This repo extends the original MAML code (link, ICML 2017 paper) [1] with CIFAR-FS (CIFAR few shot) classification based on the splits in Luca et al. [2]. It includes code for running the few-shot supervised learning domain experiments, including sinusoid regression, Omniglot classification, MiniImagenet classification, and the newly added CIFAR-FS classification.
For the experiments in the RL domain, see this codebase.
A report detailing the implementation details and reproducibility can be found here.
This code requires the following:
- python 2.* or python 3.*
- TensorFlow v1.0+
These results were achieved using TensorFlow-gpu 1.8.0, CUDA 9.0 and cudnn 7.
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 iterations, with a 95% confidence interval, and 10 finetune steps:
| Dataset, method | this code accuracy |
reported by MAML [1] & R2D2 [2] |
|---|---|---|
| CIFAR-FS, MAML 5-way, 1-shot | 56.8 ± 1.9% | 58.9 ± 1.9% |
| CIFAR-FS, MAML 5-way, 5-shot | 70.8 ± 0.9% | 71.5 ± 1.0% |
| CIFAR-FS, MAML 2-way, 1-shot | 83.1 ± 2.6% | 82.8 ± 2.7% |
| CIFAR-FS, MAML 2-way, 5-shot | 88.5 ± 1.1% | 88.3 ± 1.1% |
| miniImagenet, MAML 5-way, 1-shot | 47.6 ± 1.9% | 48.7 ± 1.8% |
| miniImagenet, MAML 5-way, 5-shot | 63.0 ± 0.9% | 63.1 ± 0.9% |
| miniImagenet, MAML 2-way, 1-shot | 78.8 ± 2.8% | 74.9 ± 3.0% |
| miniImagenet, MAML 2-way, 5-shot | 82.6 ± 1.2% | 84.4 ± 1.2% |
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] Finn, Chelsea, Pieter Abbeel, and Sergey Levine. "Model-agnostic meta-learning for fast adaptation of deep networks." arXiv preprint arXiv:1703.03400 (2017).
[2] Bertinetto, Luca, et al. "Meta-learning with differentiable closed-form solvers." arXiv preprint arXiv:1805.08136 (2018).