KevinMusgrave/pytorch-metric-learning

Features

• Metric Learning Pipelines - Provides complete training pipelines for deep metric learning with configurable loss functions and mining strategies.
• Batch Size Tuning - Implements cross-batch memory queues that store embeddings from previous iterations for contrastive learning.
• Retrieval Accuracy Metrics - Calculates retrieval and clustering metrics like precision, recall, and NMI from embedding distances.
• Clustering Evaluation Metrics - Provides clustering and retrieval metrics to evaluate embedding quality on held-out datasets.
• Loss Metric Optimizer Customizers - Swaps distance metrics, reducers, or regularizers inside loss functions to control similarity computation.
• Metric Learning Libraries - An open-source library for training similarity-preserving embeddings using modular loss functions, miners, and reducers.
• Embedding Computation - Computes clustering and retrieval metrics from query and reference embeddings with labels.
• Embedding Loss Functions - Provides loss functions that compute penalties from embeddings and labels for metric learning.
• End-to-End Training Pipelines - Provides ready-to-use trainers that handle the full training loop for metric learning models.
• Triplet Loss Functions - Provides triplet loss computation from all possible triplets within a batch for metric learning.
• Self-Supervised Label Generators - Wraps standard metric losses to train embeddings without labels using augmented views as positive pairs.
• Loss Function Customization - Provides modular loss function assembly from interchangeable distance metrics, reducers, and regularizers.
• Contrastive Loss Implementations - Trains embeddings by minimizing distance between similar pairs and maximizing distance between dissimilar pairs.
• Composable Loss Components - Provides a modular loss composition system where distance metrics, reducers, and regularizers can be swapped to customize loss behavior.
• Triplet Margin Losses - Implements triplet margin loss for metric learning to separate classes in embedding space.
• Model Training Pipelines - Provides complete training pipelines that initialize models, optimizers, losses, miners, and testers with logging.
• Metric Learning Pipelines - Orchestrates training loops combining loss functions, miners, and optimizers for similarity-preserving embeddings.
• Embedding Model Training - Trains neural networks to map similar items to nearby points in an embedding space using labeled data.
• Metric Learning Training - Runs training loops combining metric loss functions, mining strategies, and classifiers for similarity embeddings.
• Cascaded Embedding Refiners - Splits the embedder output into sections and computes separate losses for progressive refinement.
• Metric Loss Pipelines - Orchestrates a full training loop that applies loss functions, miners, and reducers in a configurable sequence.
• Self-Supervised Loss Calculators - Wraps standard metric learning losses to work with unlabeled data via augmented view pairs.
• Embedding Training Toolkits - Provides a toolkit for training embedding models with triplet, contrastive, proxy-based, and self-supervised losses.
• Metric Loss Training - Trains models using triplet and pair-based loss functions to learn similarity-preserving embeddings.
• Embedding Accuracy Evaluators - Evaluates how well the learned embedding space separates classes using retrieval and clustering metrics.
• Hard Negative Mining - Implements hard negative and positive pair mining to focus training on challenging examples.
• Batch-Level Hard Pair Miners - Provides configurable batch-level mining strategies that select informative pairs and triplets to accelerate metric learning convergence.
• Hard Pair and Triplet Miners - Provides configurable miners that select hard pairs and triplets from batches for metric learning.
• Hard Triplet Miners - Ships miners that select hard triplets from batches to accelerate metric learning convergence.
• Margin-Violation Triplet Miners - Ships miners that select triplets violating a margin constraint for metric learning.
• Multi-Similarity Miners - Provides multi-similarity mining to select the most informative pairs during metric learning training.
• Embedding Similarity Loss Functions - Implements loss functions that use configurable similarity measures to separate classes in embedding space.
• Cross-Batch Memory Queues - Implements cross-batch memory queues that store embeddings from previous iterations for contrastive learning.
• Metric Learning Workflows - Ships ready-to-use training loops and testers for end-to-end metric learning model training and evaluation.
• Contrastive Learning Wrappers - Implements MoCo-style self-supervised contrastive learning with cross-batch memory for embedding training.
• ML Batch Training Optimizations - Implements cross-batch memory queues that store embeddings from previous iterations for contrastive learning.
• Loss Value Reducers - Ships a pluggable reducer architecture that aggregates per-element or per-pair losses using interchangeable strategies like mean or threshold averaging.
• Cross-Batch Memory Queues - Implements cross-batch memory queues to enlarge negative sample pools without increasing batch size.
