def test_layer_invocation_with_float16_with_relative_pe( self, use_relative_pe, pe_max_seq_length): tf.keras.mixed_precision.set_global_policy('mixed_float16') test_layer = reuse_transformer.ReuseTransformer( num_attention_heads=10, inner_dim=2048, inner_activation='relu', use_relative_pe=use_relative_pe, pe_max_seq_length=pe_max_seq_length) 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_get_config(self): num_attention_heads = 2 encoder_block = reuse_transformer.ReuseTransformer( num_attention_heads=num_attention_heads, inner_dim=32, inner_activation='relu', output_dropout=0.1, attention_dropout=0.1, use_bias=False, norm_first=True, norm_epsilon=1e-6, inner_dropout=0.1, attention_initializer=tf.keras.initializers.RandomUniform( minval=0., maxval=1.)) encoder_block_config = encoder_block.get_config() new_encoder_block = reuse_transformer.ReuseTransformer.from_config( encoder_block_config) self.assertEqual(encoder_block_config, new_encoder_block.get_config())
def test_several_attention_axes(self, attention_axes): test_layer = reuse_transformer.ReuseTransformer( inner_dim=32, inner_activation='relu', output_dropout=0.1, attention_dropout=0.1, use_bias=False, norm_first=True, norm_epsilon=1e-6, inner_dropout=0.1, num_attention_heads=10, attention_axes=attention_axes) num_rows = 21 num_cols = 13 width = 80 # Create a 3-dimensional input (the first dimension is implicit). data_tensor = tf.keras.Input(shape=(num_rows, num_cols, 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_use_bias_norm_first(self): num_attention_heads = 2 hidden_size = 16 encoder_block = reuse_transformer.ReuseTransformer( num_attention_heads=num_attention_heads, inner_dim=32, inner_activation='relu', output_dropout=0.1, attention_dropout=0.1, use_bias=False, norm_first=True, norm_epsilon=1e-6, inner_dropout=0.1, attention_initializer=tf.keras.initializers.RandomUniform( minval=0., maxval=1.)) # Forward path. dummy_tensor = tf.zeros([2, 4, 16], dtype=tf.float32) dummy_mask = tf.zeros([2, 4, 4], dtype=tf.float32) inputs = [dummy_tensor, dummy_mask] output, _ = encoder_block(inputs) self.assertEqual(output.shape, (2, 4, hidden_size))
def test_layer_invocation_with_mask(self, reuse_attention, return_attention_scores, use_relative_pe): test_layer = reuse_transformer.ReuseTransformer( num_attention_heads=10, inner_dim=2048, inner_activation='relu', reuse_attention=reuse_attention, use_relative_pe=use_relative_pe) 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)) return_scores_tensor = tf.keras.Input(shape=(1, )) reuse_attention_scores = tf.keras.Input(shape=(10, sequence_length, sequence_length)) output_tensor, _ = test_layer( [data_tensor, mask_tensor, reuse_attention_scores]) # Create a model from the test layer. model = tf.keras.Model( ([data_tensor, mask_tensor, reuse_attention_scores ], return_scores_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)) reuse_scores = np.random.rand(batch_size, 10, sequence_length, sequence_length) _ = model.predict([input_data, mask_data, reuse_scores], return_attention_scores)