def test_throws_exception_when_weights_key_is_missing(self):
losses = {"a": -15, "b": .4, "c": .3}
weights = {"a": tf.constant(0.5),
"b": tf.constant(0.3)} # Misses the key "c".
with self.assertRaises(ValueError):
_ = scalarization.LinearlyScalarizedOptimizer(
problem=ProblemWithConstantLosses(losses), weights=weights)
def test_check_weighted_value_on_constant_losses(self):
weights = {"a": tf.constant(0.5),
"b": tf.constant(0.3),
"c": tf.constant(0.4)}
losses = {"a": -15, "b": .4, "c": .3}
optimizer = scalarization.LinearlyScalarizedOptimizer(
problem=ProblemWithConstantLosses(losses), weights=weights)
loss, _ = optimizer.compute_train_loss_and_update_op(
inputs=dict(), base_optimizer=tf.train.GradientDescentOptimizer(0.))
with self.cached_session() as session:
session.run(tf.initializers.global_variables())
self.assertAllClose(loss,
sum(weights[key] * losses[key] for key in weights))
def test_exception_thrown_when_weights_is_of_invalid_type(self):
losses = {"a": -15, "b": .4, "c": .3}
# Should fail as `weights` is neither a dict nor in the enum.
with self.assertRaises(TypeError):
_ = scalarization.LinearlyScalarizedOptimizer(
problem=ProblemWithConstantLosses(losses), weights=123)