transformer.py 21.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
# Copyright 2018 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.
# ==============================================================================
"""Defines the Transformer model in TF 2.0.

Model paper: https://arxiv.org/pdf/1706.03762.pdf
Transformer model code source: https://github.com/tensorflow/tensor2tensor
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf

from official.transformer.model import model_utils
from official.transformer.utils.tokenizer import EOS_ID
from official.transformer.v2 import attention_layer
29
from official.transformer.v2 import beam_search
30
31
32
33
34
from official.transformer.v2 import embedding_layer
from official.transformer.v2 import ffn_layer
from official.transformer.v2 import metrics


Reed's avatar
Reed committed
35
36
37
38
39
# Disable the not-callable lint error, since it claims many objects are not
# callable when they actually are.
# pylint: disable=not-callable


40
41
42
43
44
45
46
47
48
49
def create_model(params, is_train):
  """Creates transformer model."""
  with tf.name_scope("model"):
    if is_train:
      inputs = tf.keras.layers.Input((None,), dtype="int64", name="inputs")
      targets = tf.keras.layers.Input((None,), dtype="int64", name="targets")
      internal_model = Transformer(params, name="transformer_v2")
      logits = internal_model([inputs, targets], training=is_train)
      vocab_size = params["vocab_size"]
      label_smoothing = params["label_smoothing"]
50
51
      if params["enable_metrics_in_training"]:
        logits = metrics.MetricLayer(vocab_size)([logits, targets])
52
      logits = tf.keras.layers.Lambda(lambda x: x, name="logits",
Reed's avatar
Reed committed
53
                                      dtype=tf.float32)(logits)
guptapriya's avatar
guptapriya committed
54
      model = tf.keras.Model([inputs, targets], logits)
55
      # TODO(reedwm): Can we do this loss in float16 instead of float32?
guptapriya's avatar
guptapriya committed
56
57
58
59
      loss = metrics.transformer_loss(
          logits, targets, label_smoothing, vocab_size)
      model.add_loss(loss)
      return model
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89

    else:
      inputs = tf.keras.layers.Input((None,), dtype="int64", name="inputs")
      internal_model = Transformer(params, name="transformer_v2")
      ret = internal_model([inputs], training=is_train)
      outputs, scores = ret["outputs"], ret["scores"]
      return tf.keras.Model(inputs, [outputs, scores])


class Transformer(tf.keras.Model):
  """Transformer model with Keras.

  Implemented as described in: https://arxiv.org/pdf/1706.03762.pdf

  The Transformer model consists of an encoder and decoder. The input is an int
  sequence (or a batch of sequences). The encoder produces a continuous
  representation, and the decoder uses the encoder output to generate
  probabilities for the output sequence.
  """

  def __init__(self, params, name=None):
    """Initialize layers to build Transformer model.

    Args:
      params: hyperparameter object defining layer sizes, dropout values, etc.
      name: name of the model.
    """
    super(Transformer, self).__init__(name=name)
    self.params = params
    self.embedding_softmax_layer = embedding_layer.EmbeddingSharedWeights(
90
        params["vocab_size"], params["hidden_size"])
91
92
93
94
95
96
97
98
    self.encoder_stack = EncoderStack(params)
    self.decoder_stack = DecoderStack(params)

  def get_config(self):
    return {
        "params": self.params,
    }

99
  def call(self, inputs, training):
100
101
102
    """Calculate target logits or inferred target sequences.

    Args:
103
      inputs: input tensor list of size 1 or 2.
104
105
106
107
108
109
110
111
112
113
114
115
        First item, inputs: int tensor with shape [batch_size, input_length].
        Second item (optional), targets: None or int tensor with shape
          [batch_size, target_length].
      training: boolean, whether in training mode or not.

    Returns:
      If targets is defined, then return logits for each word in the target
      sequence. float tensor with shape [batch_size, target_length, vocab_size]
      If target is none, then generate output sequence one token at a time.
        returns a dictionary {
          outputs: [batch_size, decoded length]
          scores: [batch_size, float]}
116
      Even when float16 is used, the output tensor(s) are always float32.
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
117
118
119

