One Shot Learning

One Shot Learning using Omniglot data set and Siamese Network

This repository contains an implementation version of the paper koch et al, Siamese Networks for one-shot learning.

The model has been created using keras.

The codebase has been divided in the following components:

• OmniglotService: Download the Omniglot data from the repository.
• OmniglotLoader: Load the Omniglot data based on categories.
• OmniglotDataset: Expose an api to process Omniglot data set.
• SiameseNet: Create a SiameseNetwork with the default configuration.
• OneShotRunner: Run One-shot learning experiments using Omniglot data set and siamese network.
• RunnerVisualizer: To display runner results.
• Utils: I/O funtionality to read/save state/data.

Details about the machine learning problem, the implementation and experiments are explained in the one shot notebook.

Disclaimer

This is an early version of a work in progress prototype.

Roadmap

• Allow Data preprocessing (normalization, resizing, augmentation, ...)
• Allow more configuration parameters for the runner
• Allow more distance discriminator functions to compare two encoded images
• Implement a NetworkConfigurator component for hyperparameter tunning
• Implement another model using a different approach
• ...