Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

support regression models #38

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

support regression models #38

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
103 changes: 58 additions & 45 deletions src/LIBLINEAR.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,11 @@ const L2R_L2LOSS_SVR_DUAL = Cint(12)
const L2R_L1LOSS_SVR_DUAL = Cint(13)
const ONECLASS_SVM = Cint(21)

function is_regression_model(m)
solver = m.solver_type
return solver in (L2R_L2LOSS_SVR, L2R_L1LOSS_SVR_DUAL, L2R_L2LOSS_SVR_DUAL)
end

struct FeatureNode
index :: Cint
value :: Float64
Expand Down Expand Up @@ -257,65 +262,73 @@ function linear_train(
model
end

# predict
function linear_predict(
model :: LinearModel{T},
instances :: AbstractMatrix{U};
probability_estimates :: Bool=false,
verbose :: Bool=false) where {T, U<:Real}
set_print(verbose)
ninstances = size(instances, 2) # instances are in columns

size(instances, 1) != model.nr_feature &&
function predict_decision_values(model, instances, probability_estimates)
ninstances = size(instances, 2) # instances are in columns
if size(instances, 1) != model.nr_feature
error("""Model has $(model.nr_feature) features but
$(size(instances, 1)) provided (instances are in columns)""")
$(size(instances, 1)) provided (instances are in columns)""")
end

model.bias >= 0 &&
(instances = [instances; fill(model.bias, 1, ninstances)])
if model.bias >= 0
instances = [instances; fill(model.bias, 1, ninstances)]
end

m = Array{Model}(undef, 1)
m[1] = Model(Parameter(model.solver_type, .0, .0, Cint(0),
convert(Ptr{Cint}, C_NULL), convert(Ptr{Float64}, C_NULL), .0,.0,
convert(Ptr{Float64}, C_NULL), Cint(0)),
model.nr_class, model.nr_feature, pointer(model.w),
pointer(model._labels), model.bias, model.rho)
m = Model(Parameter(model.solver_type, .0, .0, Cint(0),
convert(Ptr{Cint}, C_NULL), convert(Ptr{Float64}, C_NULL), .0,.0,
convert(Ptr{Float64}, C_NULL), Cint(0)),
model.nr_class, model.nr_feature, pointer(model.w),
pointer(model._labels), model.bias, model.rho)

(nodes, nodeptrs) = instances2nodes(instances)

if model.solver_type == ONECLASS_SVM
# In this case we need to return inlier/outlier class labels
# which may not be of type `T`
class = Array{String}(undef, ninstances)
else
class = Array{T}(undef, ninstances)
end
w_number = Int(model.nr_class == 2 && model.solver_type != MCSVM_CS ?
1 : model.nr_class)
w_number = Int(model.nr_class == 2 && model.solver_type != MCSVM_CS ? 1 : model.nr_class)
decvalues = Array{Float64}(undef, w_number, ninstances)
for i = 1:ninstances
if probability_estimates
output = ccall((:predict_probability, liblinear), Float64, (Ptr{Cvoid}, Ptr{FeatureNode}, Ptr{Float64}),
pointer(m), nodeptrs[i], pointer(decvalues, w_number*(i-1)+1))
else
output = ccall((:predict_values, liblinear), Float64, (Ptr{Cvoid}, Ptr{FeatureNode}, Ptr{Float64}),
pointer(m), nodeptrs[i], pointer(decvalues, w_number*(i-1)+1))
end
output_int = round(Int,output)

# For one-class SVM, `predict_values` returns +/- 1
# corresponding to outliers or not. This doesn't seem to be documented,
# but the code clearly returns +/- 1:
# https://github.com/cjlin1/liblinear/blob/8dc206b782e07676dc0d00678bedd295ce85acf3/linear.cpp#L3295
# and that is the return from scipy as well.
if model.solver_type === ONECLASS_SVM
c = output_int == -1 ? "outlier" : output_int == 1 ? "inlier" : error("Unexpected output $output_int")
output = ccall((:predict_probability, liblinear), Float64,
(Ref{Model}, Ptr{FeatureNode}, Ptr{Float64}),
m, nodeptrs[i], pointer(decvalues, w_number*(i-1)+1))
else
c = model.labels[output_int]
output = ccall((:predict_values, liblinear), Float64,
(Ref{Model}, Ptr{FeatureNode}, Ptr{Float64}),
m, nodeptrs[i], pointer(decvalues, w_number*(i-1)+1))
end
class[i] = c
end

(class, decvalues)
return decvalues
end

# For one-class SVM, `predict_values` returns +/- 1
# corresponding to outliers or not. This doesn't seem to be documented,
# but the code clearly returns +/- 1:
# https://github.com/cjlin1/liblinear/blob/8dc206b782e07676dc0d00678bedd295ce85acf3/linear.cpp#L3295
# and that is the return from scipy as well.
oneclass_output(decval) = ifelse(only(decval) > 0, "inlier", "outlier")

# predict
function linear_predict(
model :: LinearModel{T},
instances :: AbstractMatrix{U};
probability_estimates :: Bool=false,
verbose :: Bool=false) where {T, U<:Real}
set_print(verbose)
ninstances = size(instances, 2) # instances are in columns

decvalues = predict_decision_values(model, instances, probability_estimates)

if is_regression_model(model)
return decvalues, decvalues
end

if model.solver_type === ONECLASS_SVM
classes = map(oneclass_output, eachcol(decvalues))
else
classes = map(decvals -> model.labels[argmax(decvals)],
eachcol(decvalues))
end

return classes, decvalues
end

end # module
9 changes: 9 additions & 0 deletions test/validation.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,4 +38,13 @@
@test (class .== labels[2:2:end]) == correct
end
end

@testset "Regression" begin
x = rand(1, 10000);
y = vec(2 .* x)
m = linear_train(y, x, solver_type=LIBLINEAR.L2R_L2LOSS_SVR, eps=1e-5)
y_ = first(linear_predict(m, x))
@test size(y, 1) == size(y_, 2) # TODO: provide better test
end

end
Loading