Link to DifferentialEquations.jl?

[ChrisRackauckas]

Cool project! I think it would be cool if simulation-based methods could be supported here by hooking into DifferentialEquations.jl and its DiffEqFlux fast adjoint schemes (https://diffeqflux.sciml.ai/dev/). Usually this ends up being a bit faster than discrete-then-optimize techniques (when utilizing automatic differentiation mixed adjoints), so it would be a good technique to mix in.

[sshin23]

Thanks for the great suggestion @ChrisRackauckas

It would be nice to have an option to formulate the problem as a simulation-based problem using adjoint evaluation capability. And I think DifferentialEquations.jl would be the most convenient package for doing this 🙂 We can't promise anything at this point, but this seems like a great long-term plan.

Do you have an example code of hooking up DifferentialEquations.jl with NLP solvers via JuMP?

[ChrisRackauckas]

https://diffeq.sciml.ai/stable/analysis/parameter_estimation/#Using-JuMP-with-DiffEqParamEstim is an example, but that will use forward-mode AD so it won't scale to huge problems as well. Mixing it with https://diffeq.sciml.ai/stable/analysis/sensitivity/#solve-Differentiation-Examples for the gradient definition is what you'd want.

[paulflang]

Thanks a lot for your interest, @ChrisRackauckas. I think your suggestion on Twitter to use a two step approach could make SBML2Julia much more powerful. The current Julia code output of SBML2Julia looks like the tests/fixtures/jl_code_gold.jl file. Do you think you would be able to contribute fixtures tests/fixtures/mtk_code_gold.jl (or tests/fixtures/catalyst_code_gold.jl), tests/fixtures/JuMP_code_gold.jl, tests/fixtures/optim_code_gold.jl, etc. which solve the optimization problem specified in tests/fixtures/jl_code_gold.jl to the feature_mtk branch? You will probably see from the code structure that I know very little Julia, so please feel free to make improvements. I could then take it from there and modify the Python part to auto generate MTK, JuMP, optim etc. models.

[paulflang]

I am just trying to optimize a simple ODE system specified in MTK with JuMP:

using ModelingToolkit
using OrdinaryDiffEq, DiffEqParamEstim, JuMP, NLopt

t_max = 10
dt = 0.1

t_out = 0:dt:t_max
tspan = (0,t_max)

@ModelingToolkit.parameters t, k1, k2
@ModelingToolkit.variables A(t) B(t)
@ModelingToolkit.derivatives D'~t

parameters = [k1 => 0.8, k2 => 0.6]
initial_conditions = [A => 1,
                      B => 0]
eqs = [D(A) ~ -k1*A + k2*B,
       D(B) ~ k1*A - k2*B]
sys = ODESystem(eqs)

model_ode(p_) = OrdinaryDiffEq.ODEProblem(sys,initial_conditions,tspan,p_,jac=true)
solve_model(mp_) = OrdinaryDiffEq.solve(model_ode(mp_), Tsit5(),saveat=dt)
mock_data = Array(solve_model([0.8, 0.6]))
loss_objective(mp_, dat) = build_loss_objective(model_ode(mp_), Tsit5(), L2Loss(t_out,dat))

juobj(args...) = loss_objective(args, mock_data)(args)
jumodel = Model()
JuMP.register(jumodel, :juobj, 2, juobj, autodiff=true)
@JuMP.variable(jumodel, k1, start=1)
@JuMP.variable(jumodel, k2, start=1)
@JuMP.NLobjective(jumodel, Min, juobj(k1, k2))
set_optimizer(jumodel, NLopt.Optimizer)
set_optimizer_attribute(jumodel, "algorithm", :LD_MMA)
JuMP.optimize!(jumodel)

Unfortunately, I get the following error: ERROR: MethodError: no method matching lower_mapnames(::Tuple{Float64,Float64}) I am using Julia v1.5.3 and just added the used packages yesterday. Do you have ideas how to resolve this?

