TripletLib

TripletLib is a lightweight Python library for training and evaluating embedding models using triplet loss. It is designed to help you build models that learn meaningful embeddings for images, time series, or any vector data, enabling tasks like verification, retrieval, or classification based on similarity.

What is Triplet Loss?

A "triplet" is a set of 3 samples of data (anchor, positive, negative). The anchor is a data sample from a specific class, the positive is another sample from the same class, and the negative is a sample from an entirely different class. Triplet loss is a type of metric learning objective that trains an embedding model to map inputs into a vector space such that:

• Samples from the same class (e.g., anchor and positive) are close together.
• Samples from different classes (e.g., anchor and negative) are far apart.

This is particularly useful for tasks like biometric identification (e.g., face recognition), where traditional classification cannot generalize to unseen classes (new faces).

Features

• Base EmbeddingModel class with a built-in training loop.
• TripletDataset wrapper to automatically generate triplets from any dataset.
• Utilities for evaluating embeddings with EER and AUC metrics.
• Compatible with PyTorch and supports GPU acceleration.

Installation

You can install TripletLib via pip:

pip install tripletlib

Usage

import torch
from tripletlib.triplet_utils import EmbeddingModel, TripletDataset
from torch.utils.data import DataLoader

# Example: using a simple linear embedding model
class LinearEmbeddingModel(EmbeddingModel):
    def __init__(self, input_size, embedding_dim=64, lr=1e-3, margin=1.0, device="cpu"):
        super().__init__(embedding_dim, margin=margin, device=device)
        self.fc = torch.nn.Sequential(
            torch.nn.Linear(input_size, 128),
            torch.nn.ReLU(),
            torch.nn.Linear(128, embedding_dim)
        )
        self.optimizer = torch.optim.Adam(self.parameters(), lr=lr)

    def forward(self, x):
        x = self.fc(x)
        return torch.nn.functional.normalize(x, p=2, dim=1)

# Prepare your dataset as a list of (feature, label) tuples
train_dataset = [(torch.randn(10), label) for label in range(5) for _ in range(20)]
triplet_train = TripletDataset(train_dataset)
dataloader_train = DataLoader(triplet_train, batch_size=16, shuffle=True)

# Instantiate and train the model
model = LinearEmbeddingModel(input_size=10, embedding_dim=64, lr=1e-3)
history = model.fit(dataloader_train, epochs=5)

Examples

You can find examples of real usages in /examples, including:

• Keystroke Dynamics: Comparing typing patterns to determine if they are from the same individual.
• More to come...