Transformer model for language understanding

View on TensorFlow.org Run in Google Colab View source on GitHub

This tutorial trains a Transformer model to translate Portuguese to English. This is an advanced example that assumes knowledge of text generation and attention.

The core idea behind the Transformer model is self-attention—the ability to attend to different positions of the input sequence to compute a representation of that sequence. Transformer creates stacks of self-attention layers and is explained below in the sections Scaled dot product attention and Multi-head attention.

A transformer model handles variable-sized input using stacks of self-attention layers instead of RNNs or CNNs. This general architecture has a number of advantages:

  • It make no assumptions about the temporal/spatial relationships across the data. This is ideal for processing a set of objects (for example, StarCraft units).
  • Layer outputs can be calculated in parallel, instead of a series like an RNN.
  • Distant items can affect each other's output without passing through many RNN-steps, or convolution layers (see Scene Memory Transformer for example).
  • It can learn long-range dependencies. This is a challenge in many sequence tasks.

The downsides of this architecture are:

  • For a time-series, the output for a time-step is calculated from the entire history instead of only the inputs and current hidden-state. This may be less efficient.
  • If the input does have a temporal/spatial relationship, like text, some positional encoding must be added or the model will effectively see a bag of words.

After training the model in this notebook, you will be able to input a Portuguese sentence and return the English translation.

Attention heatmap

from __future__ import absolute_import, division, print_function, unicode_literals

!pip install -q tf-nightly-gpu-2.0-preview
import tensorflow_datasets as tfds
import tensorflow as tf

import time
import numpy as np
import matplotlib.pyplot as plt

Setup input pipeline

Use TFDS to load the Portugese-English translation dataset from the TED Talks Open Translation Project.

This dataset contains approximately 50000 training examples, 1100 validation examples, and 2000 test examples.

examples, metadata = tfds.load('ted_hrlr_translate/pt_to_en', with_info=True,
                               as_supervised=True)
train_examples, val_examples = examples['train'], examples['validation']
Downloading and preparing dataset ted_hrlr_translate (124.94 MiB) to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/0.0.1...

HBox(children=(IntProgress(value=1, bar_style='info', description='Dl Completed...', max=1, style=ProgressStyl…
HBox(children=(IntProgress(value=1, bar_style='info', description='Dl Size...', max=1, style=ProgressStyle(des…
HBox(children=(IntProgress(value=1, bar_style='info', description='Extraction completed...', max=1, style=Prog…





HBox(children=(IntProgress(value=1, bar_style='info', max=1), HTML(value='')))

HBox(children=(IntProgress(value=0, description='Shuffling...', max=1, style=ProgressStyle(description_width='…
WARNING: Logging before flag parsing goes to stderr.
W0514 17:23:44.727313 140548191815424 deprecation.py:323] From /home/kbuilder/.local/lib/python3.5/site-packages/tensorflow_datasets/core/file_format_adapter.py:247: tf_record_iterator (from tensorflow.python.lib.io.tf_record) is deprecated and will be removed in a future version.
Instructions for updating:
Use eager execution and: 
`tf.data.TFRecordDataset(path)`

HBox(children=(IntProgress(value=1, bar_style='info', description='Reading...', max=1, style=ProgressStyle(des…
HBox(children=(IntProgress(value=0, description='Writing...', max=51785, style=ProgressStyle(description_width…

HBox(children=(IntProgress(value=1, bar_style='info', max=1), HTML(value='')))

HBox(children=(IntProgress(value=0, description='Shuffling...', max=1, style=ProgressStyle(description_width='…
HBox(children=(IntProgress(value=1, bar_style='info', description='Reading...', max=1, style=ProgressStyle(des…
HBox(children=(IntProgress(value=0, description='Writing...', max=1193, style=ProgressStyle(description_width=…

HBox(children=(IntProgress(value=1, bar_style='info', max=1), HTML(value='')))

HBox(children=(IntProgress(value=0, description='Shuffling...', max=1, style=ProgressStyle(description_width='…
HBox(children=(IntProgress(value=1, bar_style='info', description='Reading...', max=1, style=ProgressStyle(des…
HBox(children=(IntProgress(value=0, description='Writing...', max=1803, style=ProgressStyle(description_width=…
Dataset ted_hrlr_translate downloaded and prepared to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/0.0.1. Subsequent calls will reuse this data.

