Internal change

PiperOrigin-RevId: 293958491

official/nlp/albert/export_albert_tfhub.py

@@ -56,7 +56,7 @@ def create_albert_model(
   input_type_ids = tf.keras.layers.Input(
       shape=(None,), dtype=tf.int32, name="input_type_ids")
   transformer_encoder = bert_models.get_transformer_encoder(
-      albert_config, sequence_length=None, float_dtype=tf.float32)
+      albert_config, sequence_length=None)
   sequence_output, pooled_output = transformer_encoder(
       [input_word_ids, input_mask, input_type_ids])
   # To keep consistent with legacy hub modules, the outputs are

official/nlp/bert/bert_models.py

@@ -85,14 +85,12 @@ class BertPretrainLossAndMetricLayer(tf.keras.layers.Layer):
 
 
 def get_transformer_encoder(bert_config,
-                            sequence_length,
-                            float_dtype=tf.float32):
+                            sequence_length):
   """Gets a 'TransformerEncoder' object.
 
   Args:
     bert_config: A 'modeling.BertConfig' or 'modeling.AlbertConfig' object.
     sequence_length: Maximum sequence length of the training data.
-    float_dtype: tf.dtype, tf.float32 or tf.float16.
 
   Returns:
     A networks.TransformerEncoder object.
@@ -110,8 +108,7 @@ def get_transformer_encoder(bert_config,
       max_sequence_length=bert_config.max_position_embeddings,
       type_vocab_size=bert_config.type_vocab_size,
       initializer=tf.keras.initializers.TruncatedNormal(
-          stddev=bert_config.initializer_range),
-      float_dtype=float_dtype.name)
+          stddev=bert_config.initializer_range))
   if isinstance(bert_config, bert_modeling.AlbertConfig):
     kwargs['embedding_width'] = bert_config.embedding_size
     return networks.AlbertTransformerEncoder(**kwargs)
@@ -191,10 +188,9 @@ def pretrain_model(bert_config,
 class BertSquadLogitsLayer(tf.keras.layers.Layer):
   """Returns a layer that computes custom logits for BERT squad model."""
 
-  def __init__(self, initializer=None, float_type=tf.float32, **kwargs):
+  def __init__(self, initializer=None, **kwargs):
     super(BertSquadLogitsLayer, self).__init__(**kwargs)
     self.initializer = initializer
-    self.float_type = float_type
 
   def build(self, unused_input_shapes):
     """Implements build() for the layer."""
@@ -217,14 +213,11 @@ class BertSquadLogitsLayer(tf.keras.layers.Layer):
     logits = tf.keras.backend.reshape(logits, [-1, sequence_length, 2])
     logits = tf.transpose(logits, [2, 0, 1])
     unstacked_logits = tf.unstack(logits, axis=0)
-    if self.float_type == tf.float16:
-      unstacked_logits = tf.cast(unstacked_logits, tf.float32)
     return unstacked_logits[0], unstacked_logits[1]
 
 
 def squad_model(bert_config,
                 max_seq_length,
-                float_type,
                 initializer=None,
                 hub_module_url=None):
   """Returns BERT Squad model along with core BERT model to import weights.
@@ -232,7 +225,6 @@ def squad_model(bert_config,
   Args:
     bert_config: BertConfig, the config defines the core Bert model.
     max_seq_length: integer, the maximum input sequence length.
-    float_type: tf.dtype, tf.float32 or tf.bfloat16.
     initializer: Initializer for the final dense layer in the span labeler.
       Defaulted to TruncatedNormal initializer.
     hub_module_url: TF-Hub path/url to Bert module.
@@ -245,8 +237,7 @@ def squad_model(bert_config,
     initializer = tf.keras.initializers.TruncatedNormal(
         stddev=bert_config.initializer_range)
   if not hub_module_url:
-    bert_encoder = get_transformer_encoder(bert_config, max_seq_length,
-                                           float_type)
+    bert_encoder = get_transformer_encoder(bert_config, max_seq_length)
     return bert_span_labeler.BertSpanLabeler(
         network=bert_encoder, initializer=initializer), bert_encoder
 
@@ -261,7 +252,7 @@ def squad_model(bert_config,
       [input_word_ids, input_mask, input_type_ids])
 
   squad_logits_layer = BertSquadLogitsLayer(
-      initializer=initializer, float_type=float_type, name='squad_logits')
+      initializer=initializer, name='squad_logits')
   start_logits, end_logits = squad_logits_layer(sequence_output)
 
   squad = tf.keras.Model(
@@ -276,7 +267,6 @@ def squad_model(bert_config,
 
 
 def classifier_model(bert_config,
-                     float_type,
                      num_labels,
                      max_seq_length,
                      final_layer_initializer=None,
@@ -289,7 +279,6 @@ def classifier_model(bert_config,
   Args:
     bert_config: BertConfig or AlbertConfig, the config defines the core
       BERT or ALBERT model.
-    float_type: dtype, tf.float32 or tf.bfloat16.
     num_labels: integer, the number of classes.
     max_seq_length: integer, the maximum input sequence length.
     final_layer_initializer: Initializer for final dense layer. Defaulted
@@ -328,8 +317,7 @@ def classifier_model(bert_config,
   output = tf.keras.layers.Dense(
       num_labels,
       kernel_initializer=initializer,
-      name='output',
-      dtype=float_type)(
+      name='output')(
           output)
   return tf.keras.Model(
       inputs={

official/nlp/bert/export_tfhub.py

@@ -54,7 +54,7 @@ def create_bert_model(bert_config: bert_modeling.BertConfig) -> tf.keras.Model:
   input_type_ids = tf.keras.layers.Input(
       shape=(None,), dtype=tf.int32, name="input_type_ids")
   transformer_encoder = bert_models.get_transformer_encoder(
-      bert_config, sequence_length=None, float_dtype=tf.float32)
+      bert_config, sequence_length=None)
   sequence_output, pooled_output = transformer_encoder(
       [input_word_ids, input_mask, input_type_ids])
   # To keep consistent with legacy hub modules, the outputs are

official/nlp/bert/run_classifier.py

@@ -123,7 +123,6 @@ def run_bert_classifier(strategy,
     classifier_model, core_model = (
         bert_models.classifier_model(
             bert_config,
-            tf.float32,
             num_classes,
             max_seq_length,
             hub_module_url=FLAGS.hub_module_url))
@@ -271,8 +270,10 @@ def export_classifier(model_export_path, input_meta_data,
   if not model_dir:
     raise ValueError('Export path is not specified: %s' % model_dir)
 
+  # Export uses float32 for now, even if training uses mixed precision.
+  tf.keras.mixed_precision.experimental.set_policy('float32')
   classifier_model = bert_models.classifier_model(
-      bert_config, tf.float32, input_meta_data['num_labels'],
+      bert_config, input_meta_data['num_labels'],
       input_meta_data['max_seq_length'])[0]
 
   model_saving_utils.export_bert_model(

official/nlp/bert/run_squad.py

@@ -186,7 +186,7 @@ def predict_squad_customized(strategy, input_meta_data, bert_config,
     # Prediction always uses float32, even if training uses mixed precision.
     tf.keras.mixed_precision.experimental.set_policy('float32')
     squad_model, _ = bert_models.squad_model(
-        bert_config, input_meta_data['max_seq_length'], float_type=tf.float32)
+        bert_config, input_meta_data['max_seq_length'])
 
