def test_layer_invocation_with_mask(self): test_layer = transformer.Transformer(num_attention_heads=10, intermediate_size=2048, intermediate_activation='relu') sequence_length = 21 width = 80 # Create a 3-dimensional input (the first dimension is implicit). data_tensor = tf.keras.Input(shape=(sequence_length, width)) # Create a 2-dimensional input (the first dimension is implicit). mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length)) output_tensor = test_layer([data_tensor, mask_tensor]) # Create a model from the test layer. model = tf.keras.Model([data_tensor, mask_tensor], output_tensor) # Invoke the model on test data. We can't validate the output data itself # (the NN is too complex) but this will rule out structural runtime errors. batch_size = 6 input_data = 10 * np.random.random_sample( (batch_size, sequence_length, width)) # The attention mask should be of shape (batch, from_seq_len, to_seq_len), # which here is (batch, sequence_length, sequence_length) mask_data = np.random.randint(2, size=(batch_size, sequence_length, sequence_length)) _ = model.predict([input_data, mask_data])
def test_layer_creation(self): test_layer = transformer.Transformer(num_attention_heads=10, intermediate_size=2048, intermediate_activation='relu') sequence_length = 21 width = 80 # Create a 3-dimensional input (the first dimension is implicit). data_tensor = tf.keras.Input(shape=(sequence_length, width)) output_tensor = test_layer(data_tensor) # The default output of a transformer layer should be the same as the input. self.assertEqual(data_tensor.shape.as_list(), output_tensor.shape.as_list())
def test_layer_creation_with_incorrect_mask_fails(self): test_layer = transformer.Transformer(num_attention_heads=10, intermediate_size=2048, intermediate_activation='relu') sequence_length = 21 width = 80 # Create a 3-dimensional input (the first dimension is implicit). data_tensor = tf.keras.Input(shape=(sequence_length, width)) # Create a 2-dimensional input (the first dimension is implicit). mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length - 3)) with self.assertRaisesRegex(ValueError, 'When passing a mask tensor.*'): _ = test_layer([data_tensor, mask_tensor])
def test_dynamic_layer_sequence(self): test_layer = transformer.Transformer( num_attention_heads=10, intermediate_size=2048, intermediate_activation='relu', kernel_initializer=tf.keras.initializers.TruncatedNormal( stddev=0.02)) # Create a 3-dimensional input (the first dimension is implicit). width = 30 input_tensor = tf.keras.Input(shape=(None, width)) output_tensor = test_layer(input_tensor) model = tf.keras.Model(input_tensor, output_tensor) input_length = 17 input_data = np.ones((1, input_length, width)) output_data = model.predict(input_data) self.assertAllEqual([1, input_length, width], output_data.shape)
def test_layer_invocation(self): test_layer = transformer.Transformer(num_attention_heads=10, intermediate_size=2048, intermediate_activation='relu') sequence_length = 21 width = 80 # Create a 3-dimensional input (the first dimension is implicit). data_tensor = tf.keras.Input(shape=(sequence_length, width)) output_tensor = test_layer(data_tensor) # Create a model from the test layer. model = tf.keras.Model(data_tensor, output_tensor) # Invoke the model on test data. We can't validate the output data itself # (the NN is too complex) but this will rule out structural runtime errors. batch_size = 6 input_data = 10 * np.random.random_sample( (batch_size, sequence_length, width)) _ = model.predict(input_data)