Create a custom subwords tokenizer from the training dataset.

tokenizer_en = tfds.features.text.SubwordTextEncoder.build_from_corpus(
    (en.numpy() for pt, en in train_examples), target_vocab_size=2**13)

tokenizer_pt = tfds.features.text.SubwordTextEncoder.build_from_corpus(
    (pt.numpy() for pt, en in train_examples), target_vocab_size=2**13)
sample_string = 'Transformer is awesome.'

tokenized_string = tokenizer_en.encode(sample_string)
print ('Tokenized string is {}'.format(tokenized_string))

original_string = tokenizer_en.decode(tokenized_string)
print ('The original string: {}'.format(original_string))

assert original_string == sample_string
Tokenized string is [7915, 1248, 7946, 7194, 13, 2799, 7877]
The original string: Transformer is awesome.

The tokenizer encodes the string by breaking it into subwords if the word is not in its dictionary.

for ts in tokenized_string:
  print ('{} ----> {}'.format(ts, tokenizer_en.decode([ts])))
7915 ----> T
1248 ----> ran
7946 ----> s
7194 ----> former 
13 ----> is 
2799 ----> awesome
7877 ----> .
BUFFER_SIZE = 20000
BATCH_SIZE = 64

Add a start and end token to the input and target.

def encode(lang1, lang2):
  lang1 = [tokenizer_pt.vocab_size] + tokenizer_pt.encode(
      lang1.numpy()) + [tokenizer_pt.vocab_size+1]

  lang2 = [tokenizer_en.vocab_size] + tokenizer_en.encode(
      lang2.numpy()) + [tokenizer_en.vocab_size+1]
  
  return lang1, lang2
MAX_LENGTH = 40
def filter_max_length(x, y, max_length=MAX_LENGTH):
  return tf.logical_and(tf.size(x) <= max_length,
                        tf.size(y) <= max_length)

Operations inside .map() run in graph mode and receive a graph tensor that do not have a numpy attribute. The tokenizer expects a string or Unicode symbol to encode it into integers. Hence, you need to run the encoding inside a tf.py_function, which receives an eager tensor having a numpy attribute that contains the string value.

def tf_encode(pt, en):
  return tf.py_function(encode, [pt, en], [tf.int64, tf.int64])
train_dataset = train_examples.map(tf_encode)
train_dataset = train_dataset.filter(filter_max_length)
# cache the dataset to memory to get a speedup while reading from it.
train_dataset = train_dataset.cache()
train_dataset = train_dataset.shuffle(BUFFER_SIZE).padded_batch(
    BATCH_SIZE, padded_shapes=([-1], [-1]))
train_dataset = train_dataset.prefetch(tf.data.experimental.AUTOTUNE)


val_dataset = val_examples.map(tf_encode)
val_dataset = val_dataset.filter(filter_max_length).padded_batch(
    BATCH_SIZE, padded_shapes=([-1], [-1]))
pt_batch, en_batch = next(iter(val_dataset))
pt_batch, en_batch
W0514 17:26:46.915636 140543325624064 backprop.py:820] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string
W0514 17:26:46.917539 140543325624064 backprop.py:820] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string
W0514 17:26:46.919743 140543325624064 backprop.py:820] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string
W0514 17:26:46.921139 140543325624064 backprop.py:820] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string
W0514 17:26:46.923256 140543300445952 backprop.py:820] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string

(<tf.Tensor: id=311410, shape=(64, 40), dtype=int64, numpy=
 array([[8214, 1259,    5, ...,    0,    0,    0],
        [8214,  299,   13, ...,    0,    0,    0],
        [8214,   59,    8, ...,    0,    0,    0],
        ...,
        [8214,   95,    3, ...,    0,    0,    0],
        [8214, 5157,    1, ...,    0,    0,    0],
        [8214, 4479, 7990, ...,    0,    0,    0]])>,
 <tf.Tensor: id=311411, shape=(64, 40), dtype=int64, numpy=
 array([[8087,   18,   12, ...,    0,    0,    0],
        [8087,  634,   30, ...,    0,    0,    0],
        [8087,   16,   13, ...,    0,    0,    0],
        ...,
        [8087,   12,   20, ...,    0,    0,    0],
        [8087,   17, 4981, ...,    0,    0,    0],
        [8087,   12, 5453, ...,    0,    0,    0]])>)

Positional encoding

Since this model doesn't contain any recurrence or convolution, positional encoding is added to give the model some information about the relative position of the words in the sentence.

