jacobdwatters/jml

A simple modular machine learning library for Java.

jml's Issues

Add Softmax activation for Neural Networks

Saved Neural Network loses information

• Dropout Layers: When a neural network is saved, dropout layers are not saved in the file.
• Optimizers: When a neural network is saved, the optimizer used during training is not saved.

This means when the model is reloaded, it will be missing these components of the neural network. Neither of these will result in different predictions being made by the loaded model but if the model was to be retrained, it may not behave as the user expects.

To address this, the optimizer's details (including learning rate scheduling) and dropout layers should be saved to the .mdl file so they can be read upon loading the model and the model can be constructed correctly.

Publish as Maven Package

Add ability to freeze layers.

In order to support transfer learning, layers should be able to be frozen. I.e. their trainable parameters should be able to be made non-trainable.

Combine MultipleLinearRegression and LinearRegression classes

MultipleLinearRegression and LinearRegression should be combined into a single class called 'LinearRegression'.

All normalizers will become objects.

All methods of normalization/standardization in Normalize.Java will become their own objects. These objects will implement a FeatureScaler interface.

Encoder Objects and Interface.

All encoders should be an object with a fit(), decode(), and encode() method. Add an Encoder interface which all of these objects should inherit from.

Logistic Regression Should Allow Multiple Classes.

In the current implementation, the LogisticRegression class only works with binary classes. It should support multi-class data and implement one-vs-rest and multi-label regression (i.e. softmax regression).