Implementation of Reptile (Nichol et al.) [1] in PyTorch.

Meta learning can be understood as finding an efficient model training procedure. For this purpose, the Reptile algorithm attempts to find good model parameter initilization. A good parameter initialization here means the parameters can be easily adapted for many tasks, after a few update steps. Reptile is a variant of fo-MAML (Finn et al.) [2] (1st order Model Agnostic Meta Learning)

Intuition for Reptile and MAML:

Figure by author.

Overview of MAML algorithm:

Figure by author.

Few shot episode setting:

Adapted from https://www.borealisai.com/en/blog/tutorial-2-few-shot-learning-and-meta-learning-i/

Basic usage

Input

Critical input variables are at the top of the Jupyter notebook or python script in the src folder. By default, expected input are images and a label file. datadir contains an images folder with all images in flat structure, i.e. as direct children. datadir also contains a metadata file, having at least 2 columns: filename and label. The column names are specified by 2 variables: filecolumn, labelcolumn

An example is in the link below. info.json file is ignored.

https://github.com/ihsanullah2131/metadl_contrib/tree/master/DataFormat/mini_insect_3

resultdir: location to save result

dataname: result are saved in this folder inside the resultdir

resultprefix: prefix for output file

random_seed: can be set to None

Output

Evaluation metrics are saved in a assessment.csv file in resultdir/metric.

The meta model is saved in resultdir/model

References

[1]Reptile: https://arxiv.org/abs/1803.02999

[2]MAML: http://proceedings.mlr.press/v70/finn17a

Inspired by https://github.com/gebob19/REPTILE-Metalearning/blob/master/omniglot_trainer.ipynb

Related links

Baseline models:

https://github.com/phanav/meta-album