Skip to content
This repository was archived by the owner on Aug 25, 2025. It is now read-only.

Hack p to serve as data arg and implement stochastic gradient oracle #89

Merged
merged 3 commits into from
Aug 22, 2024
Merged

Hack p to serve as data arg and implement stochastic gradient oracle #89

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c161954
Hack p to serve as data arg and implement stochastic gradient oracle
Vaibhavdixit02 Aug 22, 2024
3e04244
use master project.toml
Vaibhavdixit02 Aug 22, 2024
36963a4
format
Vaibhavdixit02 Aug 22, 2024
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
15 changes: 10 additions & 5 deletions Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
name = "OptimizationBase"
uuid = "bca83a33-5cc9-4baa-983d-23429ab6bcbb"
authors = ["Vaibhav Dixit <[email protected]> and contributors"]
version = "1.3.3"
version = "1.5.0"

[deps]
ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
Expand Down Expand Up @@ -40,21 +40,26 @@ OptimizationZygoteExt = "Zygote"
[compat]
ADTypes = "1.5"
ArrayInterface = "7.6"
DifferentiationInterface = "0.5.2"
DifferentiationInterface = "0.5"
DocStringExtensions = "0.9"
Enzyme = "0.12.12"
FiniteDiff = "2.12"
ForwardDiff = "0.10.26"
LinearAlgebra = "1.9, 1.10"
ModelingToolkit = "9"
PDMats = "0.11"
Reexport = "1.2"
Requires = "1"
ReverseDiff = "1.14"
SciMLBase = "2"
SymbolicAnalysis = "0.1, 0.2"
SymbolicAnalysis = "0.3"
SymbolicIndexingInterface = "0.3"
Symbolics = "5.12"
Symbolics = "5.12, 6"
Zygote = "0.6.67"
julia = "1.10"

[extras]
Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

[targets]
test = ["Test"]
test = ["Test"]
4 changes: 2 additions & 2 deletions ext/OptimizationEnzymeExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ function OptimizationBase.instantiate_function(f::OptimizationFunction{true}, x,
cons_j = false, cons_vjp = false, cons_jvp = false, cons_h = false,
lag_h = false)
if g == true && f.grad === nothing
function grad(res, θ)
function grad(res, θ, p = p)
Enzyme.make_zero!(res)
Enzyme.autodiff(Enzyme.Reverse,
Const(firstapply),
Expand All @@ -111,7 +111,7 @@ function OptimizationBase.instantiate_function(f::OptimizationFunction{true}, x,
end

if fg == true && f.fg === nothing
function fg!(res, θ)
function fg!(res, θ, p = p)
Enzyme.make_zero!(res)
y = Enzyme.autodiff(Enzyme.ReverseWithPrimal,
Const(firstapply),
Expand Down
56 changes: 52 additions & 4 deletions ext/OptimizationZygoteExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,8 @@ import OptimizationBase.SciMLBase: OptimizationFunction
import OptimizationBase.LinearAlgebra: I, dot
import DifferentiationInterface
import DifferentiationInterface: prepare_gradient, prepare_hessian, prepare_hvp,
prepare_jacobian,
prepare_jacobian, value_and_gradient!,
value_derivative_and_second_derivative!,
gradient!, hessian!, hvp!, jacobian!, gradient, hessian,
hvp, jacobian
using ADTypes, SciMLBase
Expand All @@ -19,8 +20,9 @@ function OptimizationBase.instantiate_function(
g = false, h = false, hv = false, fg = false, fgh = false,
cons_j = false, cons_vjp = false, cons_jvp = false, cons_h = false,
lag_h = false)
global _p = p
function _f(θ)
return f(θ, p)[1]
return f(θ, _p)[1]
end

adtype, soadtype = OptimizationBase.generate_adtype(adtype)
Expand All @@ -30,19 +32,41 @@ function OptimizationBase.instantiate_function(
function grad(res, θ)
gradient!(_f, res, adtype, θ, extras_grad)
end
if p !== SciMLBase.NullParameters() && p !== nothing
function grad(res, θ, p)
global _p = p
gradient!(_f, res, adtype, θ)
end
end
elseif g == true
grad = (G, θ) -> f.grad(G, θ, p)
if p !== SciMLBase.NullParameters() && p !== nothing
grad = (G, θ, p) -> f.grad(G, θ, p)
end
else
grad = nothing
end

