Performance question

[itsdfish]

Hi,

A few months ago, I opened an issue in DiffEqFlux about using normalizing flows to learn likelihood functions from simulated data. I noticed you have been working on this package which seems to serve a similar purpose. Out of curiosity, I tried to run the example in the README file (commit 82149d2), but stopped the program after it had been running for more than an hour. Would you expect this algorithm to require so much time for a simple problem?

[prbzrg]

Hi,

There are some performance issues, For now, we can use compute_mode=ZygoteMatrixMode and adtype=Optimization.AutoForwardDiff():

using ICNF
using Distributions, Lux
using DifferentialEquations, SciMLSensitivity
using Optimization, ForwardDiff

# Parameters
nvars = 1
n = 1024

# Data
data_dist = Beta(2.0f0, 4.0f0)
r = rand(data_dist, nvars, n)

# Model
nn = Chain(Dense(nvars => 4 * nvars, tanh), Dense(4 * nvars => nvars, tanh))
icnf = construct(RNODE, nn, nvars; tspan = (0.0f0, 4.0f0), compute_mode=ZygoteMatrixMode)

# Training
using DataFrames, MLJBase
df = DataFrame(transpose(r), :auto)
model = ICNFModel(icnf; adtype=Optimization.AutoForwardDiff())
mach = machine(model, df)
fit!(mach)
ps, st = fitted_params(mach)

# Use It
d = ICNFDist(icnf, ps, st)
actual_pdf = pdf.(data_dist, vec(r))
estimated_pdf = pdf(d, r)
new_data = rand(d, n)

# Evaluation
using Distances
mad_ = meanad(estimated_pdf, actual_pdf)
msd_ = msd(estimated_pdf, actual_pdf)
tv_dis = totalvariation(estimated_pdf, actual_pdf) / n

[itsdfish]

Thank you for your reply. I will try that.