supervised reptile

We have hosted the application supervised reptile in order to run this application in our online workstations with Wine or directly.

Quick description about supervised reptile:

The supervised-reptile repository contains code associated with the paper �On First-Order Meta-Learning Algorithms�, which introduces Reptile, a meta-learning algorithm for learning model parameter initializations that adapt quickly to new tasks. The implementation here is aimed at supervised few-shot learning settings (e.g. Omniglot, Mini-ImageNet), not reinforcement learning, and includes scripts to run training and evaluation for few-shot classification. The fundamental idea is: sample a task, train on that task (inner loop), and then move the initialization parameters toward the adapted parameters (outer loop). Because Reptile is a first-order algorithm, it avoids computing second derivatives or full meta-gradients, making it computationally simpler while retaining good performance. The repo includes training scripts, dataset fetchers (Omniglot, Mini-ImageNet), and modules for defining the Reptile update logic, variables, and hyperparameters.

Features:

• Implementation of the Reptile algorithm for few-shot supervised meta-learning
• Support scripts for Omniglot and Mini-ImageNet experiment setups
• First-order meta-learning (no second derivatives) for computational simplicity
• Command-line interface for hyperparameter control (shots, inner/outer loops, meta steps)
• Dataset download / preprocessing utilities (e.g. fetch_data.sh)
• Modular structure: reptile.py, variables.py, dataset modules, experiment drivers

Programming Language: JavaScript.
Categories: