def testCreateAgentWithPrebuiltPreprocessingLayers(self, agent_class):
dense_layer = tf.keras.layers.Dense(2)
q_net = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
dense_layer)
with self.assertRaisesRegexp(
ValueError, 'shares weights with the original network'):
agent_class(
self._time_step_spec,
self._action_spec,
q_network=q_net,
optimizer=None)
# Explicitly share weights between q and target networks.
# This would be an unusual setup so we check that an error is thrown.
q_target_net = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
dense_layer)
with self.assertRaisesRegexp(
ValueError, 'shares weights with the original network'):
agent_class(
self._time_step_spec,
self._action_spec,
q_network=q_net,
optimizer=None,
target_q_network=q_target_net)
def test_check_matching_networks(self):
layer_1 = tf.keras.layers.Dense(3)
layer_2 = tf.keras.layers.Dense(3)
q_net_1 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
layer_1)
q_net_2 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
layer_2)
common.check_matching_networks(q_net_1, q_net_2)
def test_check_no_shared_variables_expect_fail(self):
dense_layer = tf.keras.layers.Dense(3)
q_net_1 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
dense_layer)
q_net_2 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
dense_layer)
with self.assertRaises(ValueError):
common.check_no_shared_variables(q_net_1, q_net_2)
def test_check_no_shared_variables(self):
layer_1 = tf.keras.layers.Dense(3)
layer_2 = tf.keras.layers.Dense(3)
q_net_1 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec, layer_1)
q_net_2 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec, layer_2)
q_net_1.create_variables()
q_net_2.create_variables()
common.check_no_shared_variables(q_net_1, q_net_2)
def test_check_matching_networks_different_vars(self):
layer_1 = tf.keras.layers.Dense(3)
layer_2 = tf.keras.layers.GRU(3)
q_net_1 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec, layer_1)
q_net_2 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec, layer_2)
q_net_1.create_variables()
q_net_2.create_variables()
with self.assertRaisesRegexp(ValueError, 'Variables lengths do not match'):
common.check_matching_networks(q_net_1, q_net_2)
def test_check_matching_networks_different_shape(self):
layer_1 = tf.keras.layers.Dense(3)
layer_2 = tf.keras.layers.Dense(4)
q_net_1 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
layer_1)
q_net_2 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
layer_2)
with self.assertRaisesRegexp(ValueError,
'Variable dtypes or shapes do not match'):
common.check_matching_networks(q_net_1, q_net_2)
def test_check_matching_networks_different_input_spec(self):
layer_1 = tf.keras.layers.Dense(3)
layer_2 = tf.keras.layers.Dense(3)
q_net_1 = networks_test_utils.KerasLayersNet(self._observation_spec,
self._action_spec,
layer_1)
q_net_2 = networks_test_utils.KerasLayersNet(
tensor_spec.TensorSpec([3], tf.float32), self._action_spec,
layer_2)
with self.assertRaisesRegexp(
ValueError, 'Input tensor specs of network and target network '
'do not match'):
common.check_matching_networks(q_net_1, q_net_2)
def testCreateAgentWithPrebuiltPreprocessingLayers(self, agent_class):
dense_layer = tf.keras.layers.Dense(3)
q_net = networks_test_utils.KerasLayersNet(self._observation_spec[0],
self._action_spec,
dense_layer)
with self.assertRaisesRegexp(
ValueError, 'shares weights with the original network'):
agent_class(self._time_step_spec,
self._action_spec,
q_network=q_net,
optimizer=None)
# Explicitly share weights between q and target networks; this is ok.
q_target_net = networks_test_utils.KerasLayersNet(
self._observation_spec[0], self._action_spec, dense_layer)
agent_class(self._time_step_spec,
self._action_spec,
q_network=q_net,
optimizer=None,
target_q_network=q_target_net)