mariobecerra / Active_Learning_CNNs

Active Learning with approximations of Bayesian Convolutional Neural Networks.

Active Learning with CNNs

Code for my master's thesis in Computer Science, "A comparison of frequentist methods and Bayesian approximations in the implementation of Convolutional Neural Networks in an Active Learning setting".

The repository of the final document is https://github.com/mariobecerra/msc_thesis.

And the final PDF file can be found here.

Folders:

• code: All code used for the experiments. Files:

  • utils.py: Python functions that are used from the R scripts. Here are the acquisition funtions and the dropout predictions from the models.

  • MNIST_active_learning.R: Performs the active learning acquisition steps for the MNIST dataset.

  • MNIST_utils.R: Functions used in MNIST_active_learning.R to perform the active learninng steps.

  • MNIST_utils.py: Python functions used in MNIST_active_learning.R to perform the active learninng steps. Particularly, the Keras model.

  • MNIST_plot_accuracies.R: Creates accuracy plot used in Results chapter.

  • MNIST_plot_probs.R: Creates probability plot used in Results chapter.

  • CIFAR10_active_learning.R: Performs the active learning acquisition steps for the CIFAR10 dataset.

  • CIFAR10_utils.R: Functions used in CIFAR10_active_learning.R to perform the active learninng steps.

  • CIFAR10_plot_accuracies.R: Creates accuracy plot used in Results chapter.

  • CIFAR10_plot_probs.R: Creates probability plot used in Results chapter.

  • cats_dogs_preprocess.R: Resizes original train set to a 64 by 64 pixel format. Saves the dataset in a list in an RDS file.

  • cats_dogs_active_learning.R: Performs the active learning acquisition steps for the cats and dogs dataset.

  • cats_dogs_utils.R: Functions used in cats_dogs_active_learning.R to perform the active learninng steps.

  • cats_dogs_plot_accuracies.R: Creates accuracy plot used in Results chapter.

  • cats_dogs_plot_probs.R: Creates probability plot used in Results chapter.

• data: Training and test data. Subfolders:

  • cats_dogs:

    • test1: Test data.

    • train: Training data.

• out: Objects and files created go here.