl")

diff --git a/src/counterfactuals/core_struct.jl b/src/counterfactuals/core_struct.jl
@@ -71,10 +71,7 @@ function CounterfactualExplanation(
     )
 
     # Initialization:
-    adjust_shape!(ce)                   # adjust shape to specified number of counterfactuals
-    ce.s′ = encode_state(ce)            # encode the counterfactual state
-    ce.s′ = initialize_state(ce)        # initialize the counterfactual state
-    ce.x′ = decode_state(ce)            # decode the counterfactual state
+    adjust_shape!(ce) |> encode_state! |> initialize_state! |> decode_state!
 
     ce.search[:path] = [ce.s′]
     ce.search[:times_changed_features] = zeros(size(decode_state(ce)))

diff --git a/src/counterfactuals/encodings.jl b/src/counterfactuals/encodings.jl
@@ -1,3 +1,7 @@
+using ChainRulesCore: ignore_derivatives
+using MultivariateStats: MultivariateStats
+using StatsBase: StatsBase
+
 """
     encode_array(dt::MultivariateStats.AbstractDimensionalityReduction, x::AbstractArray)
 
@@ -68,7 +72,7 @@ function encode_state(
     if !ce.generator.latent_space && data.standardize
         dt = data.dt
         idx = transformable_features(data)
-        ChainRulesCore.ignore_derivatives() do
+        ignore_derivatives() do
             s = s′[idx, :]
             s = encode_array(dt, s)
             s′[idx, :] = s
@@ -85,6 +89,19 @@ function encode_state(
     return s′
 end
 
+"""
+    encode_state!(ce::CounterfactualExplanation, x::Union{AbstractArray,Nothing}=nothing)
+
+In-place version of `encode_state`.
+"""
+function encode_state!(
+    ce::CounterfactualExplanation, x::Union{AbstractArray,Nothing}=nothing
+)
+    ce.s′ = encode_state(ce, x)
+
+    return ce
+end
+
 """
 function decode_state(
     ce::CounterfactualExplanation,
@@ -118,7 +135,7 @@ function decode_state(
 
         # Continuous:
         idx = transformable_features(data)
-        ChainRulesCore.ignore_derivatives() do
+        ignore_derivatives() do
             s = s′[idx, :]
             s = decode_array(dt, s)
             s′[idx, :] = s
@@ -137,3 +154,16 @@ function decode_state(
 
     return s′
 end
+
+"""
+    decode_state!(ce::CounterfactualExplanation, x::Union{AbstractArray,Nothing}=nothing)
+
+In-place version of `decode_state`.
+"""
+function decode_state!(
+    ce::CounterfactualExplanation, x::Union{AbstractArray,Nothing}=nothing
+)
+    ce.x′ = decode_state(ce, x)
+
+    return ce
+end
diff --git a/src/counterfactuals/initialisation.jl b/src/counterfactuals/initialisation.jl
@@ -1,3 +1,5 @@
+using Flux
+
 """
     initialize_state(ce::CounterfactualExplanation)
 
@@ -30,3 +32,14 @@ function initialize_state(ce::CounterfactualExplanation)
 
     return s′
 end
+
+"""
+    initialize_state!(ce::CounterfactualExplanation)
+
+Initializes the starting point for the factual(s) in-place.
+"""
+function initialize_state!(ce::CounterfactualExplanation)
+    ce.s′ = initialize_state(ce)
+
+    return ce
+end
diff --git a/src/counterfactuals/latent_space_mappings.jl b/src/counterfactuals/latent_space_mappings.jl
@@ -1,3 +1,5 @@
+using Flux: Flux
+
 """
     map_from_latent(
         ce::CounterfactualExplanation,

diff --git a/src/counterfactuals/utils.jl b/src/counterfactuals/utils.jl
@@ -19,7 +19,7 @@ function guess_loss(ce::CounterfactualExplanation)
         elseif ce.M.likelihood == :classification_multi
             loss_fun = Objectives.logitcrossentropy
         else
-            loss_fun = Flux.Losses.mse
+            loss_fun = Objectives.mse
         end
     else
         loss_fun = nothing
@@ -66,7 +66,9 @@ function adjust_shape!(ce::CounterfactualExplanation)
 
     search = ce.search
     search[:mutability] = adjust_shape(ce, search[:mutability])      # augment to account for specified number of counterfactuals
-    return ce.search = search
+    ce.search = search
+
+    return ce
 end
 
 """

diff --git a/src/data_preprocessing/DataPreprocessing.jl b/src/data_preprocessing/DataPreprocessing.jl
@@ -3,12 +3,12 @@ module DataPreprocessing
 using CategoricalArrays
 using CounterfactualExplanations
 using ..GenerativeModels
-using DataFrames
-using Flux
-using MultivariateStats
+using DataFrames: DataFrames
+using Flux: Flux
+using MultivariateStats: MultivariateStats
 using StatsBase
 using Tables
-using Random
+using Random: Random
 
 include("counterfactual_data.jl")
 include("utils.jl")

diff --git a/src/data_preprocessing/counterfactual_data.jl b/src/data_preprocessing/counterfactual_data.jl
@@ -1,4 +1,6 @@
-using MLJBase
+using MLJBase: MLJBase, Continuous, Finite
+using StatsBase: StatsBase, ZScoreTransform
+using Tables: Tables
 