   checkpoint_path = tf.train.latest_checkpoint(FLAGS.model_dir)
   logging.info('Restoring checkpoints from %s', checkpoint_path)
@@ -254,7 +254,6 @@ def train_squad(strategy,
     squad_model, core_model = bert_models.squad_model(
         bert_config,
         max_seq_length,
-        float_type=tf.float16 if use_float16 else tf.float32,
         hub_module_url=FLAGS.hub_module_url)
     squad_model.optimizer = optimization.create_optimizer(
         FLAGS.learning_rate, steps_per_epoch * epochs, warmup_steps)
@@ -389,8 +388,10 @@ def export_squad(model_export_path, input_meta_data):
     raise ValueError('Export path is not specified: %s' % model_export_path)
   bert_config = MODEL_CLASSES[FLAGS.model_type][0].from_json_file(
       FLAGS.bert_config_file)
-  squad_model, _ = bert_models.squad_model(
-      bert_config, input_meta_data['max_seq_length'], float_type=tf.float32)
+  # Export uses float32 for now, even if training uses mixed precision.
+  tf.keras.mixed_precision.experimental.set_policy('float32')
+  squad_model, _ = bert_models.squad_model(bert_config,
+                                           input_meta_data['max_seq_length'])
   model_saving_utils.export_bert_model(
       model_export_path, model=squad_model, checkpoint_dir=FLAGS.model_dir)
 

official/nlp/modeling/layers/transformer.py

@@ -123,6 +123,8 @@ class Transformer(tf.keras.layers.Layer):
         bias_constraint=self._bias_constraint,
         name="self_attention_output")
     self._attention_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
+    # Use float32 in layernorm for numeric stability.
+    # It is probably safe in mixed_float16, but we haven't validated this yet.
     self._attention_layer_norm = (
         tf.keras.layers.LayerNormalization(
             name="self_attention_layer_norm",
@@ -140,8 +142,10 @@ class Transformer(tf.keras.layers.Layer):
         kernel_constraint=self._kernel_constraint,
         bias_constraint=self._bias_constraint,
         name="intermediate")
+    # Use float32 in intermediate gelu activation for numeric stability.
+    # TODO(b/149117297): investigate gelu numeric stability.
     self._intermediate_activation_layer = tf.keras.layers.Activation(
-        self._intermediate_activation)
+        self._intermediate_activation, dtype=tf.float32)
     self._output_dense = dense_einsum.DenseEinsum(
         output_shape=hidden_size,
         kernel_initializer=self._kernel_initializer,
@@ -153,6 +157,7 @@ class Transformer(tf.keras.layers.Layer):
         bias_constraint=self._bias_constraint,
         name="output")
     self._output_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
+    # Use float32 in layernorm for numeric stability.
     self._output_layer_norm = tf.keras.layers.LayerNormalization(
         name="output_layer_norm", axis=-1, epsilon=1e-12, dtype=tf.float32)
 
@@ -202,30 +207,16 @@ class Transformer(tf.keras.layers.Layer):
     attention_output = self._attention_layer(attention_inputs)
     attention_output = self._attention_output_dense(attention_output)
     attention_output = self._attention_dropout(attention_output)
-    # Use float32 in keras layer norm and the gelu activation in the
-    # intermediate dense layer for numeric stability
-    if self.dtype == tf.float16:
-      input_tensor = tf.cast(input_tensor, tf.float32)
-      attention_output = tf.cast(attention_output, tf.float32)
     attention_output = self._attention_layer_norm(input_tensor +
                                                   attention_output)
     intermediate_output = self._intermediate_dense(attention_output)
-    if self.dtype == tf.float16:
-      # Casts to float32 so that activation is done in float32.
-      intermediate_output = tf.cast(intermediate_output, tf.float32)
-      intermediate_output = self._intermediate_activation_layer(
-          intermediate_output)
-      intermediate_output = tf.cast(intermediate_output, tf.float16)
-    else:
-      intermediate_output = self._intermediate_activation_layer(
-          intermediate_output)
+    intermediate_output = self._intermediate_activation_layer(
+        intermediate_output)
     layer_output = self._output_dense(intermediate_output)
     layer_output = self._output_dropout(layer_output)
-    # Use float32 in keras layer norm for numeric stability
-    if self.dtype == tf.float16:
-      layer_output = tf.cast(layer_output, tf.float32)
+    # During mixed precision training, attention_output is from layer norm and
+    # is always fp32 for now. cast layer_output to fp32 for the subsequent add.
+    layer_output = tf.cast(layer_output, tf.float32)
     layer_output = self._output_layer_norm(layer_output + attention_output)
-    if self.dtype == tf.float16:
-      layer_output = tf.cast(layer_output, tf.float16)
 
