Merge branch 'panoptic-segmentation' into panoptic-segmentation

official/modeling/hyperparams/base_config.py

@@ -49,6 +49,11 @@ class Config(params_dict.ParamsDict):
   default_params: dataclasses.InitVar[Optional[Mapping[str, Any]]] = None
   restrictions: dataclasses.InitVar[Optional[List[str]]] = None
 
+  def __post_init__(self, default_params, restrictions):
+    super().__init__(
+        default_params=default_params,
+        restrictions=restrictions)
+
   @classmethod
   def _isvalidsequence(cls, v):
     """Check if the input values are valid sequences.
@@ -140,13 +145,6 @@ class Config(params_dict.ParamsDict):
               else subconfig_type)
     return subconfig_type
 
-  def __post_init__(self, default_params, restrictions, *args, **kwargs):
-    super().__init__(
-        default_params=default_params,
-        restrictions=restrictions,
-        *args,
-        **kwargs)
-
   def _set(self, k, v):
     """Overrides same method in ParamsDict.
 

official/modeling/multitask/base_model.py

@@ -12,21 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Lint as: python3
-# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
 """Abstraction of multi-task model."""
 from typing import Text, Dict
 

official/modeling/multitask/base_trainer.py

@@ -12,21 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Lint as: python3
-# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
 """Multitask base trainer implementation.
 
 The trainer derives from the Orbit `StandardTrainer` class.

official/modeling/multitask/evaluator.py

@@ -54,8 +54,15 @@ class MultiTaskEvaluator(orbit.AbstractEvaluator):
     self._model = model
     self._global_step = global_step or orbit.utils.create_global_step()
     self._checkpoint_exporter = checkpoint_exporter
+    if hasattr(self.model, "checkpoint_items"):
+      checkpoint_items = self.model.checkpoint_items
+    else:
+      checkpoint_items = {}
+
     self._checkpoint = tf.train.Checkpoint(
-        global_step=self.global_step, model=self.model)
+        model=self.model,
+        global_step=self.global_step,
+        **checkpoint_items)
 
     self._validation_losses = None
     self._validation_metrics = None

official/modeling/multitask/task_sampler.py

@@ -78,7 +78,10 @@ class ProportionalTaskSampler(TaskSampler):
 
 
 class AnnealingTaskSampler(TaskSampler):
-  """Sample tasks according to task weights as well as training progress."""
+  """Sample tasks according to task weights as well as training progress.
+
+  See http://proceedings.mlr.press/v97/stickland19a/stickland19a.pdf
+  """
 
   def __init__(self,
                task_weights: Dict[Text, Union[float, int]],

official/nlp/bert/model_training_utils.py

@@ -22,7 +22,7 @@ from absl import logging
 import tensorflow as tf
 from tensorflow.python.util import deprecation
 from official.common import distribute_utils
-from official.staging.training import grad_utils
+from official.modeling import grad_utils
 
 _SUMMARY_TXT = 'training_summary.txt'
 _MIN_SUMMARY_STEPS = 10

official/nlp/configs/experiment_configs.py

@@ -17,3 +17,4 @@
 from official.nlp.configs import finetuning_experiments
 from official.nlp.configs import pretraining_experiments
 from official.nlp.configs import wmt_transformer_experiments
+from official.nlp.projects.teams import teams_experiments

official/nlp/data/pretrain_dynamic_dataloader_test.py

@@ -43,10 +43,11 @@ def _create_fake_dataset(output_path, seq_length, num_masked_tokens,
     f = tf.train.Feature(float_list=tf.train.FloatList(value=list(values)))
     return f
 
+  rng = np.random.default_rng(37)
   for _ in range(num_examples):
     features = {}
     padding = np.zeros(shape=(max_seq_length - seq_length), dtype=np.int32)
-    input_ids = np.random.randint(low=1, high=100, size=(seq_length))
+    input_ids = rng.integers(low=1, high=100, size=(seq_length))
     features['input_ids'] = create_int_feature(
         np.concatenate((input_ids, padding)))
     features['input_mask'] = create_int_feature(
@@ -56,9 +57,9 @@ def _create_fake_dataset(output_path, seq_length, num_masked_tokens,
     features['position_ids'] = create_int_feature(
         np.concatenate((np.ones_like(input_ids), padding)))
     features['masked_lm_positions'] = create_int_feature(
-        np.random.randint(60, size=(num_masked_tokens), dtype=np.int64))
+        rng.integers(60, size=(num_masked_tokens), dtype=np.int64))
     features['masked_lm_ids'] = create_int_feature(
-        np.random.randint(100, size=(num_masked_tokens), dtype=np.int64))
+        rng.integers(100, size=(num_masked_tokens), dtype=np.int64))
     features['masked_lm_weights'] = create_float_feature(
         np.ones((num_masked_tokens,), dtype=np.float32))
     features['next_sentence_labels'] = create_int_feature(np.array([0]))
@@ -156,6 +157,7 @@ class PretrainDynamicDataLoaderTest(tf.test.TestCase, parameterized.TestCase):
         self.assertEqual(dynamic_metrics[key], static_metrics[key])
 
