This project is a coarse replication of the paper Deep Metric Learning Using Triplet Network (Hoffer et al.).
Run the main_triplet.py file to train the model. Loss, accuracy, and encoded graphs are auto-generated using matplotlib under root. By default, the dataset (CIFAR10 Images) will be encoded into 128D vectors by the trained model.
Run the main_classifier.py file to train another independent, one-layer neural network, with inputs being the encoded images (128D vectors) and outputs being a softmax classifying vector (10D vector).
To customize loss function or model architecture, use the lib folder and modify the classes.
To apply the sampling method in the paper Sampling Matters in Deep Embedding Learning (Wu et al.), run the triplet_sampling.py file.
Using the following hyperparameters:
learning_rate=0.01, loss_margin=1, batch_size=100, num_train_instance=50000, num_test_instance=10000
The model achieves KNN-accuracy of 65% on testing dataset after around 50 epochs.
The encoded images shows clusters after applying PCA for dimension reduction.
This program uses a different architecture than the one mentioned in the paper. Additionally, data augmentation techniques are different from the one used in the original paper.