Skip to content

Implementation of "Siamese neural networks for one-shot image recognition" via Keras 2.3.1 and TensorFlow backend.

License

Notifications You must be signed in to change notification settings

mervess/oneshot-to-siamese-networks

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

46 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

One-Shot Learning via Siamese Networks with Keras

Siamese twins Siamese twins made with GIMP via mirroring a Siamese cat.

This repository contains the implementation of the code for 2015 ICML paper Siamese Neural Networks for One-shot Image Recognition via Keras 2.3.1 with TensorFlow backend.

This is an update to the fork's origin. I prefer to study on it since, in my opinion, it's been the most straightforward implementation about the paper so far. The accuracy level could be up to 83% with this version caused by the optimizer. I'm planning to work on it when I have the time.

Setup

Necessary Python Packages

I use Python 3.5.2 on Linux, necessary packages are listed below along with their versions for reference.

  • matplotlib (3.0.3)
  • numpy (1.17.2)
  • opencv-python (4.1.1.26)
  • keras (2.3.1)
  • tensorflow (2.0.0)
  • scikit-learn (0.21.3)
  • jupyter (1.0.0)
  • dill (0.3.1.1)

You can just run pip install -r requirements.txt to install them all.

Gathering Data

The Siamese network uses omniglot data for training and evaluation. You can find the paper that introduced the data here.

  1. In order to download the dataset, run the command below in a location which will serve as {omniglot_location} :
git clone https://github.com/brendenlake/omniglot.git
  1. In order to achieve the data model to be trained and tested by the network:
python load_data.py --path {omniglot_location} --save {pickles_location}

Generating the Model

After gathering the data, run OneShot_OneStepAtATime.ipynb by following the steps below.

git clone https://github.com/mervess/oneshot-to-siamese-networks.git
cd oneshot-to-siamese-networks
jupyter notebook &

Notes

Beware that the notebook needs to be updated with the path information.

The two images used in the notebook are from the paper to visualize and compare the network with theirs.

The uploaded version of the network in the notebook is trained up until 55% accuracy fyi. Change the iteration variable in the training block (n_iter) for better results.

References

  1. Koch, Gregory, Richard Zemel, and Ruslan Salakhutdinov. "Siamese neural networks for one-shot image recognition." ICML deep learning workshop. Vol. 2. 2015.
  2. Lake, Brenden, et al. "One shot learning of simple visual concepts." Proceedings of the annual meeting of the cognitive science society. Vol. 33. No. 33. 2011.

About

Implementation of "Siamese neural networks for one-shot image recognition" via Keras 2.3.1 and TensorFlow backend.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Languages

  • Jupyter Notebook 98.5%
  • Python 1.5%