     return layer_output

official/nlp/modeling/layers/transformer_test.py

@@ -30,6 +30,10 @@ from official.nlp.modeling.layers import transformer
 @keras_parameterized.run_all_keras_modes
 class TransformerLayerTest(keras_parameterized.TestCase):
 
+  def tearDown(self):
+    super(TransformerLayerTest, self).tearDown()
+    tf.keras.mixed_precision.experimental.set_policy('float32')
+
   def test_layer_creation(self):
     test_layer = transformer.Transformer(
         num_attention_heads=10,
@@ -121,16 +125,15 @@ class TransformerLayerTest(keras_parameterized.TestCase):
     _ = model.predict([input_data, mask_data])
 
   def test_layer_invocation_with_float16_dtype(self):
+    tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
     test_layer = transformer.Transformer(
         num_attention_heads=10,
         intermediate_size=2048,
-        intermediate_activation='relu',
-        dtype='float16')
+        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), dtype=tf.float16)
+    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])
@@ -142,7 +145,7 @@ class TransformerLayerTest(keras_parameterized.TestCase):
     # (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))).astype(np.float16)
+        (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(
@@ -165,4 +168,5 @@ class TransformerLayerTest(keras_parameterized.TestCase):
 
 
 if __name__ == '__main__':
+  assert tf.version.VERSION.startswith('2.')
   tf.test.main()

official/nlp/modeling/networks/albert_transformer_encoder.py

@@ -65,7 +65,6 @@ class AlbertTransformerEncoder(network.Network):
     attention_dropout_rate: The dropout rate to use for the attention layers
       within the transformer layers.
     initializer: The initialzer to use for all weights in this encoder.
-    float_dtype: The dtype of this encoder. Can be 'float32' or 'float16'.
   """
 
   def __init__(self,
@@ -82,7 +81,6 @@ class AlbertTransformerEncoder(network.Network):
                dropout_rate=0.1,
                attention_dropout_rate=0.1,
                initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02),
-               float_dtype='float32',
                **kwargs):
     activation = tf.keras.activations.get(activation)
     initializer = tf.keras.initializers.get(initializer)
@@ -104,7 +102,6 @@ class AlbertTransformerEncoder(network.Network):
         'dropout_rate': dropout_rate,
         'attention_dropout_rate': attention_dropout_rate,
         'initializer': tf.keras.initializers.serialize(initializer),
-        'float_dtype': float_dtype,
     }
 
     word_ids = tf.keras.layers.Input(
@@ -118,7 +115,6 @@ class AlbertTransformerEncoder(network.Network):
         vocab_size=vocab_size,
         embedding_width=embedding_width,
         initializer=initializer,
-        dtype=float_dtype,
         name='word_embeddings')
     word_embeddings = self._embedding_layer(word_ids)
 
@@ -126,8 +122,7 @@ class AlbertTransformerEncoder(network.Network):
     self._position_embedding_layer = layers.PositionEmbedding(
         initializer=initializer,
         use_dynamic_slicing=True,
-        max_sequence_length=max_sequence_length,
-        dtype=float_dtype)
+        max_sequence_length=max_sequence_length)
     position_embeddings = self._position_embedding_layer(word_embeddings)
 
     type_embeddings = (
@@ -136,7 +131,6 @@ class AlbertTransformerEncoder(network.Network):
             embedding_width=embedding_width,
             initializer=initializer,
             use_one_hot=True,
-            dtype=float_dtype,
             name='type_embeddings')(type_ids))
 
