aifeducation 1.1.0

Major Changes

• Removed support for ‘tensorflow’.
• Refactor of all classifiers and for FeatureExtractor.
• Added support for modernBERT.
• Removed support for DeBERTa_V2. Implementation of this model changed from ‘transformer’ version 4.46.3 to 4.47.1. This causes that the model does not produce the same results for the same data after saving and loading a model. Reproducibility is not guaranteed. In the case that this is fixed in the future the model support will be re-implemented.
• Changed the strings for a high numbers of arguments (e.g., “lstm” into “LSTM”) in order to be more in line with PyTorch. Old models are updated automatically.

Installation and Configuration

• Removed functions belonging to ‘tensorflow’.
• Added the possibility to install python packages either to a ‘conda’ environment or a virtual environment. Virtual environment is the new default.
• Added functions for a more convenient preparation of a new session and for installing/updating python packages.

LargeDataSetForText

• Added an algorithm cleaning raw texts in order to improve the performance of the following analysis. See the method’s documentation for more details. Currently only available for .pdf and .txt files.

Ai for Education Studio

• Implemented popovers to explain the different widgets. This feature will be extended in the future.
• Fixed a bug that prevented classifiers with applied pseudo labeling to visualize training history.
• Re-created the user interface for classifiers, FeatureExtractor, and base models. Now the user interfaces generates the necessary control widgets for configuration and training automatically depending on the method’s arguments.
• Added loading animations for the time widgets are generated.

TextEmbeddingModel

• Removed parameter method from method configure. The method of the model is now detected automatically.
• Fixed a bug which caused that TextEmbeddingModels saved the model without the mlm head.

Classifiers

• Added parameters for determining the learning rate and warm up ratio for training.
• Added a check for the number of unlabeled cases during training to avoid application of pseudo labeling if there are no unlabeled cases.
• Fixed an error in calculating the number of folds if the number of requested folds is greater as the minimal frequency of all classes/categories.
• Users can now choose between different activation functions and parametrizations.
• Removed argument dir_checkpoint from training method. Now the training uses a folder in the regular temp directory of the machine. After a successful training the folder is removed.
• Parameter ‘name’ in ‘configure’ is now optional. If set to NULL a unique name is generated automatically.
• Added Focal Loss as a new loss function to sequential classifiers.
• Added four new classes of classifiers.
• Fixed a bug that caused classifiers to be not order invariant if the attention type is “Fourier”.

FeatureExtractor

• The option to choose an optimizer is now working.
• Added parameters for determining the learning rate and warm up ratio for training.
• Tracking sustainability is now working.
• Removed argument ‘dir_checkpoint’ from training method. Now the training uses a folder in the regular temp directory of the machine. After a successful training the folder is removed.
• Parameter ‘name’ in ‘configure’ is now optional. If set to NULL a unique name is generated automatically.

Minorty Oversamping Techniques

• Added K-Nearest Neighbor OveRsampling approach (KNNOR) in C++ as new oversampling technique.
• Removed support for all other oversampling techniques. No dependency to the package ‘smotefamily’.
• Changed update intervall of most processes within AI for Education - Studio from 300 to 30 seconds.

Performance Measures

• Added an own implementation of Gwet’s AC1 and AC2 according to Gwet(2021).
• Removed dependency to package ‘irrCAC’.

Minor Changes

• Updated code for newer versions of numpy.
• Added the function prepare_python for a convenient set up of virtual and ‘conda’ environments.

aifeducation 1.0.2

• Fixed a bug with alpha 3 codes for sustainability tracking preventing Ai for Education Studio to start.
• The source urls of entries in LargeDataSetsForTexts are now displayed correctly within Ai for Education Studio.

aifeducation 1.0.1

Fixed a bug with the initialization of ‘codecarbon’ on linux.

aifeducation 1.0.0

First complete release of the package including major changes, bug fixes, new features, and objects.

The most important change is that we decided to use ‘PyTorch’ for several reasons. First, ‘PyTorch’ is a very flexible and stable machine learning framework. At the moment, most new architectures are based on ‘PyTorch’ as can be seen on Hugging Face. Currently (11th November 2024) there are 190,237 models for this framework compared to 13,346 models for ‘tensorflow’. Second, ‘PyTorch’ provides an easy installation and supports native GPU acceleration on Linux and Windows while tensorflow supports native GPU support only on Linux and for Windows only in version 2.10 or lower. Fourth, keras, which was an important element of ‘tensorflow’, changed to a multi-back-end framework. However, keras 3.0 does not have a native Windows support. Since we assume that many educational researchers use either Windows or Mac and are not familiar with more complex system configurations (such as using Windows subsystem for Linux (WSL)), this is problematic.

