diff --git a/dev/.documenter-siteinfo.json b/dev/.documenter-siteinfo.json index 87c2b1c..1757184 100644 --- a/dev/.documenter-siteinfo.json +++ b/dev/.documenter-siteinfo.json @@ -1 +1 @@ -{"documenter":{"julia_version":"1.10.4","generation_timestamp":"2024-07-10T15:07:57","documenter_version":"1.1.0"}} \ No newline at end of file +{"documenter":{"julia_version":"1.10.4","generation_timestamp":"2024-07-10T15:25:31","documenter_version":"1.1.0"}} \ No newline at end of file diff --git a/dev/index.html b/dev/index.html index 5411545..b534929 100644 --- a/dev/index.html +++ b/dev/index.html @@ -1,4 +1,4 @@ -Home · NeuralDELux.jl

NeuralDELux

Documentation for NeuralDELux.

NeuralDELux.ADEulerStep — Type
ADEulerStep

Does a single Euler step, does work with GPUs and AD (tested with Zygote)

It's is called with solve(model, x, ps, st, solver::ADEulerStepTest, dt; kwargs...)

source
NeuralDELux.ADNeuralDE — Type
ADNeuralDE(; model=model, alg=ADRK4Step(), dt=dt, kwargs...)

Model for setting up and training Chaotic Neural Differential Equations with Lux.jl and NeuralDELux one-step solvers

Fields:

  • prob DEProblem
  • alg Algorithm to use for the solve command
  • dt time step
  • kwargs any additonal keywords

An instance of the model is called with a trajectory pair (t,x) in t are the timesteps that NDE is integrated for and x is a trajectory N x ... x N_t in which x[:, ... , 1] is taken as the initial condition.

source
NeuralDELux.ADRK4Step — Type
ADRK4Step

Does a single Runge Kutta 4th order step, does work with GPUs and AD (tested with Zygote)

It's is called with solve(model, x, ps, st, solver::ADEulerStepTest, dt; kwargs...)

source
NeuralDELux.ANODEForecastLength — Type
ANODEForecastLength(data; threshold::Number=0.4, metric="norm")

Provides an additonal metric to measure how well a model performs on data in term of its forecast error. The length is definded as the time step it takes until metric exceeds threshold. The initialized struct can then be called with 

fl =  ForecastLength(data)
-res = fl(model, ps, st)
source
NeuralDELux.AlternativeModelLossSingleSample — Type
AlternativeModelLossSingleSample(model, loss, data)

Computes the mean loss with model on data. data is supposed to serve as an iterator. data is assumed to be batched along the last dimension, but model only gets single samples as inputs.

source
NeuralDELux.AugmentedNeuralDE — Type
AugmentedNeuralDE(node_model::Union{ADNeuralDE, SciMLNeuralDE}, size_aug::Tuple, size_orig::Tuple, cat_dim)

Construct an augmented NODE that wraps around an exisiting node_model with observales with size size_orig and adds size_aug additional dimensions along dimension cat_dim.

source
NeuralDELux.ForecastLength — Type
ForecastLength(data; threshold::Number=0.4, modes=("forecast_length",), metric="norm", N_avg::Int=30)

Provides an additonal metric to measure how well a model performs on data in term of its forecast error. The length is definded as the time step it takes until metric exceeds threshold. The initialized struct can then be called with 

fl =  ForecastLength(data)
-res = fl(model, ps, st)
source
NeuralDELux.SciMLEulerStep — Type
SciMLEulerStep

Does one Euler step using direct AD. Expected to be used like a solver algorithm from OrdinaryDiffEq.jl, so with solve(prob::AbstractDEProblem, ADEulerStep()).

source
NeuralDELux.SciMLNeuralDE — Type
SciMLNeuralDE(model; alg=ADEulerStep(), gpu=nothing, kwargs...)

Model for setting up and training Chaotic Neural Differential Equations with Lux.jl and SciMLSensitivity.jl

Fields:

  • prob DEProblem
  • alg Algorithm to use for the solve command
  • kwargs any additional keyword arguments that should be handed over (e.g. sensealg)
  • device the device the model is running on, either DeviceCPU or DeviceCUDA, used for dispatiching if Arrays or CuArrays are used

An instance of the model is called with a trajectory pair (t,x) in t are the timesteps that NDE is integrated for and x is a trajectory N x ... x N_t in which x[:, ... , 1] is taken as the initial condition.

source
NeuralDELux.DetermineDevice — Method
DetermineDevice(; gpu::Union{Nothing, Bool}=nothing)

Initializes the device that is used. Returns either DeviceCPU or DeviceCUDA. If no gpu keyword argument is given, it determines automatically if a GPU is available.

source
NeuralDELux.SamePadCircularConv — Function
SamePadCircularConv(kernel, ch, activation=identity)

Wrapper around Lux.Conv that adds circular padding so that the dimensions stay the same.

source
NeuralDELux.evolve — Method
evolve(model::ADNeuralDE, ps, st, ic; tspan::Union{T, Nothing}=nothing, N_t::Union{Integer,Nothing}=nothing) where T

