Repository provides nonofficial implementation of Siamese-Networks for the task of one-shot learning in TensorFlow 2.0. Implemenation is based upon Siamese Neural Networks for One-shot Image Recognition paper from Gregory Koch, Richard Zemel and Ruslan Salakhutdinov. Model has been tested on Omniglot dataset.
The implementation presented here was created for comaprison with Protoypical and Matching networks models, consequently uses its data handling and data splits.
- 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
siamnetlib. Runpython setup.py installto install it - To download Omniglot dataset run
bash data/download_omniglot.shfrom repository's root
Repository structured as follows. siamnet contains library with the model and data processing-loading procedures. scripts contains training and evaluation scripts. tests provides minimal tests for training. results folder serves as a directory for text logs destination as well as tensorboard data (by default). Also this folder contains .md file with configuration specifications.
Configuration of training and evaluation procedures is specified by .config files. Most important parameters in config are
data.dataset_pathpath to the directory with datadata.batchbach size (number of 2-way elements in the batch)data.episodesnumber of episodes within each epochdata.cudaflag to use CUDA accelerationtrain.epochsnumber of epochs to traintrain.patiencenumber of allowed epochs without val score improvementtrain.restoreflag to restore model from existing one (model.save_dir) All the other parameters is less configurable or less significant in current implementation.
To run training procedure run the following command from repository's root
- python scripts/train/run_train.py
To run evaluation procedure run the following command from repository's root
- python scripts/eval/run_eval.py
Training procedure differs from the one presented in original paper. Presented training organized as follows. On the every step two classes are selected randomly. For each class 1 example is chosen randomly two times resulting in two sets of samples to compare within. Thus we have 2 examples of two classes from one side and class-corresponding different samples from another. That combination of true-false pairs are sorted out resulting in 4 vs. 4 samples. That 4x4 samples are multiplied by batch size resulting in [batch * 4] & [batch * 4] "samples vectors" with corresponding labels column (1 if samples have the same class, 0 otherwise).
Basic tests can be launched by following command from root directory (for now tests required GPU support)
- python -m unittest tests/*
Presented results are different from paper's due to the difference in neural network architecture and data handling. Relatively modest metrics are caused by no hyperparameters search, absence of data transformations and short training procedure and hopefully will be improved by me in future in my spare time. However, model showed prediction capacity and thus can be improved in near future. For the evaluation phase I used batch of size 1 (which means we gather accuracy from each pair of classes) and averaged metric from 1000 trials (episodes).
| Way | 5-way | 10-way | 20-way |
|---|---|---|---|
| Accuracy | 84.9% | 75.5% | 66.1% |
[1] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov Siamese Neural Networks for One-shot Image Recognition
[2] Brenden M. Lake, Ruslan Salakhutdinov, Joshua B. Tenenbaum The Omniglot Challenge: A 3-Year Progress Report (https://arxiv.org/abs/1902.03477)