def test_network_invocation_with_internal_and_external_logits(self): """Validate that the logit outputs are correct.""" sequence_length = 5 input_width = 512 num_classes = 10 test_object = token_classification.TokenClassification( input_width=input_width, num_classes=num_classes, output='logits') # Create a 3-dimensional input (the first dimension is implicit). sequence_data = tf.keras.Input(shape=(sequence_length, input_width), dtype=tf.float32) output = test_object(sequence_data) model = tf.keras.Model(sequence_data, output) logits_model = tf.keras.Model(test_object.inputs, test_object.logits) batch_size = 3 input_data = 10 * np.random.random_sample( (batch_size, sequence_length, input_width)) outputs = model.predict(input_data) logits = logits_model.predict(input_data) # Ensure that the tensor shapes are correct. expected_output_shape = (batch_size, sequence_length, num_classes) self.assertEqual(expected_output_shape, outputs.shape) self.assertEqual(expected_output_shape, logits.shape) self.assertAllClose(outputs, logits)
def test_network_invocation_with_logit_output(self): """Validate that the logit outputs are correct.""" sequence_length = 5 input_width = 512 num_classes = 10 test_object = token_classification.TokenClassification( input_width=input_width, num_classes=num_classes, output='predictions') logit_object = token_classification.TokenClassification( input_width=input_width, num_classes=num_classes, output='logits') logit_object.set_weights(test_object.get_weights()) # Create a 3-dimensional input (the first dimension is implicit). sequence_data = tf.keras.Input(shape=(sequence_length, input_width), dtype=tf.float32) output = test_object(sequence_data) logit_output = logit_object(sequence_data) model = tf.keras.Model(sequence_data, output) logits_model = tf.keras.Model(sequence_data, logit_output) batch_size = 3 input_data = 10 * np.random.random_sample( (batch_size, sequence_length, input_width)) outputs = model.predict(input_data) logits = logits_model.predict(input_data) # Ensure that the tensor shapes are correct. expected_output_shape = (batch_size, sequence_length, num_classes) self.assertEqual(expected_output_shape, outputs.shape) self.assertEqual(expected_output_shape, logits.shape) # Ensure that the logits, when softmaxed, create the outputs. input_tensor = tf.keras.Input(expected_output_shape[1:]) output_tensor = tf.keras.layers.Activation( tf.nn.log_softmax)(input_tensor) softmax_model = tf.keras.Model(input_tensor, output_tensor) calculated_softmax = softmax_model.predict(logits) self.assertAllClose(outputs, calculated_softmax)
def test_network_creation(self): """Validate that the Keras object can be created.""" sequence_length = 5 input_width = 512 num_classes = 10 test_object = token_classification.TokenClassification( input_width=input_width, num_classes=num_classes) # Create a 3-dimensional input (the first dimension is implicit). sequence_data = tf.keras.Input(shape=(sequence_length, input_width), dtype=tf.float32) output = test_object(sequence_data) # Validate that the outputs are of the expected shape. expected_output_shape = [None, sequence_length, num_classes] self.assertEqual(expected_output_shape, output.shape.as_list())
def test_serialize_deserialize(self): # Create a network object that sets all of its config options. network = token_classification.TokenClassification( input_width=128, num_classes=10, initializer='zeros', output='predictions') # Create another network object from the first object's config. new_network = token_classification.TokenClassification.from_config( network.get_config()) # Validate that the config can be forced to JSON. _ = new_network.to_json() # If the serialization was successful, the new config should match the old. self.assertAllEqual(network.get_config(), new_network.get_config())
def test_network_invocation(self): """Validate that the Keras object can be invoked.""" sequence_length = 5 input_width = 512 num_classes = 10 test_object = token_classification.TokenClassification( input_width=input_width, num_classes=num_classes, output='predictions') # Create a 3-dimensional input (the first dimension is implicit). sequence_data = tf.keras.Input(shape=(sequence_length, input_width), dtype=tf.float32) output = test_object(sequence_data) # Invoke the network as part of a Model. model = tf.keras.Model(sequence_data, output) input_data = 10 * np.random.random_sample( (3, sequence_length, input_width)) _ = model.predict(input_data)
def test_unknown_output_type_fails(self): with self.assertRaisesRegex(ValueError, 'Unknown `output` value "bad".*'): _ = token_classification.TokenClassification(input_width=128, num_classes=10, output='bad')