Visual Contrastive Learning for Few-shot Image Classification

This repository provides the code to classify images in two different categories, i.e. Similar (1) and Dissimilar (0) based on the image similarity task performed by utilizing a Contrastive Learning-based approach (including employing a custom contrastive loss). Furthermore, Siamese Networks is being used in n-way k-shot settings considered in the current implementation.

Requirements

• Python 3.9
• PyTorch 1.10.2
• TorchVision 0.11.3
• numpy 1.22.3
• matplotlib 3.5.1

Usage

Data

Omniglot dataset is being used which is a collection of 1623 hand drawn characters from 50 different alphabets. For every character there are just 20 examples, each drawn by a different person. Each image is a gray scale image of resolution 105x105. Please clone this repo and then extract the images_background and images_evaluation folders. Finally, run DataGeneration.py file to create pickle files train.pickle and val.pickle files and store them in data folder. Here, train.pickle file contains characters from 30 different alphabets, whereas val.pickle contains characters from remaining 20 different alphabets.

Model Building and Training

• The SiameseNetwork model class for n-way k-shot learning can be found in Model.py file.
• To train the network, run Training.py file.
• The average loss for the trained model is printed after every epoch.
• All hyperparameters to control training and testing of the model are provided in the given Training.py file.

Output Samples

Image Similarity Scores

Image Comparison 1	Image Comparison 2	Image Comparion 3

Image Comparison 4	Image Comparison 5	Image Comparion 6

Image Comparison 7	Image Comparison 8	Image Comparion 9	Image Comparison 10

Results for Image Classification

Image Comparison 1	Image Comparison 2	Image Comparion 3

Analysis

Among all the 10 comparisons made under Image Similarity Scores sub-section, images 1, 6, and 8 appear more similar, thereby having predicted labels as 1, as shown in the Results for Image Classification sub-section. This way, the current implementation frames the image similarity task as the image classification task.