In addition, we changed the algorithm for saving and loading models, data, and objects to ensure that models trained with the package are working within future versions of aifeducation and can be updated to new developments. This is also necessary to allow reproducibility of models and research based on these models. To achieve this goal we had to make some changes for models created with version 0.3.3 or lower. If you still need these models, please install an older version of aifeducation.

The following changes have been made:

Major Changes

• The core machine learning framework is now ‘PyTorch’. ‘Tensorflow’ is still supported but only for some models and limited to version 2.15. Further implementation and support for ‘tensorflow’ models is currently not planned. We decided to base the package on ‘PyTorch’ because this framework is widely used in research, is very flexible, provides a broad GPU support, and offers more stable code across versions.
• Implemented a new mechanic and new methods for all objects allowing objects that were created with an older version of the package to update to the current version during loading.
• Removed the bag-of-words models from the package in order to focus the package on approaches which use AI.

Installation and Configuration

• Added a new function for a convenient installation of ‘python’ and ‘pytorch’.

Transformer Models

• Complete rewrite of all transformer functions into a modern object-oriented approach with R6 classes (AIFETransformerMaker).
• Functions of type create_xxx_model and train_xxx_model are now deprecated.
• Added support for MPNet with ‘pytorch’ and ‘tensorflow’.

TEFeatureExtractor

• Adding TEFeatureExtractor as a new class for ‘pytorch’ only.
• TEFeatureExtractor are auto-encoders that can be used to reduce the number of features of text embeddings before passing them onto classifiers. Their aim is to reduce computational time and/or increase performance of classifiers.

TextEmbeddingClassifiers

• TEClassifierRegular replaces TextEmbeddingClassifierNeuralNet. This new class provides additional methods and fixes a bug for pytorch models used to predict two classes.
• TextEmbeddingClassifierNeuralNet is now deprecated.
  
• Added TEClassifierProtoNet which is a classifier that applys methods of meta-learning based on ProtoNets.
• In comparison to TextEmbeddingClassifierNeuralNet, the training loop for the new classes was altered and reduced in its complexity for users. For example, only the type of pseudo-labeling described by Cascante-Bonilla et al. (2020) is now implemented and at the same type the technique described by Lee (2013) was removed. In addition, it is now possible to add synthetic cases within every step of pseudo-labeling. See the vignettes for more details.

Graphical User Interface Aifeducation Studio

• Complete rewrite of the user interface based on bslib while removing the dependencies to shinydashboard.
• User interface only supports pytorch and no longer tensorflow.
• Implemented long running tasks such as training a transformer as a shiny ExtendedTask. This allows the computation of the task in the background and the shiny app to stay responsive. This, in turn, avoids “greying out” of the app.
• Implemented a new reporting system for providing a feedback to the user during computations.

Data Management

• Introduced two new classes LargeDataSetForTextEmbeddings and LargeDataSetForText based on the python libraries ‘arrow’ and ‘datasets’ allowing to store and use data that would not fit into memory. LargeDataSetForText stores raw texts while LargeDataSetForTextEmbeddings contain text embeddings.
• Added support to all AI models for these new kinds of objects to allow training with large data sets.
• Added new methods to objects of class EmbeddedTexts (e.g. for converting EmbeddedTexts into a LargeDataSetForTextEmbeddings). See the corresponding documentation for more details.
• The function combine_embeddings is now deprecated. Please use the corresponding method of EmbeddedTexts.

Saving and Loading

• Introduced save_to_disk and load_from_disk as the new core functions for saving and loading objects and models of this package.
• Functions load_ai_model and save_ai_model are now deprecated. Please use these functions only for models created with version 0.3.3 or lower.

Further Changes

• Removed the dependencies to package abind and irr.
• Updated vignettes.

aifeducation 0.3.3

Graphical User Interface Aifeducation Studio

• Fixed a bug concerning the IDs of .pdf and .csv files. Now the IDs are correctly saved within a text collection file.
• Fixed a bug while checking for the selection of at least one file type during creation of a text collection.

