How kimyeonghyeon/metalearning-echocardiogram works
Project type and purpose
• This is an academic/ML research codebase implementing a few‐shot echocardiogram segmentation pipeline using model-agnostic meta‐learning (MAML, MetaSGD, MetaCurvature).
• The primary workflows are meta-training (train.py), meta-testing (test.py), and inference (inference.py), using PyTorch, learn2learn, segmentation-models-pytorch, and timm.
File structure
• Top-level scripts:
• train.py – launches the meta‐training loop
• test.py – runs evaluation on held-out tasks
• inference.py – performs prediction on new images
• configs/ – YAML files defining hyperparameters, data paths, shot/target settings
• meta_learning/ – older or notebook‐based meta‐learning implementations
• utils/ – shared modules: DataModule, dataset loaders, model definitions, optimizers, loss functions, training/evaluation routines
README insights
• Describes CLI arguments, output folder conventions (outputs-MAML…-lightning_logs), external model/data download links
• Specifies environment and dependency versions
Main components
a. Data Source
• Raw echocardiogram images (downloaded from Google Drive)
• Meta-train / meta-test splits
b. Configuration Manager
• YAML loader (configs/main.yaml)
• CLI argument parser (hydra/argparse)
c. DataModule & Dataset Loaders
• utils/DataModule.py orchestrates PyTorch DataLoaders
• utils/dataset*.py define per-view datasets, tasks sampling (few-shot episodes)
d. Model Zoo
• UNet and DeepLabV3Plus architectures (via segmentation-models-pytorch, timm)
• Wrapped in meta_learning/model.py or utils/model.py
e. Meta-Learning Engine
• learn2learn algorithm wrappers (MAML, MetaSGD, MetaCurvature) – meta_learning/utils/optimizer.py
• Inner-loop adaptation and outer-loop update logic (utils/train_metalearning.py)
f. Training & Evaluation Routines
• utils/train.py, utils/train_heatmap.py for standard training
• utils/train_metalearning.py for meta loops
• utils/evaluation.py for computing segmentation metrics, heatmap errors
g. Inference Pipeline
• inference.py calls DataModule → loads checkpoint → runs forward pass → utils/heatmaps.py to generate output
h. Logging & Outputs
• PyTorch Lightning logs under outputs-…-lightning_logs
• Model checkpoints saved for each version
Relationships and data flows
• Data Source → DataModule (batch/task generation)
• DataModule → Meta-Learning Engine ↔ Model Zoo
• Meta-Learning Engine → Loss & Optimizer → Model parameters
• Model outputs → Evaluation → Metrics → Logger
• Inference: DataModule → Model checkpoint → Output heatmaps
Architectural patterns & design principles
• Modular separation of concerns: data loading, model definitions, training loops, evaluation
• Plugin‐style config (YAML) decouples hyperparameters from code
• Meta-learning pattern: nested loops (inner adaptation vs. outer meta‐update)
Detailed elements
• Labels: “Meta-Training Script (train.py)”, “DataModule (utils/DataModule.py)”, “UNet/DeepLab Model”
• Config block pointing to configs/*.yaml, showing injection into scripts