[ChrisRackauckas]

JuMP turns parameters into tuples, so you probably don't want to use JuMP if the number of parameters gets large. That said, here's a working version of that code:

using ModelingToolkit
using OrdinaryDiffEq, DiffEqParamEstim, JuMP, NLopt

t_max = 10.0
dt = 0.1

t_out = 0:dt:t_max
tspan = (0,t_max)

@ModelingToolkit.parameters t, k1, k2
@ModelingToolkit.variables A(t) B(t)
@ModelingToolkit.derivatives D'~t

parameters = [k1 => 0.8, k2 => 0.6]
initial_conditions = [A => 1,
                      B => 0]
eqs = [D(A) ~ -k1*A + k2*B,
       D(B) ~ k1*A - k2*B]
sys = ODESystem(eqs)

model_ode(p_) = OrdinaryDiffEq.ODEProblem(sys,initial_conditions,tspan,p_,jac=true)
solve_model(mp_) = OrdinaryDiffEq.solve(model_ode(mp_), Tsit5(),saveat=dt)
mock_data = Array(solve_model([0.8, 0.6]))
loss_objective(mp_, dat) = build_loss_objective(model_ode(collect(mp_)), Tsit5(), L2Loss(t_out,dat))

juobj(args...) = loss_objective(args, mock_data)(args)
jumodel = Model()
JuMP.register(jumodel, :juobj, 2, juobj, autodiff=true)
@JuMP.variable(jumodel, k1, start=1)
@JuMP.variable(jumodel, k2, start=1)
@JuMP.NLobjective(jumodel, Min, juobj(k1, k2))
set_optimizer(jumodel, NLopt.Optimizer)
set_optimizer_attribute(jumodel, "algorithm", :LD_MMA)
JuMP.optimize!(jumodel)

The other change was to make time floating point valued, since integer-valued time usually isn't a great idea. Hopefully that helps!

[paulflang]

Thanks! That helped a lot.

[paulflang]

I cannot find a way to get multiple experimental condition to work. What I tried for instance is:

mock_data_c1 = Array(solve_model([0.8, 0.6]))
mock_data_c2 = Array(solve_model([0.8, 0.5]))
loss_objective(mp_, dat) = build_loss_objective(model_ode(collect(mp_)), Tsit5(), L2Loss(t_out,dat))

juobj(args1, args2) = loss_objective(args1, mock_data1)(args1) + loss_objective(args2, mock_data2)(args2)

jumodel = Model()
JuMP.register(jumodel, :juobj, 2, juobj, autodiff=true)
@JuMP.variable(jumodel, k1_1, start=1)
@JuMP.variable(jumodel, k2_1, start=1)
@JuMP.variable(jumodel, k1_2, start=1)
@JuMP.variable(jumodel, k2_2, start=1)
@JuMP.NLobjective(jumodel, Min, juobj((k1_1, k2_1), (k1_2, k2_2)))

But I get the following error:

ERROR: Unexpected object (k1, k2) (of type Tuple{VariableRef,VariableRef} in nonlinear expression.
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] _parse_NL_expr_runtime(::Model, ::Tuple{VariableRef,VariableRef}, ::Array{JuMP._Derivatives.NodeData,1}, ::Int64, ::Array{Float64,1}) at /home/paul/.julia/packages/JuMP/qhoVb/src/parse_nlp.jl:223
 [3] top-level scope at /home/paul/.julia/packages/JuMP/qhoVb/src/parse_nlp.jl:115
 [4] top-level scope at /home/paul/.julia/packages/JuMP/qhoVb/src/macros.jl:1368
 [5] top-level scope at REPL[23]:1

If anyone knows how to solve this, please let me know.

[ChrisRackauckas]

I would ask on #jump-dev as it's more of a JuMP question.

[paulflang]

Oh, I see. Thanks. I actually wanted to write an objective function that includes multiple experimental conditions and could be fed to any optimizer, not only JuMP. Is there really no way this can be done using build_loss_objective?

[ChrisRackauckas]

build_loss_objective isn't the issue here (and you could / probably should use DiffEqFlux.jl's more expressive form anyways). The issue is that is an ill-formed JuMP expression, so you couldn't write that with one experimental condition. As to why it says it's ill-formed, that's a JuMP question and you might want to ask a JuMP dev.

[paulflang]

Little update from my side. I tried to do the first step of the 2-step approach and wrote SbmlInterface.jl, which converts SBML models to MTK and simulates them. Thanks a lot for your help on slack, @ChrisRackauckas . Next plan is to interface with optimizers.

[ChrisRackauckas]

That is fantastic!