     embeddings = tf.keras.layers.Add()(
@@ -146,10 +140,9 @@ class AlbertTransformerEncoder(network.Network):
             name='embeddings/layer_norm',
             axis=-1,
             epsilon=1e-12,
-            dtype=float_dtype)(embeddings))
+            dtype=tf.float32)(embeddings))
     embeddings = (
-        tf.keras.layers.Dropout(rate=dropout_rate,
-                                dtype=tf.float32)(embeddings))
+        tf.keras.layers.Dropout(rate=dropout_rate)(embeddings))
     # We project the 'embedding' output to 'hidden_size' if it is not already
     # 'hidden_size'.
     if embedding_width != hidden_size:
@@ -159,9 +152,6 @@ class AlbertTransformerEncoder(network.Network):
           name='embedding_projection')(
               embeddings)
 
-    if float_dtype == 'float16':
-      embeddings = tf.cast(embeddings, tf.float16)
-
     data = embeddings
     attention_mask = layers.SelfAttentionMask()([data, mask])
     shared_layer = layers.Transformer(
@@ -171,7 +161,6 @@ class AlbertTransformerEncoder(network.Network):
         dropout_rate=dropout_rate,
         attention_dropout_rate=attention_dropout_rate,
         kernel_initializer=initializer,
-        dtype=float_dtype,
         name='transformer')
     for _ in range(num_layers):
       data = shared_layer([data, attention_mask])
@@ -183,7 +172,6 @@ class AlbertTransformerEncoder(network.Network):
         units=hidden_size,
         activation='tanh',
         kernel_initializer=initializer,
-        dtype=float_dtype,
         name='pooler_transform')(
             first_token_tensor)
 

official/nlp/modeling/networks/albert_transformer_encoder_test.py

@@ -31,14 +31,17 @@ from official.nlp.modeling.networks import albert_transformer_encoder
 @keras_parameterized.run_all_keras_modes
 class AlbertTransformerEncoderTest(keras_parameterized.TestCase):
 
+  def tearDown(self):
+    super(AlbertTransformerEncoderTest, self).tearDown()
+    tf.keras.mixed_precision.experimental.set_policy("float32")
+
   @parameterized.named_parameters(
       dict(testcase_name="default", expected_dtype=tf.float32),
       dict(
           testcase_name="with_float16_dtype",
-          expected_dtype=tf.float16,
-          float_dtype="float16"),
+          expected_dtype=tf.float16),
   )
-  def test_network_creation(self, expected_dtype, float_dtype=None):
+  def test_network_creation(self, expected_dtype):
     hidden_size = 32
     sequence_length = 21
 
@@ -48,8 +51,8 @@ class AlbertTransformerEncoderTest(keras_parameterized.TestCase):
         sequence_length=sequence_length,
         num_attention_heads=2,
         num_layers=3)
-    if float_dtype is not None:
-      kwargs["float_dtype"] = float_dtype
+    if expected_dtype == tf.float16:
+      tf.keras.mixed_precision.experimental.set_policy("mixed_float16")
 
     # Create a small TransformerEncoder for testing.
     test_network = albert_transformer_encoder.AlbertTransformerEncoder(**kwargs)
@@ -65,7 +68,9 @@ class AlbertTransformerEncoderTest(keras_parameterized.TestCase):
     self.assertAllEqual(expected_data_shape, data.shape.as_list())
     self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
 
-    self.assertEqual(expected_dtype, data.dtype)
+    # If float_dtype is set to float16, the data output is float32 (from a layer
+    # norm) and pool output should be float16.
+    self.assertEqual(tf.float32, data.dtype)
     self.assertEqual(expected_dtype, pooled.dtype)
 
