Up to the date of creating this repository I could not find a decent example of how to use triplet loss and online triplet mining with Keras, so I decided to create this repository to inspire fellow deep learners.
I used the triplet loss / triplet minig related code from omoindrot's tensorflow-triplet-loss repository (implemented in Tensorflow) and created the necessary Keras code around it.
You can see a minimalistic example of online triplet mining with keras in the uploaded notebook.
The only modification you have to do compared to a regular keras pipeline is to flatten the labels's array, as omoidrot's code expects it in a flattened shape. I.e:
#importing omoindrot's function
from triplet_loss import batch_hard_triplet_loss
#defining a loss function that works with Keras
def keras_batch_hard_triplet_loss(labels, y_pred):
labels = K.flatten(labels)
return batch_hard_triplet_loss(labels, y_pred, margin = 0.5)For all further information about what triplet loss is, and what it is good for, I recommend reading omoindrot's blog post and studying his GitHub repository.
Tensorflow
Keras (using Tensorflow backend)
Numpy