alan-turing-institute / MLJTuning.jl / 445

Committed 4 Nov 2020 - 14:02 coverage: 81.224% (-8.3%) from 89.485%

Build # 445

Build Type

Pull #83

travis-ci-com

Committed by

web-flow

Commit Message

Merge 95ae99102 into c2b8a4e85

Pull Request Pull Request #83: WIP Latin hypercube

Run Details

0 of 40 new or added lines in 1 file covered. (0.0%)

398 of 490 relevant lines covered (81.22%)

987.11 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

0.0

/src/strategies/latin_hypercube.jl

"""
LatinHypercube(nGenerations = 1, popSize = 100, nTournament = 2,
                pTournament = 0.8. interSampleWeight = 1.0,
                ae_power = 2, periodic_ae = false)

Instantiate  grid-based hyperparameter tuning strategy using the library
LatinHypercubeSampling.jl. The optimised Latin Hypercube sampling plan is
created using a genetic based optimization algorithm based on the inverse of the
Audze-Eglais function.
The optimization is run for nGenerations. The population size, number of samples
selected, probability of selection, the inter sample weight of sampling and the
norm can be choosen. There is also the possibility of using a periodic version
of the Audze-Eglais which reduces clustering along the boundaries of the
sampling plan. To enable this feature set `periodic_ae = true`.

### Supported ranges:

A single one-dimensional range or vector of one-dimensioinal ranges
can be specified. Specifically, in `LatinHypercubeSampling` search,
the `range` field of a `TunedModel` instance can be:

- A single one-dimensional range - ie, `ParamRange` object - `r`, constructed
using the `range` method.

- Any vector of objects of the above form

"""
mutable struct LatinHypercube <: TuningStrategy
    nGenerations::Int
    popSize::Int
    nTournament::Int
    pTournament::Number
    interSampleWeight::Number
    ae_power::Number
    periodic_ae::Bool
    rng::Random.AbstractRNG
end

function LatinHypercube(; nGenerations = 1, popSize = 100, nTournament = 2,
                        pTournament = 0.8, interSampleWeight = 1.0,
                        ae_power = 2, periodic_ae = false,
                        rng=Random.GLOBAL_RNG)

               _rng = rng isa Integer ? Random.MersenneTwister(rng) : rng

              return LatinHypercube(nGenerations,popSize,nTournament,
                                    pTournament, interSampleWeight, ae_power,
                                    periodic_ae, _rng)

end

function _create_bounds_and_dims(d,r)
    bounds = []
    dims = []
    for i = 1:d
        if r[i] isa NumericRange
            if !(r[i].scale isa Symbol)
                error("Callable scale not supported in LatinHyperCube tuning.")
            end
            push!(dims,Continuous())
            if isfinite(r[i].lower) && isfinite(r[i].upper)
                push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].lower),
                 transform(MLJBase.Scale,scale(r[i].scale),r[i].upper)))
            elseif !isfinite(r.lower) && isfinite(r.upper)
                push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].upper - 2*r[i].unit),
                 transform(MLJBase.Scale,scale(r[i].scale),r[i].upper)))
            elseif isfinite(r.lower) && !isfinite(r.upper)
                push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].lower),
                 transform(MLJBase.Scale,scale(r[i].scale),r[i].lower + 2*r[i].unit)))
            else
                push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].origin - r[i].unit),
                 transform(MLJBase.Scale,scale(r[i].scale),r[i].origin + r[i].unit)))
            end
        else
            push!(dims, Categorical(length(r[i].values), 1.0))
            push!(bounds,(0,length(r[i].values)))
        end
    end
    return bounds, dims
end

function setup(tuning::LatinHypercube, model, r, verbosity)
    d = length(r)
    bounds, dims = _create_bounds_and_dims(d, r)
    initial_plan = randomLHC(n,dims,nGenerations,
                              popsize = tuning.popSize,
                              ntour = tuning.nTournament,
                              ptour = tuning.pTournment,
                              interSampleWeight = tuning.interSampleWeight,
                              periodic_ae = tuning.periodic_ae,
                              ae_power = tuning.ae_power,
                              rng = tuning.rng)
    scaled_plan = scaleLHC(initial_plan, bounds)
    @inbounds for i = 1:size(scaled_plan,1)
        for j = 1:size(scaled_plan,2)
            if dims[j] isa LatinHypercubeSampling.Continuous
                scaled_plan[i][j] = inverse_transform(MLJBase.Scale,
                                                      scale(r[j].scale),
                                                      scaled_plan[i][j])
            else
                scaled_plan[i][j] = r[j].values[scaled_plan[i][j]]
            end
        end
    end
    ranges = r
    fields = map(r -> r.field, ranges)
    models = makeLatinHypercube(model, fields, scaled_plan)
    state = (models = models,
             fields = fields)
    return state
end

function MLJTuning.models(tuning::LatinHypercube,
                          model,
                          hystory,
                          state,
                          n_remaining,
                          verbosity)
     return state.models[_length(history) + 1:end], state
end

function makeLatinHypercube(prototype::Model,fields,plan)
    N = size(plan,1)
    map(1:N) do i
        clone = deepcopy(prototype)
        for k in eachindex(fields)
            recursive_setproperty(clone,fields[k],plan[i,k])
        end
        clone
    end
end

