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. The implementation is based on official PyTorch version from the author (https://github.com/jakesnell/prototypical-networks). 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 (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 default directory for datasets. prototf is the library containing the model itself (prototf/models) and logit for datasets processing (prototf/data). scripts directory contains scripts for launching the training. train/setup_train.py and eval/setup_eval.py contain all the configurations for training and evaluating respectively. tests contain basic training procedure on small-valued parameters to check that everything will be ok during training. results contains .md file with current results.

Training

• Run python scripts/train/run_train.py to run training procedure on Omniglot with default parameters.
• Training on miniImagenet requried a bit different set of parameters. Training can be launched for example as python scripts/train/run_train.py --model.x_dim '84,84,3' --data.cuda --data.dataset mini-imagenet --data.split ravi --data.train_way 20

Evaluating

• Run python scripts/eval/run_eval.py to run evaluation procedure on test set with default parameters on Omniglot.
• Run python scripts/eval/run_eval.py --model.x_dim '84,84,3' --train.model_path model_miniimagenet.h5 --data.cuda --data.dataset mini-imagenet --data.split ravi --data.test_way 20 to evaluate 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

Experiment results and future work

Conducted experiments uploaded to the results folder in the root of repository. Opening of Issues and getting in touch for implemenation details are welcome (: