01 Get started

1) Installation and Technical Requirements

Introduction

Several packages allow users to use machine learning directly in R such as nnet for single layer neural nets, rpart for decision trees, and ranger for random forests. Furthermore, with mlr3verse a series of packages exists for managing different algorithms with a unified interface.

These packages can be used with a ‘normal’ computer and provide an easy installation. In terms of natural language processing, these approaches are currently limited. State-of-the-art approaches rely on neural nets with multiple layers and consist of a huge number of parameters making them computationally demanding. With specialized libraries such as keras, PyTorch and tensorflow, graphical processing units (gpu) can help to speed up computations significantly. However, many of these specialized libraries for machine learning are written in python. Fortunately, an interface to python is provided via the R package reticulate.

The R package Artificial Intelligence for Education (aifeducation) aims to provide educators, educational researchers, and social researchers a convincing interface to these state-of-the-art models for natural language processing and tries to address the special needs and challenges of the educational and social sciences. The package currently supports the application of Artificial Intelligence (AI) for tasks such as text embedding, classification, and question answering.

Since state-of-the-art approaches in natural language processing rely on large models compared to classical statistical methods (e.g., latent class analysis, structural equation modeling) and are based largely on python, some additional installation steps are necessary.

If you would like to train and to develop your own models and AIs, a compatible graphic device is necessary. Even a low performing graphic device can speed up computations significantly. If you prefer using pre-trained models however, this is not necessary. In this case a ‘normal’ office computer without a graphic device should be sufficient in most cases.

Step 1 - Install the R Package

In order to use the package, you first need to install it. This can be done by:

install.packages("aifeducation")

With this command, all necessary R packages are installed on your machine.

Step 2 - Install Python

Since natural language processing with neural nets is based on models which are computationally intensive, keras, PyTorch, and tensorflow are used within this package together with some other specialized python libraries. To install them, you need to install python on your machine first. This may take some time.

reticulate::install_python()

You can check if everything is working by using the function reticulate::py_available(). This should return TRUE.

reticulate::py_available(initialize = TRUE)

Step 3 - Install Miniconda

The next step is to install miniconda since aifeducation uses conda environments for managing the different modules.

reticulate::install_miniconda()

Step 4 - Install Support for Graphic Devices

PyTorch and tensorflow as underlying machine learning backend run on MacOS, Linux, and Windows. However, there are some limitations for accelerate computations with graphic cards. The following table provides an overview.

Table: Possible gpu acceleration by operating system

Operating System PyTorch tensorflow
MacOS No No
Linux Yes Yes
Windows Yes Only with versions <=2.10
Windows with WSL Yes Yes

If you have a suitable machine and would like to use a graphic card for computations you need to install some further software. If not you can skip this step. A list with links to downloads can be found here if you would like to use tensorflow as machine learning framework: https://www.tensorflow.org/install/pip#linux

If you would like to use PyTorch as framework you can find further information here: https://pytorch.org/get-started/locally/

In general you need

Except the gpu drivers all components will be installed in step 5 automatically.

If you would like to use Windows with WSL (Windows Subsystem for Linux) installing gpu acceleration is a more complex topic. In this case please refer to the specific Windows or Ubuntu documentations.

Step 5 - Install Specialized Python Libraries

If everything is working, you can now install the remaining python libraries. For convenience, aifeducation comes with an auxiliary function install_py_modules() doing that for you.

#For Linux
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=FALSE,
                                 tf_version="<=2.14",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=FALSE)

#For Windows with gpu support
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=FALSE,
                                 tf_version="<=2.10",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=FALSE)
#For Windows without gpu support
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=FALSE,
                                 tf_version="<=2.14",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=TRUE)

#For MacOS
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=FALSE,
                                 tf_version="<=2.14",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=TRUE)

With install="all" you can decide which machine learning framework should be installed. Use install="all" to request the installation of both ‘PyTorch’ and ‘tensorflow’. If you would like to install only ‘PyTorch’ or ‘tensorflow’ set install="pytorch" or install="tenorflow". Starting with version 0.3.1 tensorflow of version 2.13 or higher is necessary.

It is very important that you call this function before loading the package the first time. If you load the library without installing the necessary modules an error may occur.