   def test_load_dataset(self):
+    tf.random.set_seed(0)
     max_seq_length = 128
     batch_size = 2
     input_path_1 = os.path.join(self.get_temp_dir(), 'train_1.tf_record')
@@ -178,7 +180,8 @@ class PretrainDynamicDataLoaderTest(tf.test.TestCase, parameterized.TestCase):
         input_path=input_paths,
         seq_bucket_lengths=[64, 128],
         use_position_id=True,
-        global_batch_size=batch_size)
+        global_batch_size=batch_size,
+        deterministic=True)
     dataset = pretrain_dynamic_dataloader.PretrainingDynamicDataLoader(
         data_config).load()
     dataset_it = iter(dataset)

official/nlp/data/squad_lib_sp.py

@@ -175,7 +175,7 @@ def _convert_index(index, pos, m=None, is_start=True):
     front -= 1
   assert index[front] is not None or index[rear] is not None
   if index[front] is None:
-    if index[rear] >= 1:
+    if index[rear] >= 1:  # pytype: disable=unsupported-operands
       if is_start:
         return 0
       else:

official/nlp/keras_nlp/layers/position_embedding.py

@@ -66,14 +66,8 @@ class PositionEmbedding(tf.keras.layers.Layer):
 
   def build(self, input_shape):
     dimension_list = input_shape.as_list()
-
-    seq_length = dimension_list[self._seq_axis]
     width = dimension_list[-1]
-
-    if self._max_length is not None:
-      weight_sequence_length = self._max_length
-    else:
-      weight_sequence_length = seq_length
+    weight_sequence_length = self._max_length
 
     self._position_embeddings = self.add_weight(
         "embeddings",

official/nlp/modeling/layers/cls_head.py

@@ -59,19 +59,33 @@ class ClassificationHead(tf.keras.layers.Layer):
           activation=self.activation,
           kernel_initializer=self.initializer,
           name="pooler_dense")
-      self.dropout = tf.keras.layers.Dropout(rate=self.dropout_rate)
+    self.dropout = tf.keras.layers.Dropout(rate=self.dropout_rate)
 
     self.out_proj = tf.keras.layers.Dense(
         units=num_classes, kernel_initializer=self.initializer, name="logits")
 
-  def call(self, features):
+  def call(self, features: tf.Tensor, only_project: bool = False):
+    """Implements call().
+
+    Args:
+      features: a rank-3 Tensor when self.inner_dim is specified, otherwise
+        it is a rank-2 Tensor.
+      only_project: a boolean. If True, we return the intermediate Tensor
+        before projecting to class logits.
+
+    Returns:
+      a Tensor, if only_project is True, shape= [batch size, hidden size].
+      If only_project is False, shape= [batch size, num classes].
+    """
     if not self.inner_dim:
       x = features
     else:
       x = features[:, self.cls_token_idx, :]  # take <CLS> token.
       x = self.dense(x)
-      x = self.dropout(x)
 
+    if only_project:
+      return x
+    x = self.dropout(x)
     x = self.out_proj(x)
     return x
 
@@ -134,7 +148,7 @@ class MultiClsHeads(tf.keras.layers.Layer):
           activation=self.activation,
           kernel_initializer=self.initializer,
           name="pooler_dense")
-      self.dropout = tf.keras.layers.Dropout(rate=self.dropout_rate)
+    self.dropout = tf.keras.layers.Dropout(rate=self.dropout_rate)
     self.out_projs = []
     for name, num_classes in cls_list:
       self.out_projs.append(
@@ -142,13 +156,28 @@ class MultiClsHeads(tf.keras.layers.Layer):
               units=num_classes, kernel_initializer=self.initializer,
               name=name))
 
-  def call(self, features):
+  def call(self, features: tf.Tensor, only_project: bool = False):
+    """Implements call().
+
+    Args:
+      features: a rank-3 Tensor when self.inner_dim is specified, otherwise
+        it is a rank-2 Tensor.
+      only_project: a boolean. If True, we return the intermediate Tensor
+        before projecting to class logits.
+
+    Returns:
+      If only_project is True, a Tensor with shape= [batch size, hidden size].
+      If only_project is False, a dictionary of Tensors.
+    """
     if not self.inner_dim:
       x = features
     else:
       x = features[:, self.cls_token_idx, :]  # take <CLS> token.
       x = self.dense(x)
-      x = self.dropout(x)
+
+    if only_project:
+      return x
+    x = self.dropout(x)
 
     outputs = {}
     for proj_layer in self.out_projs:

official/nlp/modeling/layers/cls_head_test.py

@@ -39,6 +39,8 @@ class ClassificationHeadTest(tf.test.TestCase, parameterized.TestCase):
     self.assertAllClose(output, [[0., 0.], [0., 0.]])
     self.assertSameElements(test_layer.checkpoint_items.keys(),
                             ["pooler_dense"])
+    outputs = test_layer(features, only_project=True)
+    self.assertEqual(outputs.shape, (2, 5))
 
   def test_layer_serialization(self):
     layer = cls_head.ClassificationHead(10, 2)
@@ -71,6 +73,9 @@ class MultiClsHeadsTest(tf.test.TestCase, parameterized.TestCase):
     self.assertSameElements(test_layer.checkpoint_items.keys(),
                             ["pooler_dense", "foo", "bar"])
 
+    outputs = test_layer(features, only_project=True)
+    self.assertEqual(outputs.shape, (2, 5))
+
   def test_layer_serialization(self):
     cls_list = [("foo", 2), ("bar", 3)]
     test_layer = cls_head.MultiClsHeads(inner_dim=5, cls_list=cls_list)

official/nlp/modeling/layers/mobile_bert_layers.py

@@ -39,23 +39,6 @@ class NoNorm(tf.keras.layers.Layer):
     return output
 
 
-@tf.keras.utils.register_keras_serializable(package='Text')
-class NoNormClipped(NoNorm):
-  """Quantization friendly implementation for the NoNorm.
-
-  The output of NoNorm layer is clipped to [-6.0, 6.0] to make it quantization
-  friendly.
-  """
-
-  def __init__(self, name=None):
-    super(NoNormClipped, self).__init__(name=name)
-
-  def call(self, feature):
-    output = feature * self.scale + self.bias
-    clipped_output = tf.clip_by_value(output, -6.0, 6.0)
-    return clipped_output
-
-
 def _get_norm_layer(normalization_type='no_norm', name=None):
   """Get normlization layer.
 