TextEmbeddingClassifiers

• Fixed the process for checking if TextEmbeddingModels are compatible.

Python Installation

• Fixed a bug which caused the installation of incompatible versions of keras and Tensorflow.

Further Changes

• Removed quanteda.textmodels as necessary library for testing the package.
• Added a dataset for testing the package based on Maas et al. (2011).

aifeducation 0.3.2

TextEmbeddingClassifiers

• Fixed a bug in GlobalAveragePooling1D_PT. Now the layer makes a correct pooling. This change has an effect on PyTorch models trained with version 0.3.1.

TextEmbeddingModel

• Replaced the parameter ‘aggregation’ with three new parameters allowing to explicitly choose the start and end layer to be included in the creation of embeddings. Furthermore, two options for the pooling method within each layer is added (“CLS” and “Average”).
• Added support for reporting the training and validation loss during training the corresponding base model.

Transformer Models

• Fixed a bug in the creation of all transformer models except funnel. Now choosing the number of layers is working.
• A file ‘history.log’ is now saved within the model’s folder reporting the loss and validation loss during training for each epoch.

EmbeddedText

• Changed the process for validating if EmbeddedTexts are compatible. Now only the model’s unique name is used for the validation.
• Added new fields and updated methods to account for the new options in creating embeddings (layer selection and pooling type).

Graphical User Interface Aifeducation Studio

• Adapted the interface according to the changes made in this version.
• Improved the read of raw texts. Reading now reduces multiple spaces characters to one single space character. Hyphenation is removed.

Python Installation

• Updated installation to account for the new version of keras.

aifeducation 0.3.1

Graphical User Interface Aifeducation Studio

• Added a shiny app to the package that serves as a graphical user interface.

Transformer Models

• Fixed a bug in all transformers except BERT concerning the unk_token.
• Switched from SentencePiece tokenizer to WordPiece tokenizer for DeBERTa_V2.
• Add the possibility to train DeBERTa_V2 and FunnelTransformer models with Whole Word Masking.

TextEmbeddingModel

• Added a method for ‘fill-mask’.
• Added a new argument to the method ‘encode’, allowing to chose between encoding into token ids or into token strings.
• Added a new argument to the method ‘decode’, allowing to chose between decoding into single tokens or into plain text.
• Fixed a bug for embedding texts when using pytorch. The fix should decrease computational time and enables gpu support (if available on machine).
• Fixed two missing columns for saving the results of sustainability tracking on machines without gpu.
• Implemented the advantages of datasets from the python library ‘datasets’ increasing computational speed and allowing the use of large datasets.

TextEmbeddingClassifiers

• Adding support for pytorch without the need for kerasV3 or keras-core. Classifiers for pytorch are now implemented in native pytorch.
• Changed the architecture for new classifiers and extended the abilities of neural nets by adding the possibility to add positional embedding.
• Changed the architecture for new classifiers and extended the abilities of neural nets by adding an alternative method for the self-attention mechanism via fourier transformation (similar to FNet).
• Added balanced_accuracy as the new metric for determining which state of a model predicts classes best.
• Fixed error that training history is not saved correctly.
• Added a record metric for the test dataset to training history with pytorch.
• Added the option to balance class weights for calculating training loss according to the Inverse Frequency method. Balance class weights is activated by default.
• Added a method for checking the compatibility of the underlying TextEmbeddingModels of a classifier and an object of class EmbeddedText.
• Added precision, recall, and f1-score as new metrics.

Python Installation

• Added an argument to ‘install_py_modules’, allowing to choose which machine learning framework should be installed.
• Updated ‘check_aif_py_modules’.

Further Changes

• Setting the machine learning framework at the start of a session is no longer necessary. The function for setting the global ml_framework remains active for convenience. The ml_framework can now be switched at any time during a session.
• Updated documentation.

aifeducation 0.3.0

• Added DeBERTa and Funnel-Transformer support.
• Fixed issues for installing the required python packages.
• Fixed issues in training transformer models.
• Fixed an issue for calculating the final iota values in classifiers if pseudo labeling is active.
• Added support for PyTorch and Tensorflow for all transformer models.
• Added support for PyTorch for classifier objects via keras 3 in the future.
• Removed augmentation of vocabulary from training BERT models.
• Updated documentation.
• Changed the reported values for kappa.

aifeducation 0.2.0

• First release on CRAN