Repository provides implementation of Matching Networks for One Shot Learning paper (https://arxiv.org/abs/1606.04080) in Tensorflow 2.0. Model has been tested on Omniglot and miniImagenet datasets.
- Project has been tested on Ubuntu 18.04 with Python 3.6.8 and TensorFflow 2.0.0-alpha0
- The dependencies are Pillow and tqdm libraries, which are included in setup requirements
- Training and evaluating require
matchnetlib. Runpython setup.py installto install it - To download Omniglot dataset run
bash data/download_omniglot.shfrom repository's root - miniImagenet was taken from excellent project (https://github.com/renmengye/few-shot-ssl-public) and placed into data/mini-imagenet folder
The repository organized as follows. matchnet folder contains library with model and data-loading routines. data serves as a default directory for the datasets (change configs to specify different data path). scripts contains training and evaluation scripts. tests provides minimal tests for training. resulst folder contains description of training configurations and results as well as tranining log info.
Configuration of training and evaluation procedures is specified by .config files (specify data.datsaet_path if dataset has different path). Default config files for Omniglot and miniImagenet are omniglot.conf and miniimagenet respectively (omniglot set as a default choice of scripts' arguments). Scripts run_train.py and run_eval.py runs prodcures while setup_train.py and setup_eval.py contain basic logic for model launching.
To run training procedure run the following commands from repository's root
python scripts/train/run_train.py --config scripts/omniglot.conffor Omniglotpython scripts/train/run_train.py --config scripts/miniimagenet.conffor miniImagent
To run evaluation procedure run the following commands from repository's root
python scripts/eval/run_eval.py --config scripts/omniglot.conffor Omniglotpython scripts/eval/run_eval.py --config scripts/miniimagenet.conffor miniImanet
Training procedure generates log file that can be found in results/logs directory after training will be finished. Name of the log file contains date and time and will be printed in stdout in the beginning.
To run basic tests run following command from root directory (for now tests required GPU support)
python -m unittest tests/*
Obtained results for Omniglot after 30 epochs with train (val part was not engaged yet)
| Environment | 5-way-1-shot | 5-way-5-shot | 20-way-1-shot | 20-way-5-shot |
|---|---|---|---|---|
| Accuracy | 97.0% | 99.3% | 91.4% | 97.4% |
- Thanks to Albert Berenguel Centeno (https://github.com/gitabcworld) for his PyTorch implementation which helped me to sort out tough parts of the training procedure.
[1] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, Daan Wierstra Matching Networks for One Shot Learning (https://arxiv.org/abs/1606.04080)
[2] Brenden M. Lake, Ruslan Salakhutdinov, Joshua B. Tenenbaum The Omniglot Challenge: A 3-Year Progress Report (https://arxiv.org/abs/1902.03477)