 """
     CounterfactualData(

diff --git a/src/data_preprocessing/utils.jl b/src/data_preprocessing/utils.jl
@@ -1,3 +1,6 @@
+using CategoricalArrays: CategoricalArrays
+using StatsBase: sample
+
 "Treat `CounterfactualData` as scalar when broadcasting."
 Base.broadcastable(data::CounterfactualData) = Ref(data)
 

diff --git a/src/evaluation/Evaluation.jl b/src/evaluation/Evaluation.jl
@@ -4,7 +4,7 @@ using ..CounterfactualExplanations
 using DataFrames
 using ..Generators
 using ..Models
-using LinearAlgebra
+using LinearAlgebra: LinearAlgebra
 using Statistics
 
 include("benchmark.jl")

diff --git a/src/evaluation/benchmark.jl b/src/evaluation/benchmark.jl
@@ -1,7 +1,9 @@
 using Base.Iterators
-using Serialization
+using DataFrames: DataFrames
+using Serialization: Serialization
+using Statistics: mean
 using TaijaBase: AbstractParallelizer, vectorize_collection, parallelize
-using UUIDs
+using UUIDs: UUIDs
 
 "A container for benchmarks of counterfactual explanations. Instead of subtyping `DataFrame`, it contains a `DataFrame` of evaluation measures (see [this discussion](https://discourse.julialang.org/t/creating-an-abstractdataframe-subtype/36451/6?u=pat-alt) for why we don't subtype `DataFrame` directly)."
 struct Benchmark

diff --git a/src/evaluation/evaluate.jl b/src/evaluation/evaluate.jl
@@ -1,4 +1,5 @@
-using UUIDs
+using DataFrames: nrow
+using UUIDs: uuid1
 
 """
     compute_measure(ce::CounterfactualExplanation, measure::Function, agg::Function)

diff --git a/src/evaluation/measures.jl b/src/evaluation/measures.jl
@@ -1,3 +1,5 @@
+using Statistics: Statistics
+
 """
     validity(ce::CounterfactualExplanation; γ=0.5)
 

diff --git a/src/generative_models/GenerativeModels.jl b/src/generative_models/GenerativeModels.jl
@@ -1,12 +1,10 @@
 module GenerativeModels
 
 using CounterfactualExplanations
-using CUDA
 using Flux
-using Parameters
-using ProgressMeter
+using ProgressMeter: ProgressMeter
 using Random
-using Statistics
+using Statistics: Statistics
 
 """
 Base type for generative model.

diff --git a/src/generative_models/encoders.jl b/src/generative_models/encoders.jl
@@ -1,3 +1,6 @@
+using Flux: sigmoid
+using Random: Random
+
 """
 Encoder
 

diff --git a/src/generative_models/vae.jl b/src/generative_models/vae.jl
@@ -1,15 +1,19 @@
+using Flux: Flux, Adam, cpu, gpu
+using ProgressMeter: Progress, next!
+using Statistics: mean
+
 """
     VAEParams <: AbstractGMParams
 
 The default VAE parameters describing both the encoder/decoder architecture and the training process.
 """
-Parameters.@with_kw mutable struct VAEParams <: AbstractGMParams
+Base.@kwdef mutable struct VAEParams <: AbstractGMParams
     η = 1e-3                # learning rate
     λ = 0.01f0              # regularization parameter
     batch_size = 50         # batch size
     epochs = 100            # number of epochs
     seed = 0                # random seed
-    cuda = true             # use GPU
+    gpu = true             # use GPU
     device = gpu            # default device
     latent_dim = 2          # latent dimension
     hidden_dim = 32         # hidden dimension
@@ -41,7 +45,7 @@ function VAE(input_dim; kws...)
     args = VAEParams(; kws...)
 
     # GPU config
-    if args.cuda && CUDA.has_cuda()
+    if args.gpu
         args.device = gpu
     else
         args.device = cpu

diff --git a/src/generators/Generators.jl b/src/generators/Generators.jl
@@ -9,15 +9,13 @@ using Flux
 using LinearAlgebra
 using ..Models
 using ..Objectives
-using Statistics
-using Parameters
+using Statistics: Statistics
 using DecisionTree
-using DataFrames
-using MLJBase
-using MLJDecisionTreeInterface
-using Distributions
+using DataFrames: DataFrames
+using MLJBase: MLJBase
+using MLJDecisionTreeInterface: MLJDecisionTreeInterface
+using Distributions: Distributions
 using Random
-using Statistics
 
 export AbstractGradientBasedGenerator
 export AbstractNonGradientBasedGenerator