How openai/supervised-reptile works
This repository implements supervised Reptile meta-learning for Omniglot and Mini-ImageNet: a user selects a dataset and experiment flags, the runner loads few-shot tasks and a classifier, training adapts on sampled tasks and updates the initialization, then evaluation reports accuracy and checkpoints preserve model state. The package separates dataset sampling, model construction, meta-learning, variable-state operations, and evaluation. A browser-side Omniglot classifier is also present. Edges follow the shown runners and training code; browser UI wiring and unsampled modules are not inferred.
Main entry points: run_omniglot.py Omniglot runner · run_miniimagenet.py Mini-ImageNet runner · train.py Training loop · reptile.py Task adaptation
Components 21
Experiment entrypoints
- Omniglot runner
run_omniglot.py - Mini-ImageNet runner
run_miniimagenet.py - Experiment options
supervised_reptile/args.py
Datasets
- Omniglot sampling
supervised_reptile/omniglot.py - Mini-ImageNet sampling
supervised_reptile/miniimagenet.py - Mini-ImageNet metadata
metadata/miniimagenet
Meta-learning
- Classifier models
supervised_reptile/models.py - Training loop
supervised_reptile/train.py - Task adaptation
supervised_reptile/reptile.py - Variable operations
supervised_reptile/variables.py - Few-shot evaluation
supervised_reptile/eval.py - Model checkpoints
supervised_reptile/train.py
Browser demo
- Browser adaptation
web/src/webworker.js - Tensor engine
web/deps/jsnet.js - Browser interface
web/src/ui.js - Image drawing
web/src/drawing.js - Browser evaluation
web/src/evaluator.js - Prediction display
web/src/predictions.js - Demo defaults
web/src/default.js - Model export
web/helpers/export.py
Other components
- Experimenter
Connections 18
- Experimenter → Omniglot runner: runs
README.md - Experimenter → Mini-ImageNet runner: runs
README.md - Omniglot runner → Experiment options: parses options
run_omniglot.py - Mini-ImageNet runner → Experiment options: parses options
run_miniimagenet.py - Omniglot runner → Omniglot sampling: loads and splits
run_omniglot.py - Mini-ImageNet runner → Mini-ImageNet sampling: loads
run_miniimagenet.py - Omniglot runner → Classifier models: constructs model
run_omniglot.py - Mini-ImageNet runner → Classifier models: constructs model
run_miniimagenet.py - Omniglot runner → Training loop: trains
run_omniglot.py - Mini-ImageNet runner → Training loop: trains
run_miniimagenet.py - Omniglot runner → Few-shot evaluation: evaluates
run_omniglot.py - Mini-ImageNet runner → Few-shot evaluation: evaluates
run_miniimagenet.py - Training loop → Task adaptation: runs task steps
supervised_reptile/train.py - Training loop → Variable operations: applies decay
supervised_reptile/train.py - Training loop → Model checkpoints: saves
supervised_reptile/train.py - Few-shot evaluation → Task adaptation: evaluates tasks
supervised_reptile/eval.py - Task adaptation → Variable operations: updates variables
supervised_reptile/reptile.py - Browser adaptation → Tensor engine: uses tensors
web/src/webworker.js