@@ -69,8 +52,6 @@ def _get_norm_layer(normalization_type='no_norm', name=None):
   """
   if normalization_type == 'no_norm':
     layer = NoNorm(name=name)
-  elif normalization_type == 'no_norm_clipped':
-    layer = NoNormClipped(name=name)
   elif normalization_type == 'layer_norm':
     layer = tf.keras.layers.LayerNormalization(
         name=name,

official/nlp/modeling/layers/mobile_bert_layers_test.py

@@ -33,22 +33,6 @@ def generate_fake_input(batch_size=1, seq_len=5, vocab_size=10000, seed=0):
   return fake_input
 
 
-class EdgeTPUNoNormTest(tf.test.TestCase):
-
-  def test_no_norm(self):
-    layer = mobile_bert_layers.NoNormClipped()
-    feature = tf.random.uniform(
-        [2, 3, 4], minval=-8, maxval=8, dtype=tf.float32)
-    output = layer(feature)
-    output_shape = output.shape.as_list()
-    expected_shape = [2, 3, 4]
-    self.assertListEqual(output_shape, expected_shape, msg=None)
-    output_min = tf.reduce_min(output)
-    output_max = tf.reduce_max(output)
-    self.assertGreaterEqual(6.0, output_max)
-    self.assertLessEqual(-6.0, output_min)
-
-
 class MobileBertEncoderTest(parameterized.TestCase, tf.test.TestCase):
 
   def test_embedding_layer_with_token_type(self):

official/nlp/modeling/models/seq2seq_transformer.py

@@ -544,7 +544,8 @@ class TransformerDecoder(tf.keras.layers.Layer):
            self_attention_mask=None,
            cross_attention_mask=None,
            cache=None,
-           decode_loop_step=None):
+           decode_loop_step=None,
+           return_all_decoder_outputs=False):
     """Return the output of the decoder layer stacks.
 
     Args:
@@ -561,6 +562,9 @@ class TransformerDecoder(tf.keras.layers.Layer):
                      ...}
       decode_loop_step: An integer, the step number of the decoding loop. Used
         only for autoregressive inference on TPU.
+      return_all_decoder_outputs: Return all decoder layer outputs.
+        Note that the outputs are layer normed.
+        This is useful when introducing per layer auxiliary loss.
 
     Returns:
       Output of decoder.
@@ -568,6 +572,7 @@ class TransformerDecoder(tf.keras.layers.Layer):
     """
 
     output_tensor = target
+    decoder_outputs = []
     for layer_idx in range(self.num_layers):
       transformer_inputs = [
           output_tensor, memory, cross_attention_mask, self_attention_mask
@@ -581,7 +586,13 @@ class TransformerDecoder(tf.keras.layers.Layer):
             transformer_inputs,
             cache=cache[cache_layer_idx],
             decode_loop_step=decode_loop_step)
-    return self.output_normalization(output_tensor)
+      if return_all_decoder_outputs:
+        decoder_outputs.append(self.output_normalization(output_tensor))
+
+    if return_all_decoder_outputs:
+      return decoder_outputs
+    else:
+      return self.output_normalization(output_tensor)
 
 
 def attention_initializer(hidden_size):

official/nlp/modeling/networks/classification.py

@@ -16,6 +16,7 @@
 # pylint: disable=g-classes-have-attributes
 import collections
 import tensorflow as tf
+from tensorflow.python.util import deprecation
 
 
 @tf.keras.utils.register_keras_serializable(package='Text')
@@ -39,6 +40,8 @@ class Classification(tf.keras.Model):
       `predictions`.
   """
 
+  @deprecation.deprecated(None, 'Classification as a network is deprecated. '
+                          'Please use the layers.ClassificationHead instead.')
   def __init__(self,
                input_width,
                num_classes,