    Raises:
      NotImplementedError: If try to use padded decode method on CPU/GPUs.
120
    """
121
122
    if len(inputs) == 2:
      inputs, targets = inputs[0], inputs[1]
123
    else:
124
      # Decoding path.
125
      inputs, targets = inputs[0], None
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
126
127
128
129
130
131
      if self.params["padded_decode"]:
        if not self.params["num_replicas"]:
          raise NotImplementedError(
              "Padded decoding on CPU/GPUs is not supported.")
        decode_batch_size = int(self.params["decode_batch_size"] /
                                self.params["num_replicas"])
132
133
134
        inputs.set_shape([
            decode_batch_size, self.params["decode_max_length"]
        ])
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168

    # Variance scaling is used here because it seems to work in many problems.
    # Other reasonable initializers may also work just as well.
    with tf.name_scope("Transformer"):
      # Calculate attention bias for encoder self-attention and decoder
      # multi-headed attention layers.
      attention_bias = model_utils.get_padding_bias(inputs)

      # Run the inputs through the encoder layer to map the symbol
      # representations to continuous representations.
      encoder_outputs = self.encode(inputs, attention_bias, training)
      # Generate output sequence if targets is None, or return logits if target
      # sequence is known.
      if targets is None:
        return self.predict(encoder_outputs, attention_bias, training)
      else:
        logits = self.decode(targets, encoder_outputs, attention_bias, training)
        return logits

  def encode(self, inputs, attention_bias, training):
    """Generate continuous representation for inputs.

    Args:
      inputs: int tensor with shape [batch_size, input_length].
      attention_bias: float tensor with shape [batch_size, 1, 1, input_length].
      training: boolean, whether in training mode or not.

    Returns:
      float tensor with shape [batch_size, input_length, hidden_size]
    """
    with tf.name_scope("encode"):
      # Prepare inputs to the layer stack by adding positional encodings and
      # applying dropout.
      embedded_inputs = self.embedding_softmax_layer(inputs)
169
      embedded_inputs = tf.cast(embedded_inputs, self.params["dtype"])
170
      inputs_padding = model_utils.get_padding(inputs)
171
      attention_bias = tf.cast(attention_bias, self.params["dtype"])
172
173
174
175
176

      with tf.name_scope("add_pos_encoding"):
        length = tf.shape(embedded_inputs)[1]
        pos_encoding = model_utils.get_position_encoding(
            length, self.params["hidden_size"])
177
        pos_encoding = tf.cast(pos_encoding, self.params["dtype"])
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
        encoder_inputs = embedded_inputs + pos_encoding

      if training:
        encoder_inputs = tf.nn.dropout(
            encoder_inputs, rate=self.params["layer_postprocess_dropout"])

      return self.encoder_stack(
          encoder_inputs, attention_bias, inputs_padding, training=training)

  def decode(self, targets, encoder_outputs, attention_bias, training):
    """Generate logits for each value in the target sequence.

    Args:
      targets: target values for the output sequence. int tensor with shape
        [batch_size, target_length]
      encoder_outputs: continuous representation of input sequence. float tensor
        with shape [batch_size, input_length, hidden_size]
      attention_bias: float tensor with shape [batch_size, 1, 1, input_length]
      training: boolean, whether in training mode or not.

    Returns:
      float32 tensor with shape [batch_size, target_length, vocab_size]
    """
    with tf.name_scope("decode"):
      # Prepare inputs to decoder layers by shifting targets, adding positional
      # encoding and applying dropout.
      decoder_inputs = self.embedding_softmax_layer(targets)
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
205
      decoder_inputs = tf.cast(decoder_inputs, self.params["dtype"])
206
      attention_bias = tf.cast(attention_bias, self.params["dtype"])
207
208
209
210
211
212
      with tf.name_scope("shift_targets"):
        # Shift targets to the right, and remove the last element
        decoder_inputs = tf.pad(decoder_inputs,
                                [[0, 0], [1, 0], [0, 0]])[:, :-1, :]
      with tf.name_scope("add_pos_encoding"):
        length = tf.shape(decoder_inputs)[1]
213
        pos_encoding = model_utils.get_position_encoding(
214
            length, self.params["hidden_size"])
215
216
        pos_encoding = tf.cast(pos_encoding, self.params["dtype"])
        decoder_inputs += pos_encoding
217
218
219
220
221
222
      if training:
        decoder_inputs = tf.nn.dropout(
            decoder_inputs, rate=self.params["layer_postprocess_dropout"])

      # Run values
      decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
223
          length, dtype=self.params["dtype"])
224
225
226
227
228
229
230
      outputs = self.decoder_stack(
          decoder_inputs,
          encoder_outputs,
          decoder_self_attention_bias,
          attention_bias,
          training=training)
      logits = self.embedding_softmax_layer(outputs, mode="linear")
231
      logits = tf.cast(logits, tf.float32)
232
233
234
235
236
237
238
      return logits

  def _get_symbols_to_logits_fn(self, max_decode_length, training):
    """Returns a decoding function that calculates logits of the next tokens."""

    timing_signal = model_utils.get_position_encoding(
        max_decode_length + 1, self.params["hidden_size"])
Reed's avatar
Reed committed
239
    timing_signal = tf.cast(timing_signal, self.params["dtype"])
240
    decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
Reed's avatar
Reed committed
241
        max_decode_length, dtype=self.params["dtype"])
242

A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
243
    # TODO(b/139770046): Refactor code with better naming of i.
244
245
246
247
248
    def symbols_to_logits_fn(ids, i, cache):
      """Generate logits for next potential IDs.

      Args:
        ids: Current decoded sequences. int tensor with shape [batch_size *
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
249
250
          beam_size, i + 1].
        i: Loop index.
251
252
253
254
255
256
257
258
259
260
261
262
263
264
        cache: dictionary of values storing the encoder output, encoder-decoder
          attention bias, and previous decoder attention values.

      Returns:
        Tuple of
          (logits with shape [batch_size * beam_size, vocab_size],
           updated cache values)
      """
      # Set decoder input to the last generated IDs
      decoder_input = ids[:, -1:]

      # Preprocess decoder input by getting embeddings and adding timing signal.
      decoder_input = self.embedding_softmax_layer(decoder_input)

A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
265
266
267
268
269
270
271
272
273
274
275
276
277
278
      if self.params["padded_decode"]:
        timing_signal_shape = timing_signal.shape.as_list()
        decoder_input += tf.slice(timing_signal, [i, 0],
                                  [1, timing_signal_shape[1]])

        bias_shape = decoder_self_attention_bias.shape.as_list()
        self_attention_bias = tf.slice(
            decoder_self_attention_bias, [0, 0, i, 0],
            [bias_shape[0], bias_shape[1], 1, bias_shape[3]])
      else:
        decoder_input += timing_signal[i:i + 1]

        self_attention_bias = decoder_self_attention_bias[:, :, i:i + 1, :i + 1]

279
280
281
282
283
284
      decoder_outputs = self.decoder_stack(
          decoder_input,
          cache.get("encoder_outputs"),
          self_attention_bias,
          cache.get("encoder_decoder_attention_bias"),
          training=training,
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
285
286
          cache=cache,
          decode_loop_step=i if self.params["padded_decode"] else None)
287
288
289
290
291
292
293
294
      logits = self.embedding_softmax_layer(decoder_outputs, mode="linear")
      logits = tf.squeeze(logits, axis=[1])
      return logits, cache

    return symbols_to_logits_fn

  def predict(self, encoder_outputs, encoder_decoder_attention_bias, training):
    """Return predicted sequence."""
295
    encoder_outputs = tf.cast(encoder_outputs, self.params["dtype"])
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
296
297
298
299
300
301
    if self.params["padded_decode"]:
      batch_size = encoder_outputs.shape.as_list()[0]
      input_length = encoder_outputs.shape.as_list()[1]
    else:
      batch_size = tf.shape(encoder_outputs)[0]
      input_length = tf.shape(encoder_outputs)[1]
302
    max_decode_length = input_length + self.params["extra_decode_length"]
Reed's avatar
Reed committed
303
304
    encoder_decoder_attention_bias = tf.cast(encoder_decoder_attention_bias,
                                             self.params["dtype"])
305
306
307
308
309
310
311
312
313

    symbols_to_logits_fn = self._get_symbols_to_logits_fn(
        max_decode_length, training)

    # Create initial set of IDs that will be passed into symbols_to_logits_fn.
    initial_ids = tf.zeros([batch_size], dtype=tf.int32)

    # Create cache storing decoder attention values for each layer.
    # pylint: disable=g-complex-comprehension
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
314
315
    init_decode_length = (
        max_decode_length if self.params["padded_decode"] else 0)
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
316
317
    num_heads = self.params["num_heads"]
    dim_per_head = self.params["hidden_size"] // num_heads
318
319
    cache = {
        "layer_%d" % layer: {
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
320
321
            "k":
                tf.zeros([
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
322
                    batch_size, init_decode_length, num_heads, dim_per_head
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
323
324
325
326
                ],
                         dtype=self.params["dtype"]),
            "v":
                tf.zeros([
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
327
                    batch_size, init_decode_length, num_heads, dim_per_head
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
328
329
                ],
                         dtype=self.params["dtype"])
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
        } for layer in range(self.params["num_hidden_layers"])
    }
    # pylint: enable=g-complex-comprehension