The positional encoding vector is added to the embedding vector. Embeddings represent a token in a d-dimensional space where tokens with similar meaning will be closer to each other. But the embeddings do not encode the relative position of words in a sentence. So after adding the positional encoding, words will be closer to each other based on the similarity of their meaning and their position in the sentence, in the d-dimensional space.

See the notebook on positional encoding to learn more about it. The formula for calculating the positional encoding is as follows:

$$\Large{PE_{(pos, 2i)} = sin(pos / 10000^{2i / d_{model}})} $$
$$\Large{PE_{(pos, 2i+1)} = cos(pos / 10000^{2i / d_{model}})} $$
def get_angles(pos, i, d_model):
  angle_rates = 1 / np.power(10000, (2 * (i//2)) / np.float32(d_model))
  return pos * angle_rates
def positional_encoding(position, d_model):
  angle_rads = get_angles(np.arange(position)[:, np.newaxis],
                          np.arange(d_model)[np.newaxis, :],
                          d_model)
  
  # apply sin to even indices in the array; 2i
  sines = np.sin(angle_rads[:, 0::2])
  
  # apply cos to odd indices in the array; 2i+1
  cosines = np.cos(angle_rads[:, 1::2])
  
  pos_encoding = np.concatenate([sines, cosines], axis=-1)
  
  pos_encoding = pos_encoding[np.newaxis, ...]
    
  return tf.cast(pos_encoding, dtype=tf.float32)
pos_encoding = positional_encoding(50, 512)
print (pos_encoding.shape)

plt.pcolormesh(pos_encoding[0], cmap='RdBu')
plt.xlabel('Depth')
plt.xlim((0, 512))
plt.ylabel('Position')
plt.colorbar()
plt.show()
(1, 50, 512)

png

Masking

Mask all the pad tokens in the batch of sequence. It ensures that the model does not treat padding as the input. The mask indicates where pad value 0 is present: it outputs a 1 at those locations, and a 0 otherwise.

def create_padding_mask(seq):
  seq = tf.cast(tf.math.equal(seq, 0), tf.float32)
  
  # add extra dimensions so that we can add the padding
  # to the attention logits.
  return seq[:, tf.newaxis, tf.newaxis, :]  # (batch_size, 1, 1, seq_len)
x = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
create_padding_mask(x)
<tf.Tensor: id=311428, shape=(3, 1, 1, 5), dtype=float32, numpy=
array([[[[0., 0., 1., 1., 0.]]],


       [[[0., 0., 0., 1., 1.]]],


       [[[1., 1., 1., 0., 0.]]]], dtype=float32)>

The look-ahead mask is used to mask the future tokens in a sequence. In other words, the mask indicates which entries should not be used.

This means that to predict the third word, only the first and second word will be used. Similarly to predict the fourth word, only the first, second and the third word will be used and so on.

def create_look_ahead_mask(size):
  mask = 1 - tf.linalg.band_part(tf.ones((size, size)), -1, 0)
  return mask  # (seq_len, seq_len)
x = tf.random.uniform((1, 3))
temp = create_look_ahead_mask(x.shape[1])
temp
<tf.Tensor: id=311444, shape=(3, 3), dtype=float32, numpy=
array([[0., 1., 1.],
       [0., 0., 1.],
       [0., 0., 0.]], dtype=float32)>

Scaled dot product attention

scaled_dot_product_attention

The attention function used by the transformer takes three inputs: Q (query), K (key), V (value). The equation used to calculate the attention weights is:

$$\Large{Attention(Q, K, V) = softmax_k(\frac{QK^T}{\sqrt{d_k}}) V} $$

The dot-product attention is scaled by a factor of square root of the depth. This is done because for large values of depth, the dot product grows large in magnitude pushing the softmax function where it has small gradients resulting in a very hard softmax.

For example, consider that Q and K have a mean of 0 and variance of 1. Their matrix multiplication will have a mean of 0 and variance of dk. Hence, square root of dk is used for scaling (and not any other number) because the matmul of Q and K should have a mean of 0 and variance of 1, so that we get a gentler softmax.

The mask is multiplied with -1e9 (close to negative infinity). This is done because the mask is summed with the scaled matrix multiplication of Q and K and is applied immediately before a softmax. The goal is to zero out these cells, and large negative inputs to softmax are near zero in the output.

def scaled_dot_product_attention(q, k, v, mask):
  """Calculate the attention weights.
  q, k, v must have matching leading dimensions.
  k, v must have matching penultimate dimension, i.e.: seq_len_k = seq_len_v.
  The mask has different shapes depending on its type(padding or look ahead) 
  but it must be broadcastable for addition.
  
