The tensorsem from frozen-jelly

tensorsem's Introduction

Structural Equation Modeling using Torch

Description

An R and python package for structural equation modeling using Torch. This package is meant for researchers who know their way around SEM, Torch, and lavaan. Structural equation modeling is implemented as a fully functional torch.nn module. A short example optimization loop would be:

import tensorsem as ts
model = ts.StructuralEquationModel(opt = opts)  # opts are of class SemOptions
optim = torch.optim.Adam(model.parameters())  # init the optimizer
for epoch in range(1000):
    optim.zero_grad()  # reset the gradients of the parameters
    Sigma = model()  # compute the model-implied covariance matrix
    loss = ts.mvn_negloglik(dat, Sigma)  # compute the negative log-likelihood, dat tensor should exist
    loss.backward()  # compute the gradients and store them in the parameter tensors
    optim.step()  # take a step in the negative gradient direction using adam

Installation

To install the latest version of tensorsem, run the following:

Install the R interface package from this repository:
```
remotes::install_github("vankesteren/tensorsem")
```
Install pytorch on your system. Use the pytorch website to do this. For example, for a windows pip cpu version, use:
```
pip install torch==1.5.0+cpu torchvision==0.6.0+cpu -f https://download.pytorch.org/whl/torch_stable.html
```

Install the tensorsem python package from this repository.

pip install https://github.com/vankesteren/tensorsem/archive/master.zip

(Optional) Install pandas and matplotlib for plotting and parameter storing
```
pip install matplotlib pandas
```

Usage

See the example directory for a full usage example, estimating the Holzinger-Swineford model using maximum likelihood, unweighted least squares, and diagonally weighted least squares.

Recommend Projects