triplet_loss.py

def triplet_loss(y_true, y_pred, alpha = 0.2):
    """
    Implementation of the triplet loss function

    Arguments:
    y_true -- true labels, required when you define a loss in Keras, not used in this function.
    y_pred -- python list containing three objects:
            anchor:   the encodings for the anchor data
            positive: the encodings for the positive data (similar to anchor)
            negative: the encodings for the negative data (different from anchor)

    Returns:
    loss -- real number, value of the loss
    """

    anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2]

    # distance between the anchor and the positive
    pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor,positive)))

    # distance between the anchor and the negative
    neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor,negative)))

    # compute loss
    basic_loss = pos_dist-neg_dist+alpha

    loss = tf.maximum(basic_loss,0.0)

    return loss