  Args:
    q: query shape == (..., seq_len_q, depth)
    k: key shape == (..., seq_len_k, depth)
    v: value shape == (..., seq_len_v, depth_v)
    mask: Float tensor with shape broadcastable 
          to (..., seq_len_q, seq_len_k). Defaults to None.
    
  Returns:
    output, attention_weights
  """

  matmul_qk = tf.matmul(q, k, transpose_b=True)  # (..., seq_len_q, seq_len_k)
  
  # scale matmul_qk
  dk = tf.cast(tf.shape(k)[-1], tf.float32)
  scaled_attention_logits = matmul_qk / tf.math.sqrt(dk)

  # add the mask to the scaled tensor.
  if mask is not None:
    scaled_attention_logits += (mask * -1e9)

  # softmax is normalized on the last axis (seq_len_k) so that the scores
  # add up to 1.
  attention_weights = tf.nn.softmax(scaled_attention_logits, axis=-1)  # (..., seq_len_q, seq_len_k)

  output = tf.matmul(attention_weights, v)  # (..., seq_len_v, depth_v)

  return output, attention_weights

As the softmax normalization is done on K, its values decide the amount of importance given to Q.

The output represents the multiplication of the attention weights and the V (value) vector. This ensures that the words we want to focus on are kept as is and the irrelevant words are flushed out.

def print_out(q, k, v):
  temp_out, temp_attn = scaled_dot_product_attention(
      q, k, v, None)
  print ('Attention weights are:')
  print (temp_attn)
  print ('Output is:')
  print (temp_out)
np.set_printoptions(suppress=True)

temp_k = tf.constant([[10,0,0],
                      [0,10,0],
                      [0,0,10],
                      [0,0,10]], dtype=tf.float32)  # (4, 3)

temp_v = tf.constant([[   1,0],
                      [  10,0],
                      [ 100,5],
                      [1000,6]], dtype=tf.float32)  # (4, 2)

# This `query` aligns with the second `key`,
# so the second `value` is returned.
temp_q = tf.constant([[0, 10, 0]], dtype=tf.float32)  # (1, 3)
print_out(temp_q, temp_k, temp_v)
Attention weights are:
tf.Tensor([[0. 1. 0. 0.]], shape=(1, 4), dtype=float32)
Output is:
tf.Tensor([[10.  0.]], shape=(1, 2), dtype=float32)
# This query aligns with a repeated key (third and fourth), 
# so all associated values get averaged.
temp_q = tf.constant([[0, 0, 10]], dtype=tf.float32)  # (1, 3)
print_out(temp_q, temp_k, temp_v)
Attention weights are:
tf.Tensor([[0.  0.  0.5 0.5]], shape=(1, 4), dtype=float32)
Output is:
tf.Tensor([[550.    5.5]], shape=(1, 2), dtype=float32)
# This query aligns equally with the first and second key, 
# so their values get averaged.
temp_q = tf.constant([[10, 10, 0]], dtype=tf.float32)  # (1, 3)
print_out(temp_q, temp_k, temp_v)
Attention weights are:
tf.Tensor([[0.5 0.5 0.  0. ]], shape=(1, 4), dtype=float32)
Output is:
tf.Tensor([[5.5 0. ]], shape=(1, 2), dtype=float32)

Pass all the queries together.

temp_q = tf.constant([[0, 0, 10], [0, 10, 0], [10, 10, 0]], dtype=tf.float32)  # (3, 3)
print_out(temp_q, temp_k, temp_v)
Attention weights are:
tf.Tensor(
[[0.  0.  0.5 0.5]
 [0.  1.  0.  0. ]
 [0.5 0.5 0.  0. ]], shape=(3, 4), dtype=float32)
Output is:
tf.Tensor(
[[550.    5.5]
 [ 10.    0. ]
 [  5.5   0. ]], shape=(3, 2), dtype=float32)

Multi-head attention

multi-head attention

Multi-head attention consists of four parts: * Linear layers and split into heads. * Scaled dot-product attention. * Concatenation of heads. * Final linear layer.

Each multi-head attention block gets three inputs; Q (query), K (key), V (value). These are put through linear (Dense) layers and split up into multiple heads.