official/nlp/modeling/networks/encoder_scaffold.py

@@ -74,9 +74,12 @@ class EncoderScaffold(tf.keras.Model):
       standard pretraining.
     num_hidden_instances: The number of times to instantiate and/or invoke the
       hidden_cls.
-    hidden_cls: The class or instance to encode the input data. If `hidden_cls`
-      is not set, a KerasBERT transformer layer will be used as the encoder
-      class.
+    hidden_cls: Three types of input are supported: (1) class (2) instance
+      (3) list of classes or instances, to encode the input data. If
+      `hidden_cls` is not set, a KerasBERT transformer layer will be used as the
+      encoder class. If `hidden_cls` is a list of classes or instances, these
+      classes (instances) are sequentially instantiated (invoked) on top of
+      embedding layer. Mixing classes and instances in the list is allowed.
     hidden_cfg: A dict of kwargs to pass to the hidden_cls, if it needs to be
       instantiated. If hidden_cls is not set, a config dict must be passed to
       `hidden_cfg` with the following values:
@@ -192,15 +195,26 @@ class EncoderScaffold(tf.keras.Model):
     layer_output_data = []
     hidden_layers = []
     hidden_cfg = hidden_cfg if hidden_cfg else {}
+
+    if isinstance(hidden_cls, list) and len(hidden_cls) != num_hidden_instances:
+      raise RuntimeError(
+          ('When input hidden_cls to EncoderScaffold %s is a list, it must '
+           'contain classes or instances with size specified by '
+           'num_hidden_instances, got %d vs %d.') % self.name, len(hidden_cls),
+          num_hidden_instances)
     for i in range(num_hidden_instances):
-      if inspect.isclass(hidden_cls):
+      if isinstance(hidden_cls, list):
+        cur_hidden_cls = hidden_cls[i]
+      else:
+        cur_hidden_cls = hidden_cls
+      if inspect.isclass(cur_hidden_cls):
         if hidden_cfg and 'attention_cfg' in hidden_cfg and (
             layer_idx_as_attention_seed):
           hidden_cfg = copy.deepcopy(hidden_cfg)
           hidden_cfg['attention_cfg']['seed'] = i
-        layer = hidden_cls(**hidden_cfg)
+        layer = cur_hidden_cls(**hidden_cfg)
       else:
-        layer = hidden_cls
+        layer = cur_hidden_cls
       data = layer([data, attention_mask])
       layer_output_data.append(data)
       hidden_layers.append(layer)
@@ -347,6 +361,15 @@ class EncoderScaffold(tf.keras.Model):
     else:
       return self._embedding_data
 
+  @property
+  def embedding_network(self):
+    if self._embedding_network is None:
+      raise RuntimeError(
+          ('The EncoderScaffold %s does not have a reference '
+           'to the embedding network. This is required when you '
+           'pass a custom embedding network to the scaffold.') % self.name)
+    return self._embedding_network
+
   @property
   def hidden_layers(self):
     """List of hidden layers in the encoder."""

official/nlp/modeling/networks/encoder_scaffold_test.py

@@ -605,6 +605,98 @@ class EncoderScaffoldHiddenInstanceTest(keras_parameterized.TestCase):
     self.assertNotEmpty(call_list)
     self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
 
+  def test_hidden_cls_list(self):
+    hidden_size = 32
+    sequence_length = 10
+    vocab_size = 57
+
+    embedding_network = Embeddings(vocab_size, hidden_size)
+
+    call_list = []
+    hidden_cfg = {
+        "num_attention_heads":
+            2,
+        "intermediate_size":
+            3072,
+        "intermediate_activation":
+            activations.gelu,
+        "dropout_rate":
+            0.1,
+        "attention_dropout_rate":
+            0.1,
+        "kernel_initializer":
+            tf.keras.initializers.TruncatedNormal(stddev=0.02),
+        "call_list":
+            call_list
+    }
+    mask_call_list = []
+    mask_cfg = {
+        "call_list": mask_call_list
+    }
+    # Create a small EncoderScaffold for testing. This time, we pass an already-
+    # instantiated layer object.
+    xformer = ValidatedTransformerLayer(**hidden_cfg)
+    xmask = ValidatedMaskLayer(**mask_cfg)
+
+    test_network_a = encoder_scaffold.EncoderScaffold(
+        num_hidden_instances=3,
+        pooled_output_dim=hidden_size,
+        pooler_layer_initializer=tf.keras.initializers.TruncatedNormal(
+            stddev=0.02),
+        hidden_cls=xformer,
+        mask_cls=xmask,
+        embedding_cls=embedding_network)
+    # Create a network b with same embedding and hidden layers as network a.
+    test_network_b = encoder_scaffold.EncoderScaffold(
+        num_hidden_instances=3,
+        pooled_output_dim=hidden_size,
+        pooler_layer_initializer=tf.keras.initializers.TruncatedNormal(
+            stddev=0.02),
+        mask_cls=xmask,
+        embedding_cls=test_network_a.embedding_network,
+        hidden_cls=test_network_a.hidden_layers)
+    # Create a network c with same embedding but fewer hidden layers compared to
+    # network a and b.
+    hidden_layers = test_network_a.hidden_layers
+    hidden_layers.pop()
+    test_network_c = encoder_scaffold.EncoderScaffold(
+        num_hidden_instances=2,
+        pooled_output_dim=hidden_size,
+        pooler_layer_initializer=tf.keras.initializers.TruncatedNormal(
+            stddev=0.02),
+        mask_cls=xmask,
+        embedding_cls=test_network_a.embedding_network,
+        hidden_cls=hidden_layers)
+
+    # 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)
+
+    # Create model based off of network a:
+    data_a, pooled_a = test_network_a([word_ids, mask])
+    model_a = tf.keras.Model([word_ids, mask], [data_a, pooled_a])
+    # Create model based off of network b:
+    data_b, pooled_b = test_network_b([word_ids, mask])
+    model_b = tf.keras.Model([word_ids, mask], [data_b, pooled_b])
+    # Create model based off of network b:
+    data_c, pooled_c = test_network_c([word_ids, mask])
+    model_c = tf.keras.Model([word_ids, mask], [data_c, pooled_c])
+
+    batch_size = 3
+    word_id_data = np.random.randint(
+        vocab_size, size=(batch_size, sequence_length))
+    mask_data = np.random.randint(2, size=(batch_size, sequence_length))
+    output_a, _ = model_a.predict([word_id_data, mask_data])
+    output_b, _ = model_b.predict([word_id_data, mask_data])
+    output_c, _ = model_c.predict([word_id_data, mask_data])
+
+    # Outputs from model a and b should be the same since they share the same
+    # embedding and hidden layers.
+    self.assertAllEqual(output_a, output_b)
+    # Outputs from model a and c shouldn't be the same since they share the same
+    # embedding layer but different number of hidden layers.
+    self.assertNotAllEqual(output_a, output_c)
+
   @parameterized.parameters(True, False)
   def test_serialize_deserialize(self, use_hidden_cls_instance):
     hidden_size = 32

