def test_task_graph(self):
with tf.Graph().as_default():
with tf.compat.v1.Session() as sess:
task = ranking.Ranking(
metrics=[tf.keras.metrics.BinaryAccuracy(name="accuracy")],
label_metrics=[tf.keras.metrics.Mean(name="label_mean")],
prediction_metrics=[
tf.keras.metrics.Mean(name="prediction_mean")
])
predictions = tf.constant([[1], [0.3]], dtype=tf.float32)
labels = tf.constant([[1], [1]], dtype=tf.float32)
expected_metrics = {
"accuracy": 0.5,
"label_mean": 1.0,
"prediction_mean": 0.65
}
loss = task(predictions=predictions, labels=labels)
sess.run([var.initializer for var in task.variables])
sess.run(loss)
metrics = {
metric.name: sess.run(metric.result())
for metric in task.metrics
}
self.assertAllClose(expected_metrics, metrics)
def test_task(self):
task = ranking.Ranking(
metrics=[tf.keras.metrics.BinaryAccuracy(name="accuracy")],
label_metrics=[tf.keras.metrics.Mean(name="label_mean")],
prediction_metrics=[tf.keras.metrics.Mean(name="prediction_mean")])
predictions = tf.constant([[1], [0.3]], dtype=tf.float32)
labels = tf.constant([[1], [1]], dtype=tf.float32)
# Standard log loss formula.
expected_loss = -(math.log(1) + math.log(0.3)) / 2.0
expected_metrics = {
"accuracy": 0.5,
"label_mean": 1.0,
"prediction_mean": 0.65
}
loss = task(predictions=predictions, labels=labels)
metrics = {
metric.name: metric.result().numpy()
for metric in task.metrics
}
self.assertIsNotNone(loss)
self.assertAllClose(expected_loss, loss)
self.assertAllClose(expected_metrics, metrics)