Merge pull request #8 from ablaom/dev

For a 0.1.2 release

Project.toml

@@ -1,7 +1,7 @@
 name = "EarlyStopping"
 uuid = "792122b4-ca99-40de-a6bc-6742525f08b6"
 authors = ["Anthony D. Blaom <[email protected]>"]
-version = "0.1.1"
+version = "0.1.2"
 
 [deps]
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
@@ -11,7 +11,8 @@ Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 julia = "^1"
 
 [extras]
+InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [targets]
-test = ["Test"]
+test = ["InteractiveUtils", "Test"]

src/EarlyStopping.jl

@@ -6,7 +6,8 @@ import Base.+
 
 export StoppingCriterion,
     Never, NotANumber, TimeLimit, GL, Patience, UP, PQ, NumberLimit,
-    Disjunction, criteria, stopping_time, EarlyStopper,
+    Threshold, Disjunction,
+    criteria, stopping_time, EarlyStopper,
     done!, message, needs_in_and_out_of_sample
 
 include("api.jl")

src/criteria.jl

@@ -124,6 +124,8 @@ GL(; alpha=2.0) = GL(alpha)
 
 update(::GL, loss) = (loss=loss, min_loss=loss)
 update(::GL, loss, state) = (loss=loss, min_loss=min(loss, state.min_loss))
+# in case first loss consumed was a training loss:
+update(criterion::GL, loss, ::Nothing) = update(criterion, loss)
 done(criterion::GL, state) =
     generalization_loss(state.loss, state.min_loss) > criterion.alpha
 
@@ -295,6 +297,8 @@ update(criterion::Patience, loss) = (loss=loss, n_increases=0)
     end
     return (loss=loss, n_increases=n)
 end
+# in case first loss consumed was a training loss:
+update(criterion::Patience, loss, ::Nothing) = update(criterion, loss)
 
 done(criterion::Patience, state) = state.n_increases == criterion.n
 
@@ -324,5 +328,30 @@ update(criterion::NumberLimit, loss) = 1
 @inline function update(criterion::NumberLimit, loss, state)
     return state+1
 end
+# in case first loss consumed was a training loss:
+update(criterion::NumberLimit, loss, ::Nothing) = update(criterion, loss)
 
 done(criterion::NumberLimit, state) = state == criterion.n
+
+
+# ## THRESHOLD
+
+"""
+    Threshold(; value=0.0)
+
+$STOPPING_DOC
+
+A stop is triggered as soon as the loss drops below `value`.
+
+"""
+mutable struct Threshold <: StoppingCriterion
+    value::Float64
+end
+Threshold(; value=0.0) = Threshold(value)
+
+update(criterion::Threshold, loss) = loss
+update(criterion::Threshold, loss, state) = loss
+# in case first loss consumed was a training loss:
+update(criterion::Threshold, loss, ::Nothing) = loss
+
+done(criterion::Threshold, state) = state < criterion.value

src/disjunction.jl

@@ -20,6 +20,7 @@ struct Disjunction{A,B} <: StoppingCriterion
     end
 end
 
+Disjunction() = Never()
 Disjunction(a) = a
 Disjunction(a, b, c...) = Disjunction(Disjunction(a,b), c...)
 

test/criteria.jl

@@ -159,4 +159,17 @@ end
     end
 end
 
+@testset "Threshold" begin
+    @test Threshold().value == 0.0
+    stopping_time(Threshold(2.5), Float64[12, 32, 3, 2, 5, 7]) == 4
+end
+
+@testset "robustness to first loss being a training loss" begin
+    for C in subtypes(StoppingCriterion)
+        losses = float.(4:-1:1)
+        is_training = [true, true, false, false]
+        stopping_time(C(), losses, is_training)
+    end
+end
+
 true

test/runtests.jl

@@ -1,4 +1,4 @@
-using EarlyStopping, Dates, Test
+using EarlyStopping, Dates, Test, InteractiveUtils
 
 @testset "criteria.jl" begin
     include("criteria.jl")