This is repository that allows one to estimate macro models with the simulation based inference (SBI) library
The code is straightforward to run. The main function is sbimacro.sbi_macro(). You run it like this:
from sbiwrapper.sbimacro import sbi_macro
import torch
def testfunct(params):
xout = params[0]*torch.rand(100).reshape((1,-1))
return xout
xin = testfunct(torch.tensor([10], dtype= torch.float32)).numpy()
prior1 = torch.distributions.Uniform(torch.tensor([0.0]),torch.tensor([20.0]))
boundmin1 = [.00001]
boundmax1 = [19.99999]
samps, post = sbi_macro(xin, boundmin1, boundmax1, prior1, testfunct, netparams = None, num_rounds = 10,
method = 'SNPE', folder = None, init_simulations = 30, round_simulations = 25,
numworkers = None, batch_size = 1000)
Note the init_simulations and the round_simulations are abnormally low for speed reasons. These numbers should be in the 1000s at least.
This package has been lightly tested. Feel free to report any errors here.