Meta-Learning Suite (MAML/FOMAML/Reptile + MAML++ Enhancements)

You can implement MAML and FOMAML by defining a model, configuring a meta-algorithm, and running a training loop. The suite executes the complete workflow, handling inner-loop adaptations and outer meta-gradients for fast adaptation.

Meta-learning trains models to rapidly adapt to new tasks from few examples. LSLR, a MAML++ enhancement, improves MAML by applying per-layer per-step learning rates within the inner loop to optimize adaptation.

Yes, the suite supports first-order and second-order inner loops through pluggable backends. It prefers torch.func for gradient flow, with higher or a custom SGD fallback available for meta-learning experiments.

Yes, you can run Reptile and MAML++ enhancements on toy datasets and real benchmarks. The framework targets episodic few-shot tasks, enabling experiments with multi-step loss, annealing, and batch-norm handling.

To ensure reproducible gradient flow in episodic few-shot training, use deterministic episode sampling and robust checkpointing. The suite standardizes the learn-to-adapt pipeline to guarantee experiment reproducibility.

You need PyTorch and pytest dependencies to run gradient-based meta-learning experiments. The suite uses these to support pluggable inner-loop engines and a consistent evaluation interface for research.