siamese-nn-oneshot-reproduction

Reproduction of "Siamese Neural Networks for One-shot Image Recognition," by Gregory Koch, Richard Zemel, Ruslan Salakhutdinov (ICML 2015)

Download data from the Omniglot repository : https://github.com/brendenlake/omniglot

Koch et. al. Keras implementation for reference: https://github.com/sorenbouma/keras-oneshot

Setting up repository for development

1. download images_background and images_evaluation from (here)[https://github.com/brendenlake/omniglot] and unzip into data/raw.
2. create a virtual environment and install from requirements.txt
3. run python create_data.py [30|60|90] to create data for training. e.g., python create_data.py 30 creates training data of size 30,000 in accordance with the paper and stores it in data/processed/trainX_30k.npy and data/processed/trainY_30k.npy. Code snippet to read the .npy file is included in the last part of create_data.py.

Data Shape

create_data.py converts the image from background and evaluation into numpy arrays of shape:

trainX_30k: (30000, 2, 105, 105), trainY_30k: (30000, 1)

trainX_90k: (90000, 2, 105, 105), trainY_90k: (90000, 1)

trainX_150k: (150000, 2, 105, 105), trainY_150k: (150000, 1)

Where index 0 is number of samples, index 1 is the pairs of images (1 pair, 2 images), index 2 and 3 are the height and width of the image