This gist is about how to create a AUC metric for tensorflow/Keras

tf_auc.py

def auc_roc(y_true, y_pred):
    # can be any tensorflow metric
    value, update_op = tf.contrib.metrics.streaming_auc(y_pred, y_true)
    # find all variables created for this metric
    metric_vars = [i for i in tf.local_variables() if 'auc_roc' in i.name.split('/')[1]]
    
    # Add metric variables to GLOBAL_VARIABLES collection.
    # They will be initialized for new session.
    for v in metric_vars:
        tf.add_to_collection(tf.GraphKeys.GLOBAL_VARIABLES, v)
    
    # force to update metric values
    with tf.control_dependencies([update_op]):
        value = tf.identity(value)
    return value

Creating a customized function to compute the AUC ROC curve in our keras or tensorflow model.