    # Add encoder output and attention bias to the cache.
    cache["encoder_outputs"] = encoder_outputs
    cache["encoder_decoder_attention_bias"] = encoder_decoder_attention_bias

    # Use beam search to find the top beam_size sequences and scores.
    decoded_ids, scores = beam_search.sequence_beam_search(
        symbols_to_logits_fn=symbols_to_logits_fn,
        initial_ids=initial_ids,
        initial_cache=cache,
        vocab_size=self.params["vocab_size"],
        beam_size=self.params["beam_size"],
        alpha=self.params["alpha"],
        max_decode_length=max_decode_length,
Reed's avatar
Reed committed
347
        eos_id=EOS_ID,
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
348
        padded_decode=self.params["padded_decode"],
Reed's avatar
Reed committed
349
        dtype=self.params["dtype"])
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368

    # Get the top sequence for each batch element
    top_decoded_ids = decoded_ids[:, 0, 1:]
    top_scores = scores[:, 0]

    return {"outputs": top_decoded_ids, "scores": top_scores}


class PrePostProcessingWrapper(tf.keras.layers.Layer):
  """Wrapper class that applies layer pre-processing and post-processing."""

  def __init__(self, layer, params):
    super(PrePostProcessingWrapper, self).__init__()
    self.layer = layer
    self.params = params
    self.postprocess_dropout = params["layer_postprocess_dropout"]

  def build(self, input_shape):
    # Create normalization layer
Hongkun Yu's avatar
Hongkun Yu committed
369
370
    self.layer_norm = tf.keras.layers.LayerNormalization(
        epsilon=1e-6, dtype="float32")
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
    super(PrePostProcessingWrapper, self).build(input_shape)

  def get_config(self):
    return {
        "params": self.params,
    }

  def call(self, x, *args, **kwargs):
    """Calls wrapped layer with same parameters."""
    # Preprocessing: apply layer normalization
    training = kwargs["training"]

    y = self.layer_norm(x)

    # Get layer output
    y = self.layer(y, *args, **kwargs)

    # Postprocessing: apply dropout and residual connection
    if training:
      y = tf.nn.dropout(y, rate=self.postprocess_dropout)
    return x + y


class EncoderStack(tf.keras.layers.Layer):
  """Transformer encoder stack.

  The encoder stack is made up of N identical layers. Each layer is composed
  of the sublayers:
    1. Self-attention layer
    2. Feedforward network (which is 2 fully-connected layers)
  """

  def __init__(self, params):
    super(EncoderStack, self).__init__()
    self.params = params
    self.layers = []

  def build(self, input_shape):
    """Builds the encoder stack."""
    params = self.params
    for _ in range(params["num_hidden_layers"]):
      # Create sublayers for each layer.
      self_attention_layer = attention_layer.SelfAttention(
          params["hidden_size"], params["num_heads"],
          params["attention_dropout"])
      feed_forward_network = ffn_layer.FeedForwardNetwork(
          params["hidden_size"], params["filter_size"], params["relu_dropout"])

      self.layers.append([
          PrePostProcessingWrapper(self_attention_layer, params),
          PrePostProcessingWrapper(feed_forward_network, params)
      ])

    # Create final layer normalization layer.
Hongkun Yu's avatar
Hongkun Yu committed
425
426
    self.output_normalization = tf.keras.layers.LayerNormalization(
        epsilon=1e-6, dtype="float32")
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
    super(EncoderStack, self).build(input_shape)

  def get_config(self):
    return {
        "params": self.params,
    }

  def call(self, encoder_inputs, attention_bias, inputs_padding, training):
    """Return the output of the encoder layer stacks.

    Args:
      encoder_inputs: tensor with shape [batch_size, input_length, hidden_size]
      attention_bias: bias for the encoder self-attention layer. [batch_size, 1,
        1, input_length]
      inputs_padding: tensor with shape [batch_size, input_length], inputs with
        zero paddings.
      training: boolean, whether in training mode or not.

