Weka is tried and tested open source machine learning software that can be accessed through a graphical user interface, standard terminal applications, or a Java API. It is widely used for teaching, research, and industrial applications, contains a plethora of built-in tools for standard machine learning tasks, and additionally gives transparent access to well-known toolboxes such as scikit-learn, R, and Deeplearning4j.
WekaDeeplearning4j: Deep Learning using Weka
WekaDeeplearning4j is a deep learning package for the Weka workbench. It is developed to incorporate the modern techniques of deep learning into Weka. The backend is provided by the Deeplearning4j Java library.
The source code for this package is available on GitHub. The java-doc can be found here.
Functionality
All functionality of this package is accessible via the Weka GUI, the commandline and programmatically in Java.
The following Neural Network Layers are available to build sophisticated architectures:
- ConvolutionLayer: applying convolution, useful for images and text embeddings
- DenseLayer: all units are connected to all units of its parent layer
- SubsamplingLayer: subsample from groups of units of the parent layer by different strategies (average, maximum, etc.)
- BatchNormalization: applies the common batch normalization strategy on the activations of the parent layer
- LSTM: uses long short term memory approach
- GlobalPoolingLayer: apply pooling over time for RNNs and pooling for CNNs applied on sequences
- OutputLayer: generates classification / regression outputs
New Features
The new release of WekaDeeplearning4j contains some few exciting features that will make it easier than ever to run experiments on your data.
Pretrained Models
A wide range of pretrained models are now available, sourced from both the DL4J model zoo and the Keras Applications module. These can easily be instantiated and used as a base for further finetuning or simply as a feature extractor, after which you can then apply any standard Weka classifier. The weights are cached locally after being initially downloaded, so it’s quick to instantiate in the future. Check out the Model Zoo for more information.
Updated Dl4jMlpFilter
The Dl4jMlpFilter takes activations from a layer in the given neural network and uses those as the output for the instance, commonly referred to as feature extraction or embedding creation.
The DL4jMlpFilter can now accept multiple layer names to use activations from, concatenating the activations together.
Pooling functions can be applied to the activations if using activations from an intermediary layer which outputs 3D activations (e.g. a convolution layer which outputs a set of feature maps).
Check out the filter tutorial for usage examples.
Image Dataset Conversion Script
Often, image classification datasets come in a folder-organized fashion i.e., instance classes are inferred from the subfolder they’re in as opposed to having a seperate file explicitly defining this. Now included is ImageDirectoryLoader, a tool to create an .arff file from this folder structure so it can be loaded into WEKA.
Check out Classify Your Own Dataset for usage examples
CUDA 10.2 Support
The new release of Deeplearning4j (1.0.0-beta7) now supports CUDA 10.2, so WekaDeeplearning4j has new installation packages for users with this CUDA version.
The new release of DL4J has dropped support for CUDA 9.2, so this is also no longer supported in WekaDeeplearning4j.
Citation
Please cite the following paper if using this package in an academic publication:
S. Lang, F. Bravo-Marquez, C. Beckham, M. Hall, and E. Frank WekaDeeplearning4j: a Deep Learning Package for Weka based on DeepLearning4j, In Knowledge-Based Systems, Volume 178, 15 August 2019, Pages 48-50. DOI: 10.1016/j.knosys.2019.04.013 (author version)
Contributing
Contributions are always welcome. If you want to contribute to the project, check out our contribution guide.
Future Work
Future work on WekaDeeplearning4j will include network weight and activation visualization, and support for multiple embeddings as input channels for textual data.