Evolve the model by tspan or N_t (only specifiy one), starting from the initial condition ic

source
NeuralDELux.evolve — Method
evolve(model::SciMLNeuralDE, ps, st, ic; tspan::Union{T, Nothing}=nothing, N_t::Union{Integer,Nothing}=nothing) where T

Evolve the model by tspan or N_t (only specifiy one), starting from the initial condition ic

source
NeuralDELux.evolve_to_blowup — Method
evolve_to_blowup(singlestep_solver, x, ps, st, dt, default_time=Inf)

Integrated a longer trajectory from a (trained) single step solver until it blows up.

source
NeuralDELux.evolve_to_blowup — Method
evolve_to_blowup(model::SciMLNeuralDE, ps, st, ic::A; default_time=Inf, kwargs...)

Evolves a model that is suspected to blowup and returns the last time step if that is the case, and if not returns default_time

source
NeuralDELux.forecast_δ — Function
forecast_δ(prediction::AbstractArray{T,N}, truth::AbstractArray{T,N}, mode::String="both") where {T,N}

Assumes that the last dimension of the input arrays is the time dimension and N_t long. Returns an N_t long array, judging how accurate the prediction is.

Supported modes:

  • "mean": mean between the arrays
  • "maximum": maximum norm
  • "norm": normalized, similar to the metric used in Pathak et al
source
NeuralDELux.train! — Method
model, ps, st, training_results = train!(model, ps, st, loss, train_data, opt_state, η_schedule; τ_range=2:2, N_epochs=1, verbose=true, save_name=nothing, save_results_name=nothing, shuffle_data_order=true, additional_metric=nothing, valid_data=nothing, test_data=nothing, scheduler_offset::Int=0, compute_initial_error::Bool=true, save_mode::Symbol=:valid)

Trains the model with parameters ps and state st with the loss function and train_data by applying a opt_state with the learning rate η_schedule for N_epochs. Returns the trained model, ps, st, results. An additional_metric with the signature (model, ps, st) -> value might be specified that is computed after every epoch. save_mode determines if the model is saved with the lowest error on the :valid set or :train set.

source
NeuralDELux.train_anode! — Method
model, ps, st, training_results = train_anode!(model, ps, st, loss, train_data, opt_state, η_schedule; τ_range=2:2, N_epochs=1, verbose=true, save_name=nothing, additional_metric=nothing)

Trains the model with parameters ps and state st with the loss function and train_data by applying a opt_state with the learning rate η_schedule for N_epochs. Returns the trained model, ps, st, results. An additional_metric with the signature (model, ps, st) -> value might be specified that is computed after every epoch.

source
NeuralDELux.trajectory — Method
trajectory(singlestep_solver, x, ps, st)

Integrated a longer trajectory from a (trained) single step solver. Not implemented for AD / training.

source
+Home · NeuralDELux.jl

NeuralDELux

Documentation for NeuralDELux.

NeuralDELux.ADEulerStep — Type
ADEulerStep

Does a single Euler step, does work with GPUs and AD (tested with Zygote)

It's is called with solve(model, x, ps, st, solver::ADEulerStepTest, dt; kwargs...)

source
NeuralDELux.ADNeuralDE — Type
ADNeuralDE(; model=model, alg=ADRK4Step(), dt=dt, kwargs...)

Model for setting up and training Chaotic Neural Differential Equations with Lux.jl and NeuralDELux one-step solvers

Fields:

  • prob DEProblem
  • alg Algorithm to use for the solve command
  • dt time step
  • kwargs any additonal keywords

An instance of the model is called with a trajectory pair (t,x) in t are the timesteps that NDE is integrated for and x is a trajectory N x ... x N_t in which x[:, ... , 1] is taken as the initial condition.

source
NeuralDELux.ADRK4Step — Type
ADRK4Step

Does a single Runge Kutta 4th order step, does work with GPUs and AD (tested with Zygote)

It's is called with solve(model, x, ps, st, solver::ADEulerStepTest, dt; kwargs...)

source
NeuralDELux.ANODEForecastLength — Type
ANODEForecastLength(data; threshold::Number=0.4, metric="norm")

Provides an additonal metric to measure how well a model performs on data in term of its forecast error. The length is definded as the time step it takes until metric exceeds threshold. The initialized struct can then be called with 

fl =  ForecastLength(data)
+res = fl(model, ps, st)
source
NeuralDELux.AlternativeModelLossSingleSample — Type
AlternativeModelLossSingleSample(model, loss, data)

Computes the mean loss with model on data. data is supposed to serve as an iterator. data is assumed to be batched along the last dimension, but model only gets single samples as inputs.

source
NeuralDELux.AugmentedNeuralDE — Type
AugmentedNeuralDE(node_model::Union{ADNeuralDE, SciMLNeuralDE}, size_aug::Tuple, size_orig::Tuple, cat_dim)

