git clone https://github.com/BilgehanMehmet/one-shot-image-recognition.git
cd one-shot-image-recognition
git clone https://github.com/brendenlake/omniglot.git
cd omniglot/python
unzip images_background.zip
unzip images_evaluation.zip
wget ftp://ftp.uk.research.att.com/pub/data/att_faces.zip
--one-shot-learning directory
main.py
--docker
--networks
--siamese
--base
--mnist
--omniglot
--models
--notebooks
--pre_trained_weights
--datasets
--images_background
--image_validation
--images_evaluation
--configs
--bin
Run main.py to train the default one-shot learning model (Siamese Convolutional Neural Network) on the Omniglot dataset. Note that main.py calls the data loader, loader.py. For tuning purposes, mode, n_examples, path and normalise flag can all be modified in the main.py file. The path of trained weights can also be changed in the filepath variable of the function train in main.py.
python3 main.py
You can also train the model on the ORL Database of Faces (AT&T Dataset) by running model_orl.py. The dataset contains 40 subjects, with 10 images per subject. We split the training and validation sets as follows: 30 subjects were used to train the model, and the remaining 10 subjects were used to generate one-shot trials. Of the 30 subjects, we had two cases, where in Case 1 the model was trained on 2 images per person, and 8 images were used for validation per person; and in Case 2, it was trained on 4 images per subject and 6 images were used per subject for validation purposes. The model can be trained by running:
python3 model_orl.py
Similarly, the data loader arguments can be modified during training and testing.
More information on the model architecture, dataset pre-processing and training details of one-shot learning can be found in the report provided for this project.