def test_psl_rule_1(self):
rule_weights = (1.0, )
rule_names = ('rule_1', )
psl_constraints = model.PSLModelMultiWoZ(rule_weights,
rule_names,
config=self.config)
logits = test_util.LOGITS
loss = psl_constraints.rule_1(logits=tf.constant(logits))
self.assertEqual(loss, 1.4)
def test_compute_loss(self):
rule_weights = (1.0, 2.0)
rule_names = ('rule_11', 'rule_12')
psl_constraints = model.PSLModelMultiWoZ(rule_weights,
rule_names,
config=self.config)
logits = test_util.LOGITS
loss = psl_constraints.compute_loss(logits=tf.constant(logits),
data=test_util.FEATURES)
self.assertNear(loss, 0.9, err=1e-6)
def test_psl_rule_12(self):
rule_weights = (1.0, )
rule_names = ('rule_12', )
psl_constraints = model.PSLModelMultiWoZ(rule_weights,
rule_names,
config=self.config)
logits = test_util.LOGITS
loss = psl_constraints.rule_12(logits=tf.constant(logits),
data=test_util.FEATURES)
self.assertNear(loss, 0.1, err=1e-6)
def test_psl_rule_7_run_model(self):
rule_weights = (1.0, )
rule_names = ('rule_7', )
psl_constraints = model.PSLModelMultiWoZ(rule_weights,
rule_names,
config=self.config)
constrained_model = test_util.build_constrained_model([
self.config['max_dialog_size'], self.config['max_utterance_size']
])
constrained_model.fit(self.train_ds,
epochs=self.config['train_epochs'])
logits = eval_model.evaluate_constrained_model(constrained_model,
self.test_ds,
psl_constraints)
predictions = tf.math.argmax(logits[0], axis=-1)
self.assertEqual(predictions[1][2], self.config['class_map']['end'])
self.assertEqual(predictions[2][3], self.config['class_map']['end'])
def test_psl_rule_1_run_model(self):
rule_weights = (1.0, )
rule_names = ('rule_1', )
psl_constraints = model.PSLModelMultiWoZ(rule_weights,
rule_names,
config=self.config)
constrained_model = test_util.build_constrained_model([
self.config['max_dialog_size'], self.config['max_utterance_size']
])
constrained_model.fit(self.train_ds,
epochs=self.config['train_epochs'])
logits = eval_model.evaluate_constrained_model(constrained_model,
self.test_ds,
psl_constraints)
predictions = tf.math.argmax(logits[0], axis=-1)
result = self.check_greet(predictions, self.test_labels[1],
self.config['class_map'])
self.assertTrue(result)
def get_psl_model(dataset: str, rule_names: List[str],
rule_weights: List[float], **kwargs) -> psl_model.PSLModel:
"""Constraints PSL constraint model."""
_check_dataset_supported(dataset)
return psl_model_multiwoz.PSLModelMultiWoZ(rule_weights, rule_names, **kwargs)
def test_loss_shape(self):
config = self._create_test_vrnn_config()
batch_size = 2
labels_1 = tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size)
labels_2 = tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size)
labels_1_mask = tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size)
labels_2_mask = tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size)
model_outputs = [
tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size),
tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size),
tf.keras.Input(shape=(config.max_dialog_length, config.num_states),
batch_size=batch_size),
tf.keras.Input(shape=(config.max_dialog_length, config.num_states),
batch_size=batch_size),
tf.keras.Input(shape=(config.max_dialog_length,
config.vae_cell.max_seq_length, None),
batch_size=batch_size),
tf.keras.Input(shape=(config.max_dialog_length,
config.vae_cell.max_seq_length, None),
batch_size=batch_size),
tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size),
tf.keras.Input(shape=(config.max_dialog_length, None),
batch_size=batch_size),
]
latent_label_id = tf.keras.Input(shape=(config.max_dialog_length, ),
batch_size=batch_size,
dtype=tf.int32)
latent_label_mask = tf.keras.Input(shape=(config.max_dialog_length, ),
batch_size=batch_size)
word_weights = np.ones((config.vocab_size), dtype=np.float32)
psl_config = psl_test_util.TEST_MULTIWOZ_CONFIG
rule_weights = (1.0, )
rule_names = ('rule_1', )
psl_constraints = psl_model_multiwoz.PSLModelMultiWoZ(
rule_weights, rule_names, config=psl_config)
psl_inputs = tf.keras.Input(shape=(config.max_dialog_length, 8),
batch_size=batch_size)
psl_constraint_loss_weight = 0.1
outputs = linear_vrnn.compute_loss(
labels_1,
labels_2,
labels_1_mask,
labels_2_mask,
model_outputs,
latent_label_id,
latent_label_mask,
word_weights,
with_bpr=True,
kl_loss_weight=0.5,
with_bow=True,
bow_loss_weight=0.3,
num_latent_states=config.num_states,
classification_loss_weight=0.8,
psl_constraint_model=psl_constraints,
psl_inputs=psl_inputs,
psl_constraint_loss_weight=psl_constraint_loss_weight)
self.assertLen(outputs, 8)
for loss in outputs[:7]:
self.assertEqual([], loss.shape.as_list())
loss_per_rule = outputs[7]
self.assertLen(loss_per_rule, len(rule_names))
for loss in loss_per_rule:
self.assertEqual([], loss.shape.as_list())