official/nlp/modeling/networks/funnel_transformer.py

+# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Funnel Transformer network."""
+# pylint: disable=g-classes-have-attributes
+from typing import Union, Collection
+from absl import logging
+import tensorflow as tf
+
+from official.nlp import keras_nlp
+
+
+def _pool_and_concat(data, unpool_length: int, stride: int,
+                     axes: Union[Collection[int], int]):
+  """Pools the data along a given axis with stride.
+
+  It also skips first unpool_length elements.
+
+  Args:
+    data: Tensor to be pooled.
+    unpool_length: Leading elements to be skipped.
+    stride: Stride for the given axis.
+    axes: Axes to pool the Tensor.
+
+  Returns:
+    Pooled and concatenated Tensor.
+  """
+  # Wraps the axes as a list.
+  if isinstance(axes, int):
+    axes = [axes]
+
+  for axis in axes:
+    # Skips first `unpool_length` tokens.
+    unpool_tensor_shape = [slice(None)] * axis + [slice(None, unpool_length)]
+    unpool_tensor = data[unpool_tensor_shape]
+    # Pools the second half.
+    pool_tensor_shape = [slice(None)] * axis + [
+        slice(unpool_length, None, stride)
+    ]
+    pool_tensor = data[pool_tensor_shape]
+    data = tf.concat((unpool_tensor, pool_tensor), axis=axis)
+  return data
+
+
+@tf.keras.utils.register_keras_serializable(package='Text')
+class FunnelTransformerEncoder(tf.keras.layers.Layer):
+  """Funnel Transformer-based encoder network.
+
+  Funnel Transformer Implementation of https://arxiv.org/abs/2006.03236.
+  This implementation utilizes the base framework with Bert
+  (https://arxiv.org/abs/1810.04805).
+  Its output is compatible with `BertEncoder`.
+
+  Args:
+    vocab_size: The size of the token vocabulary.
+    hidden_size: The size of the transformer hidden layers.
+    num_layers: The number of transformer layers.
+    num_attention_heads: The number of attention heads for each transformer. The
+      hidden size must be divisible by the number of attention heads.
+    max_sequence_length: The maximum sequence length that this encoder can
+      consume. If None, max_sequence_length uses the value from sequence length.
+      This determines the variable shape for positional embeddings.
+    type_vocab_size: The number of types that the 'type_ids' input can take.
+    inner_dim: The output dimension of the first Dense layer in a two-layer
+      feedforward network for each transformer.
+    inner_activation: The activation for the first Dense layer in a two-layer
+      feedforward network for each transformer.
+    output_dropout: Dropout probability for the post-attention and output
+      dropout.
+    attention_dropout: The dropout rate to use for the attention layers within
+      the transformer layers.
+    pool_stride: Pooling stride to compress the sequence length.
+    unpool_length: Leading n tokens to be skipped from pooling.
+    initializer: The initialzer to use for all weights in this encoder.
+    output_range: The sequence output range, [0, output_range), by slicing the
+      target sequence of the last transformer layer. `None` means the entire
+      target sequence will attend to the source sequence, which yields the full
+      output.
+    embedding_width: The width of the word embeddings. If the embedding width is
+      not equal to hidden size, embedding parameters will be factorized into two
+      matrices in the shape of ['vocab_size', 'embedding_width'] and
+      ['embedding_width', 'hidden_size'] ('embedding_width' is usually much
+      smaller than 'hidden_size').
+    embedding_layer: An optional Layer instance which will be called to generate
+      embeddings for the input word IDs.
+    norm_first: Whether to normalize inputs to attention and intermediate dense
+      layers. If set False, output of attention and intermediate dense layers is
+      normalized.
+  """
+
+  def __init__(
+      self,
+      vocab_size,
+      hidden_size=768,
+      num_layers=12,
+      num_attention_heads=12,
+      max_sequence_length=512,
+      type_vocab_size=16,
+      inner_dim=3072,
+      inner_activation=lambda x: tf.keras.activations.gelu(x, approximate=True),
+      output_dropout=0.1,
+      attention_dropout=0.1,
+      pool_stride=2,
+      unpool_length=0,
+      initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02),
+      output_range=None,
+      embedding_width=None,
+      embedding_layer=None,
+      norm_first=False,
+      **kwargs):
+    super().__init__(**kwargs)
+    activation = tf.keras.activations.get(inner_activation)
+    initializer = tf.keras.initializers.get(initializer)
+
+    if embedding_width is None:
+      embedding_width = hidden_size
+
+    if embedding_layer is None:
+      self._embedding_layer = keras_nlp.layers.OnDeviceEmbedding(
+          vocab_size=vocab_size,
+          embedding_width=embedding_width,
+          initializer=initializer,
+          name='word_embeddings')
+    else:
+      self._embedding_layer = embedding_layer
+
+    self._position_embedding_layer = keras_nlp.layers.PositionEmbedding(
+        initializer=initializer,
+        max_length=max_sequence_length,
+        name='position_embedding')
+
+    self._type_embedding_layer = keras_nlp.layers.OnDeviceEmbedding(
+        vocab_size=type_vocab_size,
+        embedding_width=embedding_width,
+        initializer=initializer,
+        use_one_hot=True,
+        name='type_embeddings')
+
+    self._embedding_norm_layer = tf.keras.layers.LayerNormalization(
+        name='embeddings/layer_norm', axis=-1, epsilon=1e-12, dtype=tf.float32)
+
+    self._embedding_dropout = tf.keras.layers.Dropout(
+        rate=output_dropout, name='embedding_dropout')
+
+    # We project the 'embedding' output to 'hidden_size' if it is not already
+    # 'hidden_size'.
+    self._embedding_projection = None