     # ALBERT has additonal 'embedding_hidden_mapping_in' weights and
@@ -128,6 +133,7 @@ class AlbertTransformerEncoderTest(keras_parameterized.TestCase):
     _ = model.predict([word_id_data, mask_data, type_id_data])
 
   def test_serialize_deserialize(self):
+    tf.keras.mixed_precision.experimental.set_policy("mixed_float16")
     # Create a network object that sets all of its config options.
     kwargs = dict(
         vocab_size=100,
@@ -142,8 +148,7 @@ class AlbertTransformerEncoderTest(keras_parameterized.TestCase):
         activation="relu",
         dropout_rate=0.05,
         attention_dropout_rate=0.22,
-        initializer="glorot_uniform",
-        float_dtype="float16")
+        initializer="glorot_uniform")
     network = albert_transformer_encoder.AlbertTransformerEncoder(**kwargs)
 
     expected_config = dict(kwargs)
@@ -166,4 +171,5 @@ class AlbertTransformerEncoderTest(keras_parameterized.TestCase):
 
 
 if __name__ == "__main__":
+  assert tf.version.VERSION.startswith('2.')
   tf.test.main()

official/nlp/modeling/networks/encoder_scaffold.py

@@ -50,7 +50,6 @@ class EncoderScaffold(network.Network):
     num_output_classes: The output size of the classification layer.
     classification_layer_initializer: The initializer for the classification
       layer.
-    classification_layer_dtype: The dtype for the classification layer.
     embedding_cls: The class or instance to use to embed the input data. This
       class or instance defines the inputs to this encoder. If embedding_cls is
       not set, a default embedding network (from the original BERT paper) will
@@ -65,7 +64,6 @@ class EncoderScaffold(network.Network):
       "seq_length": The sequence length for this encoder.
       "initializer": The initializer for the embedding portion of this encoder.
       "dropout_rate": The dropout rate to apply before the encoding layers.
-      "dtype": (Optional): The dtype of the embedding layers.
     embedding_data: A reference to the embedding weights that will be used to
       train the masked language model, if necessary. This is optional, and only
       needed if (1) you are overriding embedding_cls and (2) are doing standard
@@ -84,7 +82,6 @@ class EncoderScaffold(network.Network):
         "dropout_rate": The overall dropout rate for the transformer layers.
         "attention_dropout_rate": The dropout rate for the attention layers.
         "kernel_initializer": The initializer for the transformer layers.
-        "dtype": The dtype of the transformer.
   """
 
   def __init__(
@@ -92,7 +89,6 @@ class EncoderScaffold(network.Network):
       num_output_classes,
       classification_layer_initializer=tf.keras.initializers.TruncatedNormal(
           stddev=0.02),
-      classification_layer_dtype=tf.float32,
       embedding_cls=None,
       embedding_cfg=None,
       embedding_data=None,
@@ -168,10 +164,7 @@ class EncoderScaffold(network.Network):
               dtype=tf.float32)(embeddings))
       embeddings = (
           tf.keras.layers.Dropout(
-              rate=embedding_cfg['dropout_rate'], dtype=tf.float32)(embeddings))
-
-      if embedding_cfg.get('dtype') == 'float16':
-        embeddings = tf.cast(embeddings, tf.float16)
+              rate=embedding_cfg['dropout_rate'])(embeddings))
 
     attention_mask = layers.SelfAttentionMask()([embeddings, mask])
     data = embeddings
@@ -190,7 +183,6 @@ class EncoderScaffold(network.Network):
         units=num_output_classes,
         activation='tanh',
         kernel_initializer=classification_layer_initializer,
-        dtype=classification_layer_dtype,
         name='cls_transform')(
             first_token_tensor)
 

official/nlp/modeling/networks/encoder_scaffold_test.py

@@ -53,6 +53,10 @@ class ValidatedTransformerLayer(layers.Transformer):
 @keras_parameterized.run_all_keras_modes
 class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
 