Construct an augmented NODE that wraps around an exisiting node_model with observales with size size_orig and adds size_aug additional dimensions along dimension cat_dim.

source
NeuralDELux.ForecastLength — Type
ForecastLength(data; threshold::Number=0.4, modes=("forecast_length",), metric="norm", N_avg::Int=30)

Provides an additonal metric to measure how well a model performs on data in term of its forecast error. The length is definded as the time step it takes until metric exceeds threshold. The initialized struct can then be called with 

fl =  ForecastLength(data)
+res = fl(model, ps, st)
source
NeuralDELux.SciMLEulerStep — Type
SciMLEulerStep

Does one Euler step using direct AD. Expected to be used like a solver algorithm from OrdinaryDiffEq.jl, so with solve(prob::AbstractDEProblem, ADEulerStep()).

source
NeuralDELux.SciMLNeuralDE — Type
SciMLNeuralDE(model; alg=ADEulerStep(), gpu=nothing, kwargs...)

Model for setting up and training Chaotic Neural Differential Equations with Lux.jl and SciMLSensitivity.jl

Fields:

  • prob DEProblem
  • alg Algorithm to use for the solve command
  • kwargs any additional keyword arguments that should be handed over (e.g. sensealg)
  • device the device the model is running on, either DeviceCPU or DeviceCUDA, used for dispatiching if Arrays or CuArrays are used

An instance of the model is called with a trajectory pair (t,x) in t are the timesteps that NDE is integrated for and x is a trajectory N x ... x N_t in which x[:, ... , 1] is taken as the initial condition.

source
NeuralDELux.DetermineDevice — Method
DetermineDevice(; gpu::Union{Nothing, Bool}=nothing)

Initializes the device that is used. Returns either DeviceCPU or DeviceCUDA. If no gpu keyword argument is given, it determines automatically if a GPU is available.

source
NeuralDELux.SamePadCircularConv — Function
SamePadCircularConv(kernel, ch, activation=identity)

Wrapper around Lux.Conv that adds circular padding so that the dimensions stay the same.

source
NeuralDELux.evolve — Method
evolve(model::ADNeuralDE, ps, st, ic; tspan::Union{T, Nothing}=nothing, N_t::Union{Integer,Nothing}=nothing) where T

Evolve the model by tspan or N_t (only specifiy one), starting from the initial condition ic

source
NeuralDELux.evolve — Method
evolve(model::SciMLNeuralDE, ps, st, ic; tspan::Union{T, Nothing}=nothing, N_t::Union{Integer,Nothing}=nothing) where T

Evolve the model by tspan or N_t (only specifiy one), starting from the initial condition ic

source
NeuralDELux.evolve_to_blowup — Method
evolve_to_blowup(singlestep_solver, x, ps, st, dt, default_time=Inf)

Integrated a longer trajectory from a (trained) single step solver until it blows up.

source
NeuralDELux.evolve_to_blowup — Method
evolve_to_blowup(model::SciMLNeuralDE, ps, st, ic::A; default_time=Inf, kwargs...)

Evolves a model that is suspected to blowup and returns the last time step if that is the case, and if not returns default_time

source
NeuralDELux.forecast_δ — Function
forecast_δ(prediction::AbstractArray{T,N}, truth::AbstractArray{T,N}, mode::String="both") where {T,N}

Assumes that the last dimension of the input arrays is the time dimension and N_t long. Returns an N_t long array, judging how accurate the prediction is.

Supported modes:

  • "mean": mean between the arrays
  • "maximum": maximum norm
  • "norm": normalized, similar to the metric used in Pathak et al
source
NeuralDELux.train! — Method
model, ps, st, training_results = train!(model, ps, st, loss, train_data, opt_state, η_schedule; τ_range=2:2, N_epochs=1, verbose=true, save_name=nothing, save_results_name=nothing, shuffle_data_order=true, additional_metric=nothing, valid_data=nothing, test_data=nothing, scheduler_offset::Int=0, compute_initial_error::Bool=true, save_mode::Symbol=:valid)

Trains the model with parameters ps and state st with the loss function and train_data by applying a opt_state with the learning rate η_schedule for N_epochs. Returns the trained model, ps, st, results. An additional_metric with the signature (model, ps, st) -> value might be specified that is computed after every epoch. save_mode determines if the model is saved with the lowest error on the :valid set or :train set.

source
NeuralDELux.train_anode! — Method
model, ps, st, training_results = train_anode!(model, ps, st, loss, train_data, opt_state, η_schedule; τ_range=2:2, N_epochs=1, verbose=true, save_name=nothing, additional_metric=nothing)

Trains the model with parameters ps and state st with the loss function and train_data by applying a opt_state with the learning rate η_schedule for N_epochs. Returns the trained model, ps, st, results. An additional_metric with the signature (model, ps, st) -> value might be specified that is computed after every epoch.

source
NeuralDELux.trajectory — Method
trajectory(singlestep_solver, x, ps, st)

Integrated a longer trajectory from a (trained) single step solver. Not implemented for AD / training.

source