if fg == true && f.fg === nothing
if g == false
extras_grad = prepare_gradient(_f, adtype, x)
end
function fg!(res, θ)
(y, _) = value_and_gradient!(_f, res, adtype, θ, extras_grad)
return y
end
if p !== SciMLBase.NullParameters() && p !== nothing
function fg!(res, θ, p)
global _p = p
(y, _) = value_and_gradient!(_f, res, adtype, θ)
return y
end
end
elseif fg == true
fg! = (G, θ) -> f.fg(G, θ, p)
if p !== SciMLBase.NullParameters() && p !== nothing
fg! = (G, θ, p) -> f.fg(G, θ, p)
end
else
fg! = nothing
end
Expand Down Expand Up @@ -188,7 +212,8 @@ function OptimizationBase.instantiate_function(
lag_h! = nothing
end

return OptimizationFunction{true}(f.f, adtype; grad = grad, hess = hess, hv = hv!,
return OptimizationFunction{true}(f.f, adtype;
grad = grad, fg = fg!, hess = hess, hv = hv!, fgh = fgh!,
cons = cons, cons_j = cons_j!, cons_h = cons_h!,
cons_vjp = cons_vjp!, cons_jvp = cons_jvp!,
hess_prototype = hess_sparsity,
Expand Down Expand Up @@ -232,19 +257,41 @@ function OptimizationBase.instantiate_function(
function grad(res, θ)
gradient!(_f, res, adtype.dense_ad, θ, extras_grad)
end
if p !== SciMLBase.NullParameters() && p !== nothing
function grad(res, θ, p)
global p = p
gradient!(_f, res, adtype.dense_ad, θ)
end
end
elseif g == true
grad = (G, θ) -> f.grad(G, θ, p)
if p !== SciMLBase.NullParameters() && p !== nothing
grad = (G, θ, p) -> f.grad(G, θ, p)
end
else
grad = nothing
end

if fg == true && f.fg !== nothing
if g == false
extras_grad = prepare_gradient(_f, adtype.dense_ad, x)
end
function fg!(res, θ)
(y, _) = value_and_gradient!(_f, res, adtype.dense_ad, θ, extras_grad)
return y
end
if p !== SciMLBase.NullParameters() && p !== nothing
function fg!(res, θ, p)
global p = p
(y, _) = value_and_gradient!(_f, res, adtype.dense_ad, θ)
return y
end
end
elseif fg == true
fg! = (G, θ) -> f.fg(G, θ, p)
if p !== SciMLBase.NullParameters() && p !== nothing
fg! = (G, θ, p) -> f.fg(G, θ, p)
end
else
fg! = nothing
end
Expand Down Expand Up @@ -398,7 +445,8 @@ function OptimizationBase.instantiate_function(
else
lag_h! = nothing
end
return OptimizationFunction{true}(f.f, adtype; grad = grad, hess = hess, hv = hv!,
return OptimizationFunction{true}(f.f, adtype;
grad = grad, fg = fg!, hess = hess, hv = hv!, fgh = fgh!,
cons = cons, cons_j = cons_j!, cons_h = cons_h!,
hess_prototype = hess_sparsity,
hess_colorvec = hess_colors,
Expand Down
60 changes: 55 additions & 5 deletions src/OptimizationDIExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,9 @@ import OptimizationBase.SciMLBase: OptimizationFunction
import OptimizationBase.LinearAlgebra: I
import DifferentiationInterface
import DifferentiationInterface: prepare_gradient, prepare_hessian, prepare_hvp,
prepare_jacobian,
prepare_jacobian, value_and_gradient!, value_and_gradient,
value_derivative_and_second_derivative!,
value_derivative_and_second_derivative,
gradient!, hessian!, hvp!, jacobian!, gradient, hessian,
hvp, jacobian
using ADTypes, SciMLBase
Expand All @@ -26,8 +28,9 @@ function instantiate_function(
g = false, h = false, hv = false, fg = false, fgh = false,
cons_j = false, cons_vjp = false, cons_jvp = false, cons_h = false,
lag_h = false)
global _p = p
function _f(θ)
return f(θ, p)[1]
return f(θ, _p)[1]
end