• In-Memory Databases with Persistence - Implements cross-batch memory queues that store embeddings from previous iterations for contrastive learning.
• Similarity Learning Libraries - Provides a library for learning similarity-preserving embeddings through configurable mining, losses, and metrics.
• Triplet Loss Training - Trains similarity models by computing triplet loss from all possible anchor-positive and anchor-negative pairs within a batch.
• Embedding Quality Diagnostics - Evaluates embedding quality using retrieval and clustering metrics like precision, recall, and NMI.
• Embedding Pairwise Distance Calculators - Computes pairwise distance or similarity matrices between embeddings using multiple configurable metrics.
• Embedding Evaluation Suites - Ships a suite of tools for evaluating embedding quality using retrieval and clustering metrics.
• Metric Learning Frameworks - Provides a modular framework for composing and training metric learning models with interchangeable components.
• Metric Learning Classifier Heads - Combines metric and classification losses for architectures with trunk, embedder, and classifier heads.
• Balanced Class Samplers - Implements balanced batch sampling that draws a fixed number of samples per class to ensure balanced representation during training.
• Distributed GPU Training - Supports multi-GPU distributed training with correctly wrapped loss functions and miners for PyTorch.
• Distributed Training - Wraps loss functions and miners to operate correctly under PyTorch's DistributedDataParallel.
• Data-Parallel Training - Wraps loss functions and miners to operate correctly under PyTorch's DistributedDataParallel.
• Distributed Training - Wraps loss functions and miners to operate correctly under PyTorch's DistributedDataParallel.
• Fixed-Depth Sampling - Provides fixed-depth sampling strategies that limit neighbor selection to constrain computational growth during training.
• Metric Learning Classifier Heads - Trains embeddings jointly with a classification layer to combine metric learning and classification objectives.
• Asymmetric Embedding Comparisons - Enables loss computation with separate anchor and reference embedding sets for asymmetric comparisons.
• Per-Class Instance Samplers - Implements per-class instance sampling that draws a fixed number of samples from each class for balanced batch construction.
• Angular Softmax Loss Training - Trains embeddings by applying a normalized softmax with angular margins for class separation.
• Combined Loss Functions - Wraps multiple loss functions together, summing their outputs with optional per-loss weights and miners.
• Configurable Loss Combinations - Sums or averages outputs of several loss functions with per-loss miners and weights in a single forward pass.
• Metric and Classification Loss Combiners - Supports combining metric and classification losses for joint embedding training.
• Anchor-Reference Separation - Supports loss computation with separate anchor and reference embedding sources for asymmetric comparisons.
• Multi-Strategy Training Loops - Orchestrates training loops that combine metric losses, classifiers, cascaded embeddings, or adversarial objectives.
• Embedding Inference Utilities - Ships utilities to encode new data points into a learned embedding space using a trained model.
• Adversarial Robustness Training - Generates adversarial samples during training to make embeddings robust to small perturbations.
• Precomputed Pair and Triplet Loss Acceptors - Accepts precomputed pair or triplet indices directly for loss computation without labels.
• Angular Loss Training - Trains embeddings by optimizing the angle between feature vectors for class separation.
• Margin-Based Loss Training - Trains embeddings by applying a margin constraint on pairwise distances for class separation.
• Distribution Regularizers - Ships distribution regularizers that encourage uniform Lp norms or zero-mean embeddings during metric learning training.
• Norm Penalizers - Provides norm-based penalties to shrink embedding magnitudes during metric learning training.
• Space Shaping Regularizers - Applies center invariance and sparsity penalties to shape the learned embedding space distribution.
• Distance-Range Pair Miners - Provides miners that select pairs based on distance ranges for focused metric learning.
• Distance-Weighted Pair Miners - Implements distance-weighted sampling that selects pairs or triplets with probability proportional to their distance for stable training.
• Hardness-Percentage Pair Miners - Implements miners that select the hardest percentage of pairs for focused training.
• Uniform-Distance Pair Miners - Implements miners that select pairs with uniform distance distribution for balanced training.
• Cross-Batch Embedding Queues - Ships cross-batch memory queues that retain embeddings from previous iterations to enlarge the negative sample pool.
• Similarity-Threshold Pair Miners - Provides miners that select pairs by similarity thresholds for metric learning training.
• Developer Tools - Metric learning library.
• More to explore - Library for metric learning applications.
• PyTorch Utilities - Listed in the “PyTorch Utilities” section of the The Incredible Pytorch awesome list.

Star history