accuracy metric for online triplet

recall_top_k.py

def recall_top_k(y_true, y_pred):
    # get batch size
    batch_size = y_pred.shape[0]
    
    # get pairwise distances
    dists  = _pairwise_distances(y_pred)
    
    # get indexes of the closest item in the batch predictions 
    # (item itself is always the top_k, so hence getting the second top K)
    # top_k function gives indexes for the highest values so subtracting from 2(or a high enough number) to invert the matrix
    _, pred_indexes = tf.nn.top_k(2-dists, 2, sorted=False)
    
    # get the indexes of truth vector, multiplying with inverted identity matrix to mask out the item itself
    _, true_indexes = tf.nn.top_k((1-tf.eye(batch_size))*y_true)
    
    # sum of equality divided by batchsize to get accuracy between 0 to 1
    return tf.math.reduce_sum(tf.cast(tf.equal(true_indexes[:,0], true_indexes[:,0]), tf.float32))/batch_size