1	"""
2	LatinHypercube(nGenerations = 1, popSize = 100, nTournament = 2,
3	pTournament = 0.8. interSampleWeight = 1.0,
4	ae_power = 2, periodic_ae = false)
5
6	Instantiate grid-based hyperparameter tuning strategy using the library
7	LatinHypercubeSampling.jl. The optimised Latin Hypercube sampling plan is
8	created using a genetic based optimization algorithm based on the inverse of the
9	Audze-Eglais function.
10	The optimization is run for nGenerations. The population size, number of samples
11	selected, probability of selection, the inter sample weight of sampling and the
12	norm can be choosen. There is also the possibility of using a periodic version
13	of the Audze-Eglais which reduces clustering along the boundaries of the
14	sampling plan. To enable this feature set `periodic_ae = true`.
15
16	### Supported ranges:
17
18	A single one-dimensional range or vector of one-dimensioinal ranges
19	can be specified. Specifically, in `LatinHypercubeSampling` search,
20	the `range` field of a `TunedModel` instance can be:
21
22	- A single one-dimensional range - ie, `ParamRange` object - `r`, constructed
23	using the `range` method.
24
25	- Any vector of objects of the above form
26
27	"""
28	mutable struct LatinHypercube <: TuningStrategy
29	nGenerations::Int
30	popSize::Int
31	nTournament::Int
32	pTournament::Number
33	interSampleWeight::Number
34	ae_power::Number
35	periodic_ae::Bool
36	rng::Random.AbstractRNG
37	end
38
39	function LatinHypercube(; nGenerations = 1, popSize = 100, nTournament = 2,
40	pTournament = 0.8, interSampleWeight = 1.0,
41	ae_power = 2, periodic_ae = false,
42	rng=Random.GLOBAL_RNG)
43
NEW 44	_rng = rng isa Integer ? Random.MersenneTwister(rng) : rng	!
45
NEW 46	return LatinHypercube(nGenerations,popSize,nTournament,	!
47	pTournament, interSampleWeight, ae_power,
48	periodic_ae, _rng)
49
50	end
51
52	function _create_bounds_and_dims(d,r)
NEW 53	bounds = []	!
NEW 54	dims = []	!
NEW 55	for i = 1:d	!
NEW 56	if r[i] isa NumericRange	!
NEW 57	if !(r[i].scale isa Symbol)	!
NEW 58	error("Callable scale not supported in LatinHyperCube tuning.")	!
59	end
NEW 60	push!(dims,Continuous())	!
NEW 61	if isfinite(r[i].lower) && isfinite(r[i].upper)	!
NEW 62	push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].lower),	!
63	transform(MLJBase.Scale,scale(r[i].scale),r[i].upper)))
NEW 64	elseif !isfinite(r.lower) && isfinite(r.upper)	!
NEW 65	push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].upper - 2*r[i].unit),	!
66	transform(MLJBase.Scale,scale(r[i].scale),r[i].upper)))
NEW 67	elseif isfinite(r.lower) && !isfinite(r.upper)	!
NEW 68	push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].lower),	!
69	transform(MLJBase.Scale,scale(r[i].scale),r[i].lower + 2*r[i].unit)))
70	else
NEW 71	push!(bounds,(transform(MLJBase.Scale,scale(r[i].scale),r[i].origin - r[i].unit),	!
72	transform(MLJBase.Scale,scale(r[i].scale),r[i].origin + r[i].unit)))
73	end
74	else
NEW 75	push!(dims, Categorical(length(r[i].values), 1.0))	!
NEW 76	push!(bounds,(0,length(r[i].values)))	!
77	end
78	end
NEW 79	return bounds, dims	!
80	end
81
82	function setup(tuning::LatinHypercube, model, r, verbosity)
NEW 83	d = length(r)	!
NEW 84	bounds, dims = _create_bounds_and_dims(d, r)	!
NEW 85	initial_plan = randomLHC(n,dims,nGenerations,	!
86	popsize = tuning.popSize,
87	ntour = tuning.nTournament,
88	ptour = tuning.pTournment,
89	interSampleWeight = tuning.interSampleWeight,
90	periodic_ae = tuning.periodic_ae,
91	ae_power = tuning.ae_power,
92	rng = tuning.rng)
NEW 93	scaled_plan = scaleLHC(initial_plan, bounds)	!
NEW 94	@inbounds for i = 1:size(scaled_plan,1)	!
NEW 95	for j = 1:size(scaled_plan,2)	!
NEW 96	if dims[j] isa LatinHypercubeSampling.Continuous	!
NEW 97	scaled_plan[i][j] = inverse_transform(MLJBase.Scale,	!
98	scale(r[j].scale),
99	scaled_plan[i][j])
100	else
NEW 101	scaled_plan[i][j] = r[j].values[scaled_plan[i][j]]	!
102	end
103	end
104	end
NEW 105	ranges = r	!
NEW 106	fields = map(r -> r.field, ranges)	!
NEW 107	models = makeLatinHypercube(model, fields, scaled_plan)	!
NEW 108	state = (models = models,	!
109	fields = fields)
NEW 110	return state	!
111	end
112
113	function MLJTuning.models(tuning::LatinHypercube,
114	model,
115	hystory,
116	state,
117	n_remaining,
118	verbosity)
NEW 119	return state.models[_length(history) + 1:end], state	!
120	end
121
122	function makeLatinHypercube(prototype::Model,fields,plan)
NEW 123	N = size(plan,1)	!
NEW 124	map(1:N) do i	!
NEW 125	clone = deepcopy(prototype)	!
NEW 126	for k in eachindex(fields)	!
NEW 127	recursive_setproperty(clone,fields[k],plan[i,k])	!
128	end
NEW 129	clone	!
130	end
131	end

alan-turing-institute / MLJTuning.jl / 445

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous