I treat this project as an introduction to Siamese Networks before facing a more sophisticated problem, which I will describe in my Bachelor thesis.
Dataset consists of 12500 cat and 12500 dogs jpg images.
https://www.kaggle.com/c/dogs-vs-cats/data
TODO: show example input image and describe what are triples (anchor, negative, positive)
TODO: describe, citing paper
Model is evaluated on a validation set after each epoch. Validation set consists of a group of random triplets. I measure distances of ANCHOR-POSITIVE and ANCHOR-NEGATIVE, and count the ratio of positives being closer than negative to the whole size of validation set.
https://arxiv.org/pdf/1503.03832.pdf
https://towardsdatascience.com/one-shot-learning-with-siamese-networks-using-keras-17f34e75bb3d
https://thelonenutblog.wordpress.com/2017/12/14/do-telecom-networks-dreams-of-siamese-memories/
https://github.com/noelcodella/tripletloss-keras-tensorflow/blob/master/tripletloss.py
https://towardsdatascience.com/lossless-triplet-loss-7e932f990b24
stucking at loss==margin issue
omoindrot/tensorflow-triplet-loss#6
None so far :)
Python 3.6, tensorflow-gpu 1.8.0, GTX 850M
online triples generation: https://omoindrot.github.io/triplet-loss
generators: https://stanford.edu/~shervine/blog/keras-how-to-generate-data-on-the-fly
consider lossless triplet loss function
visualisations like https://youtu.be/uWODMSMTIJQ