The scaled_dot_product_attention defined above is applied to each head (broadcasted for efficiency). An appropriate mask must be used in the attention step. The attention output for each head is then concatenated (using tf.transpose, and tf.reshape) and put through a final Dense layer.

Instead of one single attention head, Q, K, and V are split into multiple heads because it allows the model to jointly attend to information at different positions from different representational spaces. After the split each head has a reduced dimensionality, so the total computation cost is the same as a single head attention with full dimensionality.

class MultiHeadAttention(tf.keras.layers.Layer):
  def __init__(self, d_model, num_heads):
    super(MultiHeadAttention, self).__init__()
    self.num_heads = num_heads
    self.d_model = d_model
    
    assert d_model % self.num_heads == 0
    
    self.depth = d_model // self.num_heads
    
    self.wq = tf.keras.layers.Dense(d_model)
    self.wk = tf.keras.layers.Dense(d_model)
    self.wv = tf.keras.layers.Dense(d_model)
    
    self.dense = tf.keras.layers.Dense(d_model)
        
  def split_heads(self, x, batch_size):
    """Split the last dimension into (num_heads, depth).
    Transpose the result such that the shape is (batch_size, num_heads, seq_len, depth)
    """
    x = tf.reshape(x, (batch_size, -1, self.num_heads, self.depth))
    return tf.transpose(x, perm=[0, 2, 1, 3])
    
  def call(self, v, k, q, mask):
    batch_size = tf.shape(q)[0]
    
    q = self.wq(q)  # (batch_size, seq_len, d_model)
    k = self.wk(k)  # (batch_size, seq_len, d_model)
    v = self.wv(v)  # (batch_size, seq_len, d_model)
    
    q = self.split_heads(q, batch_size)  # (batch_size, num_heads, seq_len_q, depth)
    k = self.split_heads(k, batch_size)  # (batch_size, num_heads, seq_len_k, depth)
    v = self.split_heads(v, batch_size)  # (batch_size, num_heads, seq_len_v, depth)
    
    # scaled_attention.shape == (batch_size, num_heads, seq_len_v, depth)
    # attention_weights.shape == (batch_size, num_heads, seq_len_q, seq_len_k)
    scaled_attention, attention_weights = scaled_dot_product_attention(
        q, k, v, mask)
    
    scaled_attention = tf.transpose(scaled_attention, perm=[0, 2, 1, 3])  # (batch_size, seq_len_v, num_heads, depth)

    concat_attention = tf.reshape(scaled_attention, 
                                  (batch_size, -1, self.d_model))  # (batch_size, seq_len_v, d_model)

    output = self.dense(concat_attention)  # (batch_size, seq_len_v, d_model)
        
    return output, attention_weights

Create a MultiHeadAttention layer to try out. At each location in the sequence, y, the MultiHeadAttention runs all 8 attention heads across all other locations in the sequence, returning a new vector of the same length at each location.

temp_mha = MultiHeadAttention(d_model=512, num_heads=8)
y = tf.random.uniform((1, 60, 512))  # (batch_size, encoder_sequence, d_model)
out, attn = temp_mha(y, k=y, q=y, mask=None)
out.shape, attn.shape
(TensorShape([1, 60, 512]), TensorShape([1, 8, 60, 60]))

Point wise feed forward network

Point wise feed forward network consists of two fully-connected layers with a ReLU activation in between.

def point_wise_feed_forward_network(d_model, dff):
  return tf.keras.Sequential([
      tf.keras.layers.Dense(dff, activation='relu'),  # (batch_size, seq_len, dff)
      tf.keras.layers.Dense(d_model)  # (batch_size, seq_len, d_model)
  ])
sample_ffn = point_wise_feed_forward_network(512, 2048)
sample_ffn(tf.random.uniform((64, 50, 512))).shape
TensorShape([64, 50, 512])

Encoder and decoder

transformer

The transformer model follows the same general pattern as a standard sequence to sequence with attention model.

  • The input sentence is passed through N encoder layers that generates an output for each word/token in the sequence.
  • The decoder attends on the encoder's output and its own input (self-attention) to predict the next word.

Encoder layer

Each encoder layer consists of sublayers:

  1. Multi-head attention (with padding mask)
  2. Point wise feed forward networks.

Each of these sublayers has a residual connection around it followed by a layer normalization. Residual connections help in avoiding the vanishing gradient problem in deep networks.