+    if embedding_width != hidden_size:
+      self._embedding_projection = tf.keras.layers.experimental.EinsumDense(
+          '...x,xy->...y',
+          output_shape=hidden_size,
+          bias_axes='y',
+          kernel_initializer=initializer,
+          name='embedding_projection')
+
+    self._transformer_layers = []
+    self._attention_mask_layer = keras_nlp.layers.SelfAttentionMask(
+        name='self_attention_mask')
+    for i in range(num_layers):
+      layer = keras_nlp.layers.TransformerEncoderBlock(
+          num_attention_heads=num_attention_heads,
+          inner_dim=inner_dim,
+          inner_activation=inner_activation,
+          output_dropout=output_dropout,
+          attention_dropout=attention_dropout,
+          norm_first=norm_first,
+          output_range=output_range if i == num_layers - 1 else None,
+          kernel_initializer=initializer,
+          name='transformer/layer_%d' % i)
+      self._transformer_layers.append(layer)
+
+    self._pooler_layer = tf.keras.layers.Dense(
+        units=hidden_size,
+        activation='tanh',
+        kernel_initializer=initializer,
+        name='pooler_transform')
+    self._att_input_pool_layer = tf.keras.layers.MaxPooling1D(
+        pool_size=pool_stride,
+        strides=pool_stride,
+        padding='same',
+        name='att_input_pool_layer')
+    self._pool_stride = pool_stride
+    self._unpool_length = unpool_length
+
+    self._config = {
+        'vocab_size': vocab_size,
+        'hidden_size': hidden_size,
+        'num_layers': num_layers,
+        'num_attention_heads': num_attention_heads,
+        'max_sequence_length': max_sequence_length,
+        'type_vocab_size': type_vocab_size,
+        'inner_dim': inner_dim,
+        'inner_activation': tf.keras.activations.serialize(activation),
+        'output_dropout': output_dropout,
+        'attention_dropout': attention_dropout,
+        'initializer': tf.keras.initializers.serialize(initializer),
+        'output_range': output_range,
+        'embedding_width': embedding_width,
+        'embedding_layer': embedding_layer,
+        'norm_first': norm_first,
+        'pool_stride': pool_stride,
+        'unpool_length': unpool_length,
+    }
+
+  def call(self, inputs):
+    # inputs are [word_ids, mask, type_ids]
+    if isinstance(inputs, (list, tuple)):
+      logging.warning('List inputs to  %s are discouraged.', self.__class__)
+      if len(inputs) == 3:
+        word_ids, mask, type_ids = inputs
+      else:
+        raise ValueError('Unexpected inputs to %s with length at %d.' %
+                         (self.__class__, len(inputs)))
+    elif isinstance(inputs, dict):
+      word_ids = inputs.get('input_word_ids')
+      mask = inputs.get('input_mask')
+      type_ids = inputs.get('input_type_ids')
+    else:
+      raise ValueError('Unexpected inputs type to %s.' % self.__class__)
+
+    word_embeddings = self._embedding_layer(word_ids)
+    # absolute position embeddings
+    position_embeddings = self._position_embedding_layer(word_embeddings)
+    type_embeddings = self._type_embedding_layer(type_ids)
+
+    embeddings = tf.keras.layers.add(
+        [word_embeddings, position_embeddings, type_embeddings])
+    embeddings = self._embedding_norm_layer(embeddings)
+    embeddings = self._embedding_dropout(embeddings)
+
+    if self._embedding_projection is not None:
+      embeddings = self._embedding_projection(embeddings)
+
+    attention_mask = self._attention_mask_layer(embeddings, mask)
+
+    encoder_outputs = []
+    x = embeddings
+    # TODO(b/195972228): attention_mask can be co-generated with pooling.
+    attention_mask = _pool_and_concat(
+        attention_mask,
+        unpool_length=self._unpool_length,
+        stride=self._pool_stride,
+        axes=[1])
+    for layer in self._transformer_layers:
+      # Pools layer for compressing the query length.
+      pooled_inputs = self._att_input_pool_layer(x[:, self._unpool_length:, :])
+      query_inputs = tf.concat(
+          values=(tf.cast(
+              x[:, :self._unpool_length, :],
+              dtype=pooled_inputs.dtype), pooled_inputs),
+          axis=1)
+      x = layer([query_inputs, x, attention_mask])
+      # Pools the corresponding attention_mask.
+      attention_mask = _pool_and_concat(
+          attention_mask,
+          unpool_length=self._unpool_length,
+          stride=self._pool_stride,
+          axes=[1, 2])
+      encoder_outputs.append(x)
+
+    last_encoder_output = encoder_outputs[-1]
+    first_token_tensor = last_encoder_output[:, 0, :]
+    pooled_output = self._pooler_layer(first_token_tensor)
+
+    return dict(
+        sequence_output=encoder_outputs[-1],
+        pooled_output=pooled_output,
+        encoder_outputs=encoder_outputs)
+
+  def get_embedding_table(self):
+    return self._embedding_layer.embeddings
+
+  def get_embedding_layer(self):
+    return self._embedding_layer
+
+  def get_config(self):
+    return dict(self._config)
+
+  @property
+  def transformer_layers(self):
+    """List of Transformer layers in the encoder."""
+    return self._transformer_layers
+
+  @property
+  def pooler_layer(self):
+    """The pooler dense layer after the transformer layers."""
+    return self._pooler_layer
+
+  @classmethod
+  def from_config(cls, config, custom_objects=None):
+    if 'embedding_layer' in config and config['embedding_layer'] is not None:
+      warn_string = (
+          'You are reloading a model that was saved with a '
+          'potentially-shared embedding layer object. If you contine to '
+          'train this model, the embedding layer will no longer be shared. '
+          'To work around this, load the model outside of the Keras API.')
+      print('WARNING: ' + warn_string)
+      logging.warn(warn_string)
+
+    return cls(**config)

