# Error in building run time function by Lux.jl model with other equations

**URL:** <https://discourse.julialang.org/t/error-in-building-run-time-function-by-lux-jl-model-with-other-equations/116615>\
**Category:** General Usage\
**Tags:** symbolics, lux\
**Created:** [July 4, 2024, 9:25am UTC](https://discourse.julialang.org/t/error-in-building-run-time-function-by-lux-jl-model-with-other-equations/116615 "2024-07-04T09:25:07Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![chooron](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/chooron/32/208300_2.png) [@chooron](https://discourse.julialang.org/u/chooron)\
**Post date:** [July 4, 2024, 9:25am UTC](https://discourse.julialang.org/t/error-in-building-run-time-function-by-lux-jl-model-with-other-equations/116615/1 "2024-07-04T09:25:07Z")

</div>

I want to build an Equation using a neural network (based on Lux.jl) as follows:

```julia
using Lux
using LuxCore
using Symbolics
using StableRNGs
using ModelingToolkit
using ModelingToolkitNeuralNets
using ModelingToolkit:t_nounits as t
using ComponentArrays
using ModelingToolkitStandardLibrary.Blocks: RealInputArray

nn = Lux.Chain(
    Lux.Dense(3 => 16), # , Lux.tanh
    Lux.Dense(16 => 16), # , Lux.leakyrelu
    Lux.Dense(16 => 2)#, Lux.leakyrelu
)

init_params = Lux.initialparameters(StableRNG(42), nn)
@parameters p[1:length(init_params)] = Vector(ComponentVector(init_params))
@parameters ptype::typeof(typeof(init_params)) = typeof(init_params) [tunable = false]
lazyconvert_p = Symbolics.array_term(convert, ptype, p, size=size(p))

@variables v1(t), v2(t), v3(t)
exprs = LuxCore.stateless_apply(nn, [v1, v2, v3], lazyconvert_p)[1]

```

This code is written according to the ModelingToolkitNeuralNets.jl reference, and then I want to build a RuntimeGeneratedFunction based on this exprs, the result is as follows:

```julia
temp_func = build_function(exprs, [v1, v2, v3], [p, ptype], expression=Val{false})
>RuntimeGeneratedFunction(#=in Symbolics=#, #=using Symbolics=#, :((ˍ₋arg1, ˍ₋arg2)->begin
          #= D:\Julia\Julia-1.10.0\packages\packages\SymbolicUtils\c0xQb\src\code.jl:373 =#
          #= D:\Julia\Julia-1.10.0\packages\packages\SymbolicUtils\c0xQb\src\code.jl:374 =#
          #= D:\Julia\Julia-1.10.0\packages\packages\SymbolicUtils\c0xQb\src\code.jl:375 =#
          begin
              (getindex)((LuxCore.stateless_apply)(Chain(), Num[v1(t), v2(t), v3(t)], (convert)(ˍ₋arg2[2], ˍ₋arg2[1])), 1)
          end
      end))

```

As can be seen from the results, the input variables [v1, v2, v3] are not replaced by the parameter arg.  
In addition, when this function is called with actual data, parameter type conversion cannot convert model parameters through boardcast:

```julia
temp_func([1,2,3], [ComponentVector(init_params), eltype(ComponentVector(init_params))])

>ERROR: MethodError: Cannot `convert` an object of type ComponentVector{Float32, Vector{Float32}, Tuple{Axis{(layer_1 = ViewAxis(1:64, Axis(weight = ViewAxis(1:48, ShapedAxis((16, 3))), bias = ViewAxis(49:64, ShapedAxis((16, 1))))), layer_2 = ViewAxis(65:98, Axis(weight = ViewAxis(1:32, ShapedAxis((2, 16))), bias = ViewAxis(33:34, ShapedAxis((2, 1))))))}}} to an object of type Float32

```

---

<div class="post-metadata">

**Author:** ![SebastianM-C](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/sebastianm-c/32/2480_2.png) [@SebastianM-C](https://discourse.julialang.org/u/SebastianM-C)\
**Post date:** [July 7, 2024, 1:42am UTC](https://discourse.julialang.org/t/error-in-building-run-time-function-by-lux-jl-model-with-other-equations/116615/2 "2024-07-07T01:42:01Z")

</div>

As pointed out in [unexpect result when use build\_function for Lux.jl-based expr · Issue #1187 · JuliaSymbolics/Symbolics.jl · GitHub](https://github.com/JuliaSymbolics/Symbolics.jl/issues/1187), the major differences from ModelingToolkitNeuralNets.jl are that the input needs to be a symbolic vector, not a vector of symbols and the parameter type (`ptype`) is the type of the `ComponentVector(init_params)` not the `NamedTuple` type.

The error that you get is due to calling the generated function with `eltype` instead of `typeof`. More specifically, the error is due to trying to convert `ComponentVector(init_params)` to its element type, which is a `Float32`.

---

<div class="post-metadata">

**Author:** ![chooron](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/chooron/32/208300_2.png) [@chooron](https://discourse.julialang.org/u/chooron)\
**Post date:** [July 8, 2024, 12:54pm UTC](https://discourse.julialang.org/t/error-in-building-run-time-function-by-lux-jl-model-with-other-equations/116615/3 "2024-07-08T12:54:36Z")

</div>

After receiving your suggestion, I realized that by changing the input to `@variables (v(t))[1:4]` in this way, I can ensure the code runs normally. However, if I use this expression, I won’t be able to substitute the individual elements of the vector `v` (which are obtained from other expressions) using the `substitute` method. Is there a possible way to replace `v` with `[v1, v2, v3, v4]` so that I can get the resulting Function?

---

<div class="post-metadata">

**Author:** ![SebastianM-C](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/sebastianm-c/32/2480_2.png) [@SebastianM-C](https://discourse.julialang.org/u/SebastianM-C)\
**Post date:** [July 15, 2024, 3:20pm UTC](https://discourse.julialang.org/t/error-in-building-run-time-function-by-lux-jl-model-with-other-equations/116615/4 "2024-07-15T15:20:06Z")

</div>

Does adding additional equations work? For example

```julia
v[1] ~ v1
v[2] ~ v2
...

```

---

<div class="post-metadata">

**Author:** ![chooron](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/chooron/32/208300_2.png) [@chooron](https://discourse.julialang.org/u/chooron)\
**Post date:** [July 16, 2024, 1:50am UTC](https://discourse.julialang.org/t/error-in-building-run-time-function-by-lux-jl-model-with-other-equations/116615/5 "2024-07-16T01:50:37Z")

</div>

Thanks for your suggestions! After many attempts, I found that the build\_function of Symbolics.jl could not meet the problem I raised. However, I tried SymbolicUtils.jl later and it met my needs by using Func, Assignment to build functions.