Prototypical Networks for Few-shot in TensorFlow 2.0

Implementation of Prototypical Networks for Few-shot Learning paper (https://arxiv.org/abs/1703.05175) in TensorFlow 2.0. Model has been tested on Omniglot and miniImagenet datasets with the same splitting as in the paper.

Dependencies and Installation

• The code has been tested on Ubuntu 18.04 with Python 3.6.8 and TensorFflow 2.0.0-alpha0
• The two main dependencies are TensorFlow and Pillow package (Pillow is included in dependencies)
• To install prototf lib run pytnon setup.py install
• Run bash data/download_omniglot.sh from repo's root directory to download Omniglot dataset
• miniImagenet was downloaded from brilliant repo from renmengye (https://github.com/renmengye/few-shot-ssl-public) and placed into data/mini-imagenet folder

Repository Structure

The repository organized as follows. data directory contains scripts for dataset downloading and used as a default directory for datasets. prototf is the library containing the model itself (prototf/models) and logic for datasets loading and processing (prototf/data). scripts directory contains scripts for launching the training. train/run_train.py and eval/run_eval.py launch training and evaluation respectively. tests folder contains basic training procedure on small-valued parameters to check general correctness. results folder contains .md file with current configuration and details of conducted experiments.

Training

• Training and evaluation configurations are specified through config files, each config describes single train+eval evnironment.
• Run python scripts/train/run_train.py --config scripts/config_omniglot.conf to run training on Omniglot with default parameters.
• Run python scripts/train/run_train.py --config scripts/config_miniimagenet.conf to run training on miniImagenet with default parmeters

Evaluating

• Run python scripts/eval/run_eval.py --config scripts/config_omniglot.conf to run evaluation on Omniglot
• Run python scripts/eval/run_eval.py --config scripts/config_miniimagenet.conf to run evaluation on miniImagenet

Tests

• Run python -m unittest tests/test_omniglot.py from repo's root to test Omniglot
• Run python -m unittest tests/test_mini_imagenet.py from repo's root test miniImagenet

Results

Omniglot:

Evnironment	5-way-5-shot	5-way-1-shot	20-way-5-shot	20-way-1shot
Accuracy	99.4%	97.4%	98.4%	92.2%

miniImagenet

Evnironment	5-way-5-shot	5-way-1-shot
Accuracy	66.0%	43.5%

Additional settings can be found in results folder in the root of repository.