    Returns:
      Output of encoder layer stack.
      float32 tensor with shape [batch_size, input_length, hidden_size]
    """
    for n, layer in enumerate(self.layers):
      # Run inputs through the sublayers.
      self_attention_layer = layer[0]
      feed_forward_network = layer[1]

      with tf.name_scope("layer_%d" % n):
        with tf.name_scope("self_attention"):
          encoder_inputs = self_attention_layer(
              encoder_inputs, attention_bias, training=training)
        with tf.name_scope("ffn"):
          encoder_inputs = feed_forward_network(
              encoder_inputs, training=training)

    return self.output_normalization(encoder_inputs)


class DecoderStack(tf.keras.layers.Layer):
  """Transformer decoder stack.

  Like the encoder stack, the decoder stack is made up of N identical layers.
  Each layer is composed of the sublayers:
    1. Self-attention layer
    2. Multi-headed attention layer combining encoder outputs with results from
       the previous self-attention layer.
    3. Feedforward network (2 fully-connected layers)
  """

  def __init__(self, params):
    super(DecoderStack, self).__init__()
    self.params = params
    self.layers = []

  def build(self, input_shape):
    """Builds the decoder stack."""
    params = self.params
    for _ in range(params["num_hidden_layers"]):
      self_attention_layer = attention_layer.SelfAttention(
          params["hidden_size"], params["num_heads"],
          params["attention_dropout"])
      enc_dec_attention_layer = attention_layer.Attention(
          params["hidden_size"], params["num_heads"],
          params["attention_dropout"])
      feed_forward_network = ffn_layer.FeedForwardNetwork(
          params["hidden_size"], params["filter_size"], params["relu_dropout"])

      self.layers.append([
          PrePostProcessingWrapper(self_attention_layer, params),
          PrePostProcessingWrapper(enc_dec_attention_layer, params),
          PrePostProcessingWrapper(feed_forward_network, params)
      ])
Hongkun Yu's avatar
Hongkun Yu committed
499
500
    self.output_normalization = tf.keras.layers.LayerNormalization(
        epsilon=1e-6, dtype="float32")
501
502
503
504
505
506
507
508
509
510
511
512
513
    super(DecoderStack, self).build(input_shape)

  def get_config(self):
    return {
        "params": self.params,
    }

  def call(self,
           decoder_inputs,
           encoder_outputs,
           decoder_self_attention_bias,
           attention_bias,
           training,
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
514
515
           cache=None,
           decode_loop_step=None):
516
517
518
    """Return the output of the decoder layer stacks.

    Args:
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
519
520
521
522
523
524
525
526
527
528
      decoder_inputs: A tensor with shape
        [batch_size, target_length, hidden_size].
      encoder_outputs: A tensor with shape
        [batch_size, input_length, hidden_size]
      decoder_self_attention_bias: A tensor with shape
        [1, 1, target_len, target_length], the bias for decoder self-attention
        layer.
      attention_bias: A tensor with shape [batch_size, 1, 1, input_length],
        the bias for encoder-decoder attention layer.
      training: A bool, whether in training mode or not.
529
530
      cache: (Used for fast decoding) A nested dictionary storing previous
        decoder self-attention values. The items are:
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
531
532
          {layer_n: {"k": A tensor with shape [batch_size, i, key_channels],
                     "v": A tensor with shape [batch_size, i, value_channels]},
533
                       ...}
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
534
535
      decode_loop_step: An integer, the step number of the decoding loop. Used
        only for autoregressive inference on TPU.
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554

    Returns:
      Output of decoder layer stack.
      float32 tensor with shape [batch_size, target_length, hidden_size]
    """
    for n, layer in enumerate(self.layers):
      self_attention_layer = layer[0]
      enc_dec_attention_layer = layer[1]
      feed_forward_network = layer[2]

      # Run inputs through the sublayers.
      layer_name = "layer_%d" % n
      layer_cache = cache[layer_name] if cache is not None else None
      with tf.name_scope(layer_name):
        with tf.name_scope("self_attention"):
          decoder_inputs = self_attention_layer(
              decoder_inputs,
              decoder_self_attention_bias,
              training=training,
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
555
556
              cache=layer_cache,
              decode_loop_step=decode_loop_step)
557
558
559
560
561
562
563
564
565
566
567
        with tf.name_scope("encdec_attention"):
          decoder_inputs = enc_dec_attention_layer(
              decoder_inputs,
              encoder_outputs,
              attention_bias,
              training=training)
        with tf.name_scope("ffn"):
          decoder_inputs = feed_forward_network(
              decoder_inputs, training=training)

    return self.output_normalization(decoder_inputs)