Implementation for Cosine similarity for triplet loss in keras

cosine_triplet_loss.py

# Cosine Similarity
def triplet_loss(alpha, emb_dim):
    def loss(y_true, y_pred):
        anc, pos, neg = y_pred[:,:emb_size], y_pred[:,emb_size:2*emb_size], y_pred[:,2*emb_size:]
        distance1 = tf.keras.losses.cosine_similarity(anc, pos)
        distance2 = tf.keras.losses.cosine_similarity(anc, neg)
        return tf.keras.backend.clip(distance1 - distance2 + alpha, 0., None)
    return loss