DeepSimulatedMoments

Documentation	Build Status

DeepSimulatedMoments.jl provides an implementation of the methods proposed in Constructing Efficient Simulated Moments Using Temporal Convolutional Networks by Chassot, J. and Creel, M. (2023). The package allows to define your own data-generating processes, set up and train neural networks to generate moment conditions and proceed with simulation-based inference. For information on how to use this package, please refer to the documentation

Quick Example

using DeepSimulatedMoments
using Flux # Flux provides the optimizer used in this example, ADAMW

# Create a moving-average process of order 2 with n=100 observations
dgp = MA2(100)
# Build a TCN to generate moment conditions for this DGP
tcn = build_tcn(dgp)

# Set up hyperparameters
hp = HyperParameters(
    validation_size=1_000, # Use 1'000 samples to validate the final network
    loss=rmse_conv, # The loss function to use in the training of the network
    nsamples=100, # Number of samples (of `n=100` observations) per epoch
    epochs=5, # Number of total epochs
    print_every=5, # Print train/test loss every 5 samples
    dev=cpu # Use the CPU as a device for the network
)

# Create the moment network
net = MomentNetwork(
    tcn |> hp.dev, # Specify the network to use and pass the TCN to the device
    hp, # Specify the hyperparameters used for training and validation
    ADAMW(), # Specify the optimizer used for training
    # Specify a transformation applied to the parameters of the DGP pre-training
    parameter_transform=datatransform(dgp, 100_000, dev=hp.dev)
)