The output of each sublayer is LayerNorm(x + Sublayer(x)). The normalization is done on the d_model (last) axis. There are N encoder layers in the transformer.

class EncoderLayer(tf.keras.layers.Layer):
  def __init__(self, d_model, num_heads, dff, rate=0.1):
    super(EncoderLayer, self).__init__()

    self.mha = MultiHeadAttention(d_model, num_heads)
    self.ffn = point_wise_feed_forward_network(d_model, dff)

    self.layernorm1 = tf.keras.layers.LayerNormalization(epsilon=1e-6)
    self.layernorm2 = tf.keras.layers.LayerNormalization(epsilon=1e-6)
    
    self.dropout1 = tf.keras.layers.Dropout(rate)
    self.dropout2 = tf.keras.layers.Dropout(rate)
    
  def call(self, x, training, mask):

    attn_output, _ = self.mha(x, x, x, mask)  # (batch_size, input_seq_len, d_model)
    attn_output = self.dropout1(attn_output, training=training)
    out1 = self.layernorm1(x + attn_output)  # (batch_size, input_seq_len, d_model)
    
    ffn_output = self.ffn(out1)  # (batch_size, input_seq_len, d_model)
    ffn_output = self.dropout2(ffn_output, training=training)
    out2 = self.layernorm2(out1 + ffn_output)  # (batch_size, input_seq_len, d_model)
    
    return out2
sample_encoder_layer = EncoderLayer(512, 8, 2048)

sample_encoder_layer_output = sample_encoder_layer(
    tf.random.uniform((64, 43, 512)), False, None)

sample_encoder_layer_output.shape  # (batch_size, input_seq_len, d_model)
TensorShape([64, 43, 512])

Decoder layer

Each decoder layer consists of sublayers:

  1. Masked multi-head attention (with look ahead mask and padding mask)
  2. Multi-head attention (with padding mask). V (value) and K (key) receive the encoder output as inputs. Q (query) receives the output from the masked multi-head attention sublayer.
  3. Point wise feed forward networks

Each of these sublayers has a residual connection around it followed by a layer normalization. The output of each sublayer is LayerNorm(x + Sublayer(x)). The normalization is done on the d_model (last) axis.

There are N decoder layers in the transformer.

As Q receives the output from decoder's first attention block, and K receives the encoder output, the attention weights represent the importance given to the decoder's input based on the encoder's output. In other words, the decoder predicts the next word by looking at the encoder output and self-attending to its own output. See the demonstration above in the scaled dot product attention section.

class DecoderLayer(tf.keras.layers.Layer):
  def __init__(self, d_model, num_heads, dff, rate=0.1):
    super(DecoderLayer, self).__init__()

    self.mha1 = MultiHeadAttention(d_model, num_heads)
    self.mha2 = MultiHeadAttention(d_model, num_heads)

    self.ffn = point_wise_feed_forward_network(d_model, dff)
 
    self.layernorm1 = tf.keras.layers.LayerNormalization(epsilon=1e-6)
    self.layernorm2 = tf.keras.layers.LayerNormalization(epsilon=1e-6)
    self.layernorm3 = tf.keras.layers.LayerNormalization(epsilon=1e-6)
    
    self.dropout1 = tf.keras.layers.Dropout(rate)
    self.dropout2 = tf.keras.layers.Dropout(rate)
    self.dropout3 = tf.keras.layers.Dropout(rate)
    
    
  def call(self, x, enc_output, training, 
           look_ahead_mask, padding_mask):
    # enc_output.shape == (batch_size, input_seq_len, d_model)

    attn1, attn_weights_block1 = self.mha1(x, x, x, look_ahead_mask)  # (batch_size, target_seq_len, d_model)
    attn1 = self.dropout1(attn1, training=training)
    out1 = self.layernorm1(attn1 + x)
    
    attn2, attn_weights_block2 = self.mha2(
        enc_output, enc_output, out1, padding_mask)  # (batch_size, target_seq_len, d_model)
    attn2 = self.dropout2(attn2, training=training)
    out2 = self.layernorm2(attn2 + out1)  # (batch_size, target_seq_len, d_model)
    
    ffn_output = self.ffn(out2)  # (batch_size, target_seq_len, d_model)
    ffn_output = self.dropout3(ffn_output, training=training)
    out3 = self.layernorm3(ffn_output + out2)  # (batch_size, target_seq_len, d_model)
    
    return out3, attn_weights_block1, attn_weights_block2
sample_decoder_layer = DecoderLayer(512, 8, 2048)

sample_decoder_layer_output, _, _ = sample_decoder_layer(
    tf.random.uniform((64, 50, 512)), sample_encoder_layer_output, 
    False, None, None)

sample_decoder_layer_output.shape  # (batch_size, target_seq_len, d_model)
TensorShape([64, 50, 512])

