MLE-Image-Classification

Meta-learning ensemble framework for image classification with convolutional neural networks

Purpose

This project creates an ensemble of CNN models for the MINST, Fashion-MNIST, and CIFAR10 dataset. The predictions from the ensemble are evaluated by a CNN super-learner and a weight matrix optimized by a genetic algorithm.

A special case is used for binary classification where a MLP detects errors in the aggregation of prediction matrices and flips the final predictions accordingly.

All base learners in the ensemble are saved to disk along with the prediction matrices and final predictions from the GA and CNN meta learners.

---

Set-up

Set up a virtual environment (source)

A virtual environment, a self-contained directory tree that contains a Python installation for a particular version of Python, plus a number of additional packages.

Unix/MacOS

• python3 -m pip install --user virtualenv - if you don't have virtualenv installed
• python3 -m venv venv - Create a virtual environment
• source venv/bin/activate - Activate virtual environment
• python3 -m pip install -r requirements.txt - Install required libraries for project

Windows

• py -m pip install --user virtualenv - if you don't have virtualenv installed
• py -m venv venv - Create a virtual environment
• .\venv\Scripts\activate - Activate virtual environment
• py -m pip install -r requirements.txt - Install required libraries for project

---

Config settings

The project uses dafault.ini to set configurations for dataset, ensemble and GA prediction. The following table shows the allowed values for the configurations:

Section	Setting	Allowed values	Description
General	Dataset	any one of MNIST, MNIST_BIN, F_MNIST, F_MNIST_BIN, CIFAR10, CIFAR10_BIN	Dataset to use
General	ValidationProportion	decimal number in range [0.0, 1.0]. Recommended 0.2	Proportion of training data to use to train meta-learners
Ensemble	Type	full or stacked	full: no meta-learner and full training set is used to train ensemble; stacked: validation data is left for training meta-learners
Ensemble	NumBaseLearners	Any positive integer > 2. Recommended a value above 30	Number of base learners in ensemble
Ensemble	UseDynamicLoss	True or False	Boolean value specifying whether or not to use dynamic loss to train base learners in the ensemble
Ensemble	DynamicLossThreshold	decimal number in range [0.0, 1.0]. Recommended a value between 0.6 and 0.9	Threshold to use for dynamic loss
GA	GAGenerations	Any positive integer. Recommended a value above 150	Number of generations to run GA meta-learner for
GA	GAPopulation	Any positive integer. Recommended a value above 100	Population to maintain in GA optimizer