+  def tearDown(self):
+    super(EncoderScaffoldLayerClassTest, self).tearDown()
+    tf.keras.mixed_precision.experimental.set_policy("float32")
+
   def test_network_creation(self):
     hidden_size = 32
     sequence_length = 21
@@ -81,8 +85,6 @@ class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float32",
         "call_list":
             call_list
     }
@@ -127,7 +129,6 @@ class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
         "max_seq_length": sequence_length,
         "initializer": tf.keras.initializers.TruncatedNormal(stddev=0.02),
         "dropout_rate": 0.1,
-        "dtype": "float16",
     }
     hidden_cfg = {
         "num_attention_heads":
@@ -142,8 +143,6 @@ class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float16",
     }
     # Create a small EncoderScaffold for testing.
     test_network = encoder_scaffold.EncoderScaffold(
@@ -151,7 +150,6 @@ class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
         num_output_classes=hidden_size,
         classification_layer_initializer=tf.keras.initializers.TruncatedNormal(
             stddev=0.02),
-        classification_layer_dtype=tf.float16,
         hidden_cfg=hidden_cfg,
         embedding_cfg=embedding_cfg)
     # Create the inputs (note that the first dimension is implicit).
@@ -165,8 +163,9 @@ class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
     self.assertAllEqual(expected_data_shape, data.shape.as_list())
     self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
 
-    # If float_dtype is set to float16, the output should always be float16.
-    self.assertAllEqual(tf.float16, data.dtype)
+    # If float_dtype is set to float16, the data output is float32 (from a layer
+    # norm) and pool output should be float16.
+    self.assertAllEqual(tf.float32, data.dtype)
     self.assertAllEqual(tf.float16, pooled.dtype)
 
   def test_network_invocation(self):
@@ -196,10 +195,7 @@ class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float32",
     }
-    tf.keras.mixed_precision.experimental.set_policy("float32")
     print(hidden_cfg)
     print(embedding_cfg)
     # Create a small EncoderScaffold for testing.
@@ -293,8 +289,6 @@ class EncoderScaffoldLayerClassTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float32",
     }
     # Create a small EncoderScaffold for testing.
     network = encoder_scaffold.EncoderScaffold(
@@ -353,8 +347,6 @@ class EncoderScaffoldEmbeddingNetworkTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float32",
     }
 
     # Create a small EncoderScaffold for testing.
@@ -422,8 +414,6 @@ class EncoderScaffoldEmbeddingNetworkTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float32",
     }
 
     # Create a small EncoderScaffold for testing.
@@ -493,7 +483,6 @@ class EncoderScaffoldHiddenInstanceTest(keras_parameterized.TestCase):
         "max_seq_length": sequence_length,
         "initializer": tf.keras.initializers.TruncatedNormal(stddev=0.02),
         "dropout_rate": 0.1,
-        "dtype": "float32",
     }
 
     call_list = []
@@ -510,8 +499,6 @@ class EncoderScaffoldHiddenInstanceTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float32",
         "call_list":
             call_list
     }
@@ -566,7 +553,6 @@ class EncoderScaffoldHiddenInstanceTest(keras_parameterized.TestCase):
         "max_seq_length": sequence_length,
         "initializer": tf.keras.initializers.TruncatedNormal(stddev=0.02),
         "dropout_rate": 0.1,
-        "dtype": "float32",
     }
 
     call_list = []
@@ -583,8 +569,6 @@ class EncoderScaffoldHiddenInstanceTest(keras_parameterized.TestCase):
             0.1,
         "kernel_initializer":
             tf.keras.initializers.TruncatedNormal(stddev=0.02),
-        "dtype":
-            "float32",
         "call_list":
             call_list
     }

official/nlp/modeling/networks/transformer_encoder.py

@@ -59,7 +59,6 @@ class TransformerEncoder(network.Network):
     attention_dropout_rate: The dropout rate to use for the attention layers
       within the transformer layers.
     initializer: The initialzer to use for all weights in this encoder.
-    float_dtype: The dtype of this encoder. Can be 'float32' or 'float16'.
     return_all_encoder_outputs: Whether to output sequence embedding outputs of
       all encoder transformer layers.
   """
@@ -77,7 +76,6 @@ class TransformerEncoder(network.Network):
                dropout_rate=0.1,
                attention_dropout_rate=0.1,
                initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02),
-               float_dtype='float32',
                return_all_encoder_outputs=False,
                **kwargs):
     activation = tf.keras.activations.get(activation)
@@ -99,7 +97,6 @@ class TransformerEncoder(network.Network):
         'dropout_rate': dropout_rate,
         'attention_dropout_rate': attention_dropout_rate,
         'initializer': tf.keras.initializers.serialize(initializer),
-        'float_dtype': float_dtype,
         'return_all_encoder_outputs': return_all_encoder_outputs,
     }
 