Encoder

The Encoder consists of: 1. Input Embedding 2. Positional Encoding 3. N encoder layers

The input is put through an embedding which is summed with the positional encoding. The output of this summation is the input to the encoder layers. The output of the encoder is the input to the decoder.

class Encoder(tf.keras.layers.Layer):
  def __init__(self, num_layers, d_model, num_heads, dff, input_vocab_size, 
               rate=0.1):
    super(Encoder, self).__init__()

    self.d_model = d_model
    self.num_layers = num_layers
    
    self.embedding = tf.keras.layers.Embedding(input_vocab_size, d_model)
    self.pos_encoding = positional_encoding(input_vocab_size, self.d_model)
    
    
    self.enc_layers = [EncoderLayer(d_model, num_heads, dff, rate) 
                       for _ in range(num_layers)]
  
    self.dropout = tf.keras.layers.Dropout(rate)
        
  def call(self, x, training, mask):

    seq_len = tf.shape(x)[1]
    
    # adding embedding and position encoding.
    x = self.embedding(x)  # (batch_size, input_seq_len, d_model)
    x *= tf.math.sqrt(tf.cast(self.d_model, tf.float32))
    x += self.pos_encoding[:, :seq_len, :]

    x = self.dropout(x, training=training)
    
    for i in range(self.num_layers):
      x = self.enc_layers[i](x, training, mask)
    
    return x  # (batch_size, input_seq_len, d_model)
sample_encoder = Encoder(num_layers=2, d_model=512, num_heads=8, 
                         dff=2048, input_vocab_size=8500)

sample_encoder_output = sample_encoder(tf.random.uniform((64, 62)), 
                                       training=False, mask=None)

print (sample_encoder_output.shape)  # (batch_size, input_seq_len, d_model)
(64, 62, 512)

Decoder

The Decoder consists of: 1. Output Embedding 2. Positional Encoding 3. N decoder layers

The target is put through an embedding which is summed with the positional encoding. The output of this summation is the input to the decoder layers. The output of the decoder is the input to the final linear layer.

class Decoder(tf.keras.layers.Layer):
  def __init__(self, num_layers, d_model, num_heads, dff, target_vocab_size, 
               rate=0.1):
    super(Decoder, self).__init__()

    self.d_model = d_model
    self.num_layers = num_layers
    
    self.embedding = tf.keras.layers.Embedding(target_vocab_size, d_model)
    self.pos_encoding = positional_encoding(target_vocab_size, self.d_model)
    
    self.dec_layers = [DecoderLayer(d_model, num_heads, dff, rate) 
                       for _ in range(num_layers)]
    self.dropout = tf.keras.layers.Dropout(rate)
    
  def call(self, x, enc_output, training, 
           look_ahead_mask, padding_mask):

    seq_len = tf.shape(x)[1]
    attention_weights = {}
    
    x = self.embedding(x)  # (batch_size, target_seq_len, d_model)
    x *= tf.math.sqrt(tf.cast(self.d_model, tf.float32))
    x += self.pos_encoding[:, :seq_len, :]
    
    x = self.dropout(x, training=training)

    for i in range(self.num_layers):
      x, block1, block2 = self.dec_layers[i](x, enc_output, training,
                                             look_ahead_mask, padding_mask)
      
      attention_weights['decoder_layer{}_block1'.format(i+1)] = block1
      attention_weights['decoder_layer{}_block2'.format(i+1)] = block2
    
    # x.shape == (batch_size, target_seq_len, d_model)
    return x, attention_weights
sample_decoder = Decoder(num_layers=2, d_model=512, num_heads=8, 
                         dff=2048, target_vocab_size=8000)

output, attn = sample_decoder(tf.random.uniform((64, 26)), 
                              enc_output=sample_encoder_output, 
                              training=False, look_ahead_mask=None, 
                              padding_mask=None)

output.shape, attn['decoder_layer2_block2'].shape
(TensorShape([64, 26, 512]), TensorShape([64, 8, 26, 62]))

