MAML-Pytorch

PyTorch implementation of the supervised learning experiments from the paper: Model-Agnostic Meta-Learning (MAML).

Version 1.0: Both MiniImagenet and Omniglot Datasets are supported! Have Fun~

Version 2.0: Re-write meta learner and basic learner. Solved some serious bugs in version 1.0.

For Tensorflow Implementation, please visit official HERE and simplier version HERE.

For First-Order Approximation Implementation, Reptile namely, please visit HERE.

Platform

• python: 3.x
• Pytorch: 0.4+

MiniImagenet

Howto

For 5-way 1-shot exp., it allocates nearly 6GB GPU memory.

1. download MiniImagenet dataset from here, splitting: train/val/test.csv from here.
2. extract it like:

miniimagenet/
├── images
	├── n0210891500001298.jpg  
	├── n0287152500001298.jpg 
	...
├── test.csv
├── val.csv
└── train.csv

3. modify the path in miniimagenet_train.py:

        mini = MiniImagenet('miniimagenet/', mode='train', n_way=args.n_way, k_shot=args.k_spt,
                    k_query=args.k_qry,
                    batchsz=10000, resize=args.imgsz)
		...
        mini_test = MiniImagenet('miniimagenet/', mode='test', n_way=args.n_way, k_shot=args.k_spt,
                    k_query=args.k_qry,
                    batchsz=100, resize=args.imgsz)

to your actual data path.

4. just run python miniimagenet_train.py and the running screenshot is as follows:

If your reproducation perf. is not so good, maybe you can enlarge your training epoch to get longer training. And MAML is notorious for its hard training. Therefore, this implementation only provide you a basic start point to begin your research. and the performance below is true and achieved on my machine.

Benchmark

Model	Fine Tune	5-way Acc.		20-way Acc.
		1-shot	5-shot	1-shot	5-shot
Matching Nets	N	43.56%	55.31%	17.31%	22.69%
Meta-LSTM		43.44%	60.60%	16.70%	26.06%
MAML	Y	48.7%	63.11%	16.49%	19.29%
Ours	Y	46.2%	60.3%	-	-

Ominiglot

Howto

run python omniglot_train.py, the program will download omniglot dataset automatically.

decrease the value of args.task_num to fit your GPU memory capacity.

For 5-way 1-shot exp., it allocates nearly 3GB GPU memory.

Refer to this Rep.

@misc{MAML_Pytorch,
  author = {Liangqu Long},
  title = {MAML-Pytorch Implementation},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/dragen1860/MAML-Pytorch}},
  commit = {master}
}