adtype, soadtype = generate_adtype(adtype)
Expand All @@ -37,19 +40,41 @@ function instantiate_function(
function grad(res, θ)
gradient!(_f, res, adtype, θ, extras_grad)
end
if p !== SciMLBase.NullParameters() && p !== nothing
function grad(res, θ, p)
global _p = p
gradient!(_f, res, adtype, θ)
end
end
elseif g == true
grad = (G, θ) -> f.grad(G, θ, p)
if p !== SciMLBase.NullParameters() && p !== nothing
grad = (G, θ, p) -> f.grad(G, θ, p)
end
else
grad = nothing
end

if fg == true && f.fg === nothing
if g == false
extras_grad = prepare_gradient(_f, adtype, x)
end
function fg!(res, θ)
(y, _) = value_and_gradient!(_f, res, adtype, θ, extras_grad)
return y
end
if p !== SciMLBase.NullParameters() && p !== nothing
function fg!(res, θ, p)
global _p = p
(y, _) = value_and_gradient!(_f, res, adtype, θ)
return y
end
end
elseif fg == true
fg! = (G, θ) -> f.fg(G, θ, p)
if p !== SciMLBase.NullParameters()
fg! = (G, θ, p) -> f.fg(G, θ, p)
end
else
fg! = nothing
end
Expand Down Expand Up @@ -196,7 +221,8 @@ function instantiate_function(
lag_h! = nothing
end

return OptimizationFunction{true}(f.f, adtype; grad = grad, hess = hess, hv = hv!,
return OptimizationFunction{true}(f.f, adtype;
grad = grad, fg = fg!, hess = hess, hv = hv!, fgh = fgh!,
cons = cons, cons_j = cons_j!, cons_h = cons_h!,
cons_vjp = cons_vjp!, cons_jvp = cons_jvp!,
hess_prototype = hess_sparsity,
Expand Down Expand Up @@ -232,8 +258,9 @@ function instantiate_function(
g = false, h = false, hv = false, fg = false, fgh = false,
cons_j = false, cons_vjp = false, cons_jvp = false, cons_h = false,
lag_h = false)
global _p = p
function _f(θ)
return f(θ, p)[1]
return f(θ, _p)[1]
end

adtype, soadtype = generate_adtype(adtype)
Expand All @@ -243,19 +270,41 @@ function instantiate_function(
function grad(θ)
gradient(_f, adtype, θ, extras_grad)
end
if p !== SciMLBase.NullParameters() && p !== nothing
function grad(θ, p)
global _p = p
gradient(_f, adtype, θ)
end
end
elseif g == true
grad = (θ) -> f.grad(θ, p)
if p !== SciMLBase.NullParameters() && p !== nothing
grad = (θ, p) -> f.grad(θ, p)
end
else
grad = nothing
end

if fg == true && f.fg === nothing
if g == false
extras_grad = prepare_gradient(_f, adtype, x)
end
function fg!(θ)
(y, res) = value_and_gradient(_f, adtype, θ, extras_grad)
return y, res
end
if p !== SciMLBase.NullParameters() && p !== nothing
function fg!(θ, p)
global _p = p
(y, res) = value_and_gradient(_f, adtype, θ)
return y, res
end
end
elseif fg == true
fg! = (θ) -> f.fg(θ, p)
if p !== SciMLBase.NullParameters() && p !== nothing
fg! = (θ, p) -> f.fg(θ, p)
end
else
fg! = nothing
end
Expand Down Expand Up @@ -387,7 +436,8 @@ function instantiate_function(
lag_h! = nothing
end

return OptimizationFunction{false}(f.f, adtype; grad = grad, hess = hess, hv = hv!,
return OptimizationFunction{false}(f.f, adtype;
grad = grad, fg = fg!, hess = hess, hv = hv!, fgh = fgh!,
cons = cons, cons_j = cons_j!, cons_h = cons_h!,
cons_vjp = cons_vjp!, cons_jvp = cons_jvp!,
hess_prototype = hess_sparsity,
Expand Down
Loading