One Shot Image Recognition

An implementation of the paper Siamese Neural Networks for One-shot Image Recognition. The model architecture is as below (taken from the paper):

With the only difference being same learning rate for all parameters, instead of layer wise learning rate as in the paper. With this, achieves 93% accuracy with 30k pairs.

Additional features:

• Training loss and validation accuracy are logged via TensorBoard.
• Best model as per validation accuracy saved at the end of training.

Requirements (Python 3.6+):

• PyTorch
• Augmentor

Usage:

• Download the Omniglot 'background' and 'evaluation' zip files from here.
• Run data_prep.py to split the dataset into train-val-test:

python data_prep.py <DATA_DIR>

where DATA_DIR is the directory containing both the unzipped files.

• Run train.py to train the model:

python train.py <DATA_DIR>

• Hyperparameters such as learning rate, number of pairs for training, augmentation etc. can be changed in train.py .