official/nlp/modeling/networks/funnel_transformer_test.py

+# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for transformer-based bert encoder network."""
+
+from absl.testing import parameterized
+import numpy as np
+import tensorflow as tf
+
+from official.nlp.modeling.networks import funnel_transformer
+
+
+class SingleLayerModel(tf.keras.Model):
+
+  def __init__(self, layer):
+    super().__init__()
+    self.layer = layer
+
+  def call(self, inputs):
+    return self.layer(inputs)
+
+
+class FunnelTransformerEncoderTest(parameterized.TestCase, tf.test.TestCase):
+
+  def tearDown(self):
+    super(FunnelTransformerEncoderTest, self).tearDown()
+    tf.keras.mixed_precision.set_global_policy("float32")
+
+  @parameterized.named_parameters(("mix", "mixed_float16", tf.float16),
+                                  ("float32", "float32", tf.float32))
+  def test_network_creation(self, policy, pooled_dtype):
+    tf.keras.mixed_precision.set_global_policy(policy)
+
+    hidden_size = 32
+    sequence_length = 21
+    pool_stride = 2
+    num_layers = 3
+    # Create a small FunnelTransformerEncoder for testing.
+    test_network = funnel_transformer.FunnelTransformerEncoder(
+        vocab_size=100,
+        hidden_size=hidden_size,
+        num_attention_heads=2,
+        num_layers=num_layers,
+        pool_stride=pool_stride,
+        unpool_length=0)
+    # 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)
+    type_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
+    dict_outputs = test_network([word_ids, mask, type_ids])
+    data = dict_outputs["sequence_output"]
+    pooled = dict_outputs["pooled_output"]
+
+    self.assertIsInstance(test_network.transformer_layers, list)
+    self.assertLen(test_network.transformer_layers, num_layers)
+    self.assertIsInstance(test_network.pooler_layer, tf.keras.layers.Dense)
+
+    # Stride=2 compresses sequence length to half the size at each layer.
+    # This configuration gives each layer of seq length: 21->11->6->3.
+    expected_data_shape = [None, 3, hidden_size]
+    expected_pooled_shape = [None, hidden_size]
+
+    self.assertAllEqual(expected_data_shape, data.shape.as_list())
+    self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
+
+    # The default output dtype is float32.
+    # 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(pooled_dtype, pooled.dtype)
+
+  @parameterized.named_parameters(
+      ("no_stride_no_unpool", 1, 0),
+      ("large_stride_with_unpool", 3, 1),
+      ("large_stride_with_large_unpool", 5, 10),
+      ("no_stride_with_unpool", 1, 1),
+  )
+  def test_all_encoder_outputs_network_creation(self, pool_stride,
+                                                unpool_length):
+    hidden_size = 32
+    sequence_length = 21
+    num_layers = 3
+    # Create a small FunnelTransformerEncoder for testing.
+    test_network = funnel_transformer.FunnelTransformerEncoder(
+        vocab_size=100,
+        hidden_size=hidden_size,
+        num_attention_heads=2,
+        num_layers=num_layers,
+        pool_stride=pool_stride,
+        unpool_length=unpool_length)
+    # 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)
+    type_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
+    dict_outputs = test_network([word_ids, mask, type_ids])
+    all_encoder_outputs = dict_outputs["encoder_outputs"]
+    pooled = dict_outputs["pooled_output"]
+
+    expected_data_shape = [None, sequence_length, hidden_size]
+    expected_pooled_shape = [None, hidden_size]
+    self.assertLen(all_encoder_outputs, num_layers)
+    for data in all_encoder_outputs:
+      expected_data_shape[1] = unpool_length + (expected_data_shape[1] +
+                                                pool_stride - 1 -
+                                                unpool_length) // pool_stride
+      print("shapes:", expected_data_shape, data.shape.as_list())
+      self.assertAllEqual(expected_data_shape, data.shape.as_list())
+    self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
+
+    # The default output dtype is float32.
+    self.assertAllEqual(tf.float32, all_encoder_outputs[-1].dtype)
+    self.assertAllEqual(tf.float32, pooled.dtype)
+
+  @parameterized.named_parameters(
+      ("all_sequence", None, 3, 0),
+      ("output_range", 1, 1, 0),
+      ("all_sequence_wit_unpool", None, 4, 1),
+      ("output_range_with_unpool", 1, 1, 1),
+      ("output_range_with_large_unpool", 1, 1, 2),
+  )
+  def test_network_invocation(self, output_range, out_seq_len, unpool_length):
+    hidden_size = 32
+    sequence_length = 21
+    vocab_size = 57
+    num_types = 7
+    pool_stride = 2
+    # Create a small FunnelTransformerEncoder for testing.
+    test_network = funnel_transformer.FunnelTransformerEncoder(
+        vocab_size=vocab_size,
+        hidden_size=hidden_size,
+        num_attention_heads=2,
