How fabian57fabian/prototypical-networks-few-shot-learning works
Project type and purpose
• This is an open-source Python toolkit for few-shot image classification using Prototypical Networks.
• It provides command-line scripts for meta-training, meta-testing, centroid learning, and prediction against multiple image datasets.
File structure
• Top-level scripts:
• meta_train.py – orchestrates episode-based training
• meta_test.py – runs evaluation episodes on a saved model
• learn_centroids.py – computes class centroids from a trained embedding
• predict.py – classifies new images using stored centroids
• Core library in src/:
• cfg/defaults.yaml – configuration defaults loader
• core.py – main training/validation loop
• data/… – dataset interfaces (AbstractClassificationDataset, CustomDataset, specific datasets)
• prototypical_net.py – embedding network definition
• prototypical_loss.py – loss/metric implementation
• utils.py, EarlyStopping.py, centroids.py – helpers for training, checkpointing, and centroid computation
• Persistence:
• models/ – saved PyTorch model checkpoints
• results/ – logged training and evaluation outputs
• Tests:
• tests/ – unit tests for each core component
• Configuration files, CI workflow, and docs provide setup and reproducibility.
README for additional insights
• Confirms usage patterns, supported datasets, and script parameters (number of ways, shots, query samples, optimization settings).
• Highlights extensibility by allowing a “toy” custom dataset via the CustomDataset interface.
Main components and interactions
• MetaTrain Script
• MetaTest Script
• LearnCentroids Script
• Predict Script
• Configuration Loader (cfg)
• Data Loader (AbstractClassificationDataset + Specific Datasets)
• Embedding Network (prototypical_net)
• Loss & Metric (prototypical_loss)
• Training Engine (core)
• Utilities (utils, EarlyStopping, centroids)
• Model Checkpoints (models/)
• Centroid Files (runs/centroids_*)
• Results & Logs (results/)
• PyTorch / TorchVision
• File System / Dataset Storage (train/val/test folders)
• CLI → Configuration Loader: scripts parse CLI args and load defaults
• Configuration Loader → Training Engine / Data Loader
• Data Loader → Training Engine: yields episodes of (support, query) image tensors
• Training Engine → Embedding Network & Loss: forward pass, loss computation, backward pass
• Training Engine → Model Checkpoints: save best checkpoints via EarlyStopping
• LearnCentroids Script → Embedding Network → centroids utility → Centroid Files
• Predict Script → Embedding Network + Centroid Files → classification outputs → Results
• MetaTest Script uses same pipeline as MetaTrain but skips weight updates, writes evaluation metrics to Results
Architectural patterns and design principles
• Modular separation: CLI, configuration, data abstraction, model, loss, utilities
• Plugin-style dataset interfaces for extensibility
• Clean pipeline orchestration in core.py following the “engine” pattern
• Isolation of metric computation (prototypical_loss) from model definition
• Reusable utilities for early stopping and centroid computation