This repo implements the paper "Meta-learning with differentiable closed-form solvers" [1] in TensorFlow. It builds on code from MAML (link) [2].
A paper with a reproducible explanation of implementation details of this code can be found here (oral presentation @ ICLR 2019 RML).
This code requires the following:
- python 2.* or python 3.*
- TensorFlow v1.0+
For the Omniglot, MiniImagenet and CIFAR-FS data, see the usage instructions in data/omniglot_resized/resize_images.py and data/miniImagenet/proc_images.py and data/CIFARFS/get_cifarfs.py respectively.
To run the code, see the usage instructions at the top of main.py.
After 60,000 (60k) iterations, with a 95% confidence interval:
| Dataset, method | this code MAML model accuracy |
this code R2D2 model accuracy |
reported by R2D2 [2] |
|---|---|---|---|
| CIFAR-FS, R2D2 5-way, 1-shot | 57.0 ± 1.7% | 62.7 ± 1.8% | 65.3 ± 0.2% |
| CIFAR-FS, R2D2 5-way, 5-shot | 68.9 ± 0.9% | 75.1 ± 0.9% | 79.4 ± 0.1% |
| CIFAR-FS, R2D2 2-way, 1-shot | 82.4 ± 2.6% | 87.3 ± 2.3% | 83.4 ± 0.3% |
| CIFAR-FS, R2D2 2-way, 5-shot | 88.0 ± 1.1% | 90.7 ± 1.0% | 91.1 ± 0.2% |
| miniImagenet, R2D2 5-way, 1-shot | 48.1 ± 1.8% | 51.7 ± 1.8% | 51.5 ± 0.2% |
| miniImagenet, R2D2 5-way, 5-shot | 63.1 ± 0.9% | 66.2 ± 0.9% | 68.8 ± 0.2% |
| miniImagenet, R2D2 2-way, 1-shot | 77.3 ± 2.8% | 79.5 ± 2.6% | 76.7 ± 0.3% |
| miniImagenet, R2D2 2-way, 5-shot | 85.4 ± 1.1% | 87.3 ± 1.1% | 86.8 ± 0.2% |
If you use (part of) this code or work, please cite the following work:
@article{devosreproducing,
title={Reproducing Meta-learning with differentiable closed-form solvers},
author={Devos, Arnout and Chatel, Sylvain and Grossglauser, Matthias},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=BJx0N2I6IN},
}
[1] Bertinetto, Luca, et al. "Meta-learning with differentiable closed-form solvers." arXiv preprint arXiv:1805.08136 (2018).
[2] Finn, Chelsea, Pieter Abbeel, and Sergey Levine. "Model-agnostic meta-learning for fast adaptation of deep networks." arXiv preprint arXiv:1703.03400 (2017). 7