lungd / ltc.jl Goto Github PK

Return the full sequence instead of the last time step only.
kwarg or separate struct/functor?

Implement some simple reinforcement learning examples.

cartpole, mountaincar, etc.

Find a way to automatically merge multiple consecutive MTKLayers resulting in a big system.
connected_sys.eqs = [ sys2.in1 ~ sys1.out1 ]?

Verify

• Creation of synapses
• Input/output mapping: equations connecting inpins/outpins with other components
• Number of parameters, lower/upper bounds
• Bounds generation (based on metadata).
• Computation of gradients (can serve as benchmark)

Create modules (or package?) for predefined MTK Systems to prevent naming conflicts
There could be an example package with separate dependencies

Should work with Float64s as well.
eltype(training_set)

Add connector equations based on custom metadata

Is this error package specific? Or maybe am I doing something wrong?

LoadError: Failed to precompile LTC [91bbf6af-ea09-4d53-996a-47377b466f23] to /home/solar/.julia/compiled/v1.6/LTC/jl_dPz0a8.
in expression starting at /home/solar/Desktop/julia-projects-new/LTC.jl/example/sine-flux.jl:1
error(s::String) at error.jl:33
compilecache(pkg::Base.PkgId, path::String, internal_stderr::Base.TTY, internal_stdout::Base.TTY) at loading.jl:1360
compilecache(pkg::Base.PkgId, path::String) at loading.jl:1306
_require(pkg::Base.PkgId) at loading.jl:1021
require(uuidkey::Base.PkgId) at loading.jl:914
require(into::Module, mod::Symbol) at loading.jl:901
eval at boot.jl:360 [inlined]
include_string(mapexpr::typeof(identity), mod::Module, code::String, filename::String) at loading.jl:1094

ERROR: LoadError: Failed to precompile LTC [91bbf6af-ea09-4d53-996a-47377b466f23] to /home/solar/.julia/compiled/v1.6/LTC/jl_qAOh87.
Stacktrace:
 [1] error(s::String)
   @ Base ./error.jl:33
 [2] compilecache(pkg::Base.PkgId, path::String, internal_stderr::Base.TTY, internal_stdout::Base.TTY)
   @ Base ./loading.jl:1360
 [3] compilecache(pkg::Base.PkgId, path::String)
   @ Base ./loading.jl:1306
 [4] _require(pkg::Base.PkgId)
   @ Base ./loading.jl:1021
 [5] require(uuidkey::Base.PkgId)
   @ Base ./loading.jl:914
 [6] require(into::Module, mod::Symbol)
   @ Base ./loading.jl:901
 [7] eval
   @ ./boot.jl:360 [inlined]
 [8] include_string(mapexpr::typeof(identity), mod::Module, code::String, filename::String)
   @ Base ./loading.jl:1094
in expression starting at /home/solar/Desktop/LTC.jl/example/sine-flux.jl:1

Reduce compilation time (of Zygote.gradient).
Type instabilities?

Metadata for parameters fixed during optimization ?
lower=upper=initial ?

Try out a system with batchisize independent unconnected subsystems (instances of the system) instead of solving an EnsembleProblem.

Currently, the system gets solved sequence_length times from 0 to 1 ( tspan=(0,1) )
Compare performance with a model which takes the complete sequence as input and solves the system from 0 to sequence_length.

Approaches for handling input:

1. Define a callback for changing corresponding parameters (prob.p[1:n_in])
2. Define input as a external input function (@register somehow ?)

"Default" ODESystem with an additional synapse defined by a PDE (or SDE #100)

Create some SDEs
Use DiffEqUncertainty.jl

Variables / parameters with units (Unitful.jl)

Check how to add NN to the ModelingToolkit system (UDEs)

modelingtoolkitize ?
parameter array ?

Data (num features, sequence length, batchsize, etc), Model (num params, etc), ODE (solver, sensealg, tspan, saveat, etc)