+        num_layers=3,
+        type_vocab_size=num_types,
+        output_range=output_range,
+        pool_stride=pool_stride,
+        unpool_length=unpool_length)
+    # 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)
+    type_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
+    dict_outputs = test_network([word_ids, mask, type_ids])
+    data = dict_outputs["sequence_output"]
+    pooled = dict_outputs["pooled_output"]
+
+    # Create a model based off of this network:
+    model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
+
+    # Invoke the model. We can't validate the output data here (the model is too
+    # complex) but this will catch structural runtime errors.
+    batch_size = 3
+    word_id_data = np.random.randint(
+        vocab_size, size=(batch_size, sequence_length))
+    mask_data = np.random.randint(2, size=(batch_size, sequence_length))
+    type_id_data = np.random.randint(
+        num_types, size=(batch_size, sequence_length))
+    outputs = model.predict([word_id_data, mask_data, type_id_data])
+    self.assertEqual(outputs[0].shape[1], out_seq_len)  # output_range
+
+    # Creates a FunnelTransformerEncoder with max_sequence_length !=
+    # sequence_length
+    max_sequence_length = 128
+    test_network = funnel_transformer.FunnelTransformerEncoder(
+        vocab_size=vocab_size,
+        hidden_size=hidden_size,
+        max_sequence_length=max_sequence_length,
+        num_attention_heads=2,
+        num_layers=3,
+        type_vocab_size=num_types,
+        pool_stride=pool_stride)
+    dict_outputs = test_network([word_ids, mask, type_ids])
+    data = dict_outputs["sequence_output"]
+    pooled = dict_outputs["pooled_output"]
+    model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
+    outputs = model.predict([word_id_data, mask_data, type_id_data])
+    self.assertEqual(outputs[0].shape[1], 3)
+
+    # Creates a FunnelTransformerEncoder with embedding_width != hidden_size
+    test_network = funnel_transformer.FunnelTransformerEncoder(
+        vocab_size=vocab_size,
+        hidden_size=hidden_size,
+        max_sequence_length=max_sequence_length,
+        num_attention_heads=2,
+        num_layers=3,
+        type_vocab_size=num_types,
+        embedding_width=16,
+        pool_stride=pool_stride)
+    dict_outputs = test_network([word_ids, mask, type_ids])
+    data = dict_outputs["sequence_output"]
+    pooled = dict_outputs["pooled_output"]
+    model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
+    outputs = model.predict([word_id_data, mask_data, type_id_data])
+    self.assertEqual(outputs[0].shape[-1], hidden_size)
+    self.assertTrue(hasattr(test_network, "_embedding_projection"))
+
+  def test_serialize_deserialize(self):
+    # Create a network object that sets all of its config options.
+    kwargs = dict(
+        vocab_size=100,
+        hidden_size=32,
+        num_layers=3,
+        num_attention_heads=2,
+        max_sequence_length=21,
+        type_vocab_size=12,
+        inner_dim=1223,
+        inner_activation="relu",
+        output_dropout=0.05,
+        attention_dropout=0.22,
+        initializer="glorot_uniform",
+        output_range=-1,
+        embedding_width=16,
+        embedding_layer=None,
+        norm_first=False,
+        pool_stride=2,
+        unpool_length=0)
+    network = funnel_transformer.FunnelTransformerEncoder(**kwargs)
+    expected_config = dict(kwargs)
+    expected_config["inner_activation"] = tf.keras.activations.serialize(
+        tf.keras.activations.get(expected_config["inner_activation"]))
+    expected_config["initializer"] = tf.keras.initializers.serialize(
+        tf.keras.initializers.get(expected_config["initializer"]))
+    self.assertEqual(network.get_config(), expected_config)
+    # Create another network object from the first object's config.
+    new_network = funnel_transformer.FunnelTransformerEncoder.from_config(
+        network.get_config())
+
+    # If the serialization was successful, the new config should match the old.
+    self.assertAllEqual(network.get_config(), new_network.get_config())
+
+    # Tests model saving/loading.
+    model_path = self.get_temp_dir() + "/model"
+    network_wrapper = SingleLayerModel(network)
+    # One forward-path to ensure input_shape.
+    batch_size = 3
+    sequence_length = 21
+    vocab_size = 100
+    num_types = 12
+    word_id_data = np.random.randint(
+        vocab_size, size=(batch_size, sequence_length))
+    mask_data = np.random.randint(2, size=(batch_size, sequence_length))
+    type_id_data = np.random.randint(
+        num_types, size=(batch_size, sequence_length))
+
+    _ = network_wrapper.predict([word_id_data, mask_data, type_id_data])
+    network_wrapper.save(model_path)
+    _ = tf.keras.models.load_model(model_path)
+
+
+if __name__ == "__main__":
+  tf.test.main()