@@ -141,11 +138,7 @@ class TransformerEncoder(network.Network):
             epsilon=1e-12,
             dtype=tf.float32)(embeddings))
     embeddings = (
-        tf.keras.layers.Dropout(rate=dropout_rate,
-                                dtype=tf.float32)(embeddings))
-
-    if float_dtype == 'float16':
-      embeddings = tf.cast(embeddings, tf.float16)
+        tf.keras.layers.Dropout(rate=dropout_rate)(embeddings))
 
     self._transformer_layers = []
     data = embeddings
@@ -159,7 +152,6 @@ class TransformerEncoder(network.Network):
           dropout_rate=dropout_rate,
           attention_dropout_rate=attention_dropout_rate,
           kernel_initializer=initializer,
-          dtype=float_dtype,
           name='transformer/layer_%d' % i)
       self._transformer_layers.append(layer)
       data = layer([data, attention_mask])

official/nlp/modeling/networks/transformer_encoder_test.py

@@ -30,6 +30,10 @@ from official.nlp.modeling.networks import transformer_encoder
 @keras_parameterized.run_all_keras_modes
 class TransformerEncoderTest(keras_parameterized.TestCase):
 
+  def tearDown(self):
+    super(TransformerEncoderTest, self).tearDown()
+    tf.keras.mixed_precision.experimental.set_policy("float32")
+
   def test_network_creation(self):
     hidden_size = 32
     sequence_length = 21
@@ -93,8 +97,7 @@ class TransformerEncoderTest(keras_parameterized.TestCase):
         hidden_size=hidden_size,
         sequence_length=sequence_length,
         num_attention_heads=2,
-        num_layers=3,
-        float_dtype="float16")
+        num_layers=3)
     # Create the inputs (note that the first dimension is implicit).
     word_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
     mask = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
@@ -106,8 +109,9 @@ class TransformerEncoderTest(keras_parameterized.TestCase):
     self.assertAllEqual(expected_data_shape, data.shape.as_list())
     self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
 
-    # If float_dtype is set to float16, the output should always be float16.
-    self.assertAllEqual(tf.float16, data.dtype)
+    # If float_dtype is set to float16, the data output is float32 (from a layer
+    # norm) and pool output should be float16.
+    self.assertAllEqual(tf.float32, data.dtype)
     self.assertAllEqual(tf.float16, pooled.dtype)
 
   def test_network_invocation(self):
@@ -115,7 +119,6 @@ class TransformerEncoderTest(keras_parameterized.TestCase):
     sequence_length = 21
     vocab_size = 57
     num_types = 7
-    tf.keras.mixed_precision.experimental.set_policy("float32")
     # Create a small TransformerEncoder for testing.
     test_network = transformer_encoder.TransformerEncoder(
         vocab_size=vocab_size,
@@ -160,6 +163,7 @@ class TransformerEncoderTest(keras_parameterized.TestCase):
     _ = model.predict([word_id_data, mask_data, type_id_data])
 
   def test_serialize_deserialize(self):
+    tf.keras.mixed_precision.experimental.set_policy("mixed_float16")
     # Create a network object that sets all of its config options.
     kwargs = dict(
         vocab_size=100,
@@ -174,7 +178,6 @@ class TransformerEncoderTest(keras_parameterized.TestCase):
         dropout_rate=0.05,
         attention_dropout_rate=0.22,
         initializer="glorot_uniform",
-        float_dtype="float16",
         return_all_encoder_outputs=False)
     network = transformer_encoder.TransformerEncoder(**kwargs)