This function installs the following python modules:

both frameworks: - transformers, - tokenizers, - datasets, - codecarbon

Pytorch - torch, - torcheval, - safetensors, - accelerate

Tensorflow - keras, - tensorflow

and its dependencies in the environment “aifeducation”.

If you would like to use aifeducation with other packages or within other environments, please ensure that these python modules are available. For gpu support some further packages are installed.

With check_aif_py_modules() you can check, if all modules are successfully installed or a specific machine learning framework.

aifeducation::check_aif_py_modules(print=TRUE,
                                   check="pytorch")

aifeducation::check_aif_py_modules(print=TRUE,
                                   check="tensorflow")

Now everything is ready to use the package.

Important note: When you start a new R session, please note that you have to call reticulate::use_condaenv(condaenv = "aifeducation") before loading the library to make the python modules available for work.

2) Configuration of Tensorflow

In general, educators and educational researchers neither have access to high performance computing nor do they own computers with a performing graphic device for their work. Thus, some additional configuration can be done to get computations working on your machine.

If you do use a computer that does own a graphic device, but you would like to use cpu only you can disable the graphic device support of tensorflow with the function set_config_cpu_only().

aifeducation::set_config_cpu_only()

Now your machine only uses cpus only for computations.

If your machine has a graphic card but with limited memory, it is recommended to change the configuration of the memory usage with set_config_gpu_low_memory()

aifeducation::set_config_gpu_low_memory()

This enables your machine to compute ‘large’ models with limited resources. For ‘small’ models, this option is not relevant since it decreases the computational speed.

Finally, in some cases you might want to disable tensorflow to print information on the console. You can change the behavior with the function set_config_tf_logger().

aifeducation::set_config_tf_logger()

You can choose between five levels “FATAL”, “ERROR”, “WARN”, “INFO”, and “DEBUG”, setting the minimal level for logging.

3 Starting a New Session

Before you can work with aifeducation you must set up a new R session. First, it is necessary that you load the library. Second, you must set up python via reticulate. In case you installed python as suggested in this vignette you may start a new session like this:

reticulate::use_condaenv(condaenv = "aifeducation")
library(aifeducation)
set_transformers_logger("ERROR")

Next you have to choose the machine learning framework you would like to use. You can set the framework for the complete session with

#For tensorflow
aifeducation_config$set_global_ml_backend("tensorflow")

#For PyTorch
aifeducation_config$set_global_ml_backend("pytorch")

You can change the framework at anytime during a session by calling this method again or by passing the framework to the ml_framework argument of a function or method. Please note that not all models are available for both frameworks and that the weights of trained models cannot be shared across frameworks for all models.

In the case that you would like to use tensorflow now is a good time to configure that backend, since some configurations can only be done before tensorflow is used the first time.

#if you would like to use only cpus
set_config_cpu_only()

#if you have a graphic device with low memory
set_config_gpu_low_memory()

#if you would like to reduce the tensorflow output to errors
set_config_os_environ_logger(level = "ERROR")

Note: Please remember: Every time you start a new session in R you have to to set the correct conda environment, to load the library aifeducation, and to chose your machine learning framework.

4) Tutorials and Guides

A guide how to use the graphical user interface can be found in vignette 02a classification tasks.

A short introduction into the package with examples for classification tasks can be found in vignette 02b classification tasks.

Documenting and sharing your work is described in vignette 03 sharing and using trained AI/models

5) Update aifeducation

In the case you already use aifeducation and you want to update to a newer version of this package it is recommended to update the used python libraries. The easiest way is to remove the conda environment “aifeducation” and to install the libraries into a fresh environment. This can be done by setting remove_first=TRUE in install_py_modules.

#For Linux
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=TRUE,
                                 tf_version="<=2.14",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=FALSE)

#For Windows with gpu support
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=TRUE,
                                 tf_version="<=2.10",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=FALSE)
#For Windows without gpu support
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=TRUE,
                                 tf_version="<=2.14",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=TRUE)

#For MacOS
aifeducation::install_py_modules(envname="aifeducation",
                                 install="all",
                                 remove_first=TRUE,
                                 tf_version="<=2.14",
                                 pytorch_cuda_version = "12.1"
                                 cpu_only=TRUE)