Create the Transformer

Transformer consists of the encoder, decoder and a final linear layer. The output of the decoder is the input to the linear layer and its output is returned.

class Transformer(tf.keras.Model):
  def __init__(self, num_layers, d_model, num_heads, dff, input_vocab_size, 
               target_vocab_size, rate=0.1):
    super(Transformer, self).__init__()

    self.encoder = Encoder(num_layers, d_model, num_heads, dff, 
                           input_vocab_size, rate)

    self.decoder = Decoder(num_layers, d_model, num_heads, dff, 
                           target_vocab_size, rate)

    self.final_layer = tf.keras.layers.Dense(target_vocab_size)
    
  def call(self, inp, tar, training, enc_padding_mask, 
           look_ahead_mask, dec_padding_mask):

    enc_output = self.encoder(inp, training, enc_padding_mask)  # (batch_size, inp_seq_len, d_model)
    
    # dec_output.shape == (batch_size, tar_seq_len, d_model)
    dec_output, attention_weights = self.decoder(
        tar, enc_output, training, look_ahead_mask, dec_padding_mask)
    
    final_output = self.final_layer(dec_output)  # (batch_size, tar_seq_len, target_vocab_size)
    
    return final_output, attention_weights
sample_transformer = Transformer(
    num_layers=2, d_model=512, num_heads=8, dff=2048, 
    input_vocab_size=8500, target_vocab_size=8000)

temp_input = tf.random.uniform((64, 62))
temp_target = tf.random.uniform((64, 26))

fn_out, _ = sample_transformer(temp_input, temp_target, training=False, 
                               enc_padding_mask=None, 
                               look_ahead_mask=None,
                               dec_padding_mask=None)

fn_out.shape  # (batch_size, tar_seq_len, target_vocab_size)
TensorShape([64, 26, 8000])

Set hyperparameters

To keep this example small and relatively fast, the values for num_layers, d_model, and dff have been reduced.

The values used in the base model of transformer were; num_layers=6, d_model = 512, dff = 2048. See the paper for all the other versions of the transformer.

num_layers = 4
d_model = 128
dff = 512
num_heads = 8

input_vocab_size = tokenizer_pt.vocab_size + 2
target_vocab_size = tokenizer_en.vocab_size + 2
dropout_rate = 0.1

Optimizer

Use the Adam optimizer with a custom learning rate scheduler according to the formula in the paper.

$$\Large{lrate = d_{model}^{-0.5} * min(step{\_}num^{-0.5}, step{\_}num * warmup{\_}steps^{-1.5})}$$
class CustomSchedule(tf.keras.optimizers.schedules.LearningRateSchedule):
  def __init__(self, d_model, warmup_steps=4000):
    super(CustomSchedule, self).__init__()
    
    self.d_model = d_model
    self.d_model = tf.cast(self.d_model, tf.float32)

    self.warmup_steps = warmup_steps
    
  def __call__(self, step):
    arg1 = tf.math.rsqrt(step)
    arg2 = step * (self.warmup_steps ** -1.5)
    
    return tf.math.rsqrt(self.d_model) * tf.math.minimum(arg1, arg2)
learning_rate = CustomSchedule(d_model)

optimizer = tf.keras.optimizers.Adam(learning_rate, beta_1=0.9, beta_2=0.98, 
                                     epsilon=1e-9)
temp_learning_rate_schedule = CustomSchedule(d_model)

plt.plot(temp_learning_rate_schedule(tf.range(40000, dtype=tf.float32)))
plt.ylabel("Learning Rate")
plt.xlabel("Train Step")
Text(0.5, 0, 'Train Step')

png

Loss and metrics

Since the target sequences are padded, it is important to apply a padding mask when calculating the loss.

loss_object = tf.keras.losses.SparseCategoricalCrossentropy(
    from_logits=True, reduction='none')
def loss_function(real, pred):
  mask = tf.math.logical_not(tf.math.equal(real, 0))
  loss_ = loss_object(real, pred)

  mask = tf.cast(mask, dtype=loss_.dtype)
  loss_ *= mask
  
  return tf.reduce_mean(loss_)
train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
    name='train_accuracy')

Training and checkpointing

transformer = Transformer(num_layers, d_model, num_heads, dff,
                          input_vocab_size, target_vocab_size, dropout_rate)
def create_masks(inp, tar):
  # Encoder padding mask
  enc_padding_mask = create_padding_mask(inