This repository is provides a hands on approach to train a Pytorch model using Metric Learning method. In short the model learns to differentiate between images of different classes, it returns a N dimensional vector which can be used to calculate distance between different embedding to find most similar image or label.
- Metric learning are specially useful when number of classes are very large and you may have less number of images per class.
- Please make sure that the number of images per class or the class distribution is balanced before training the model, You can apply augmentation method to over sample some classes which are less in number.
You can start by installing modules from requirements.txt
Some module like Pytorch or torchvision is not included as you can use latest version
- To train the model, follow the Training Notebook
- While training, you can use the Visualization Notebook to see training graphs or test how model is performing by predicting on test images.
- Finally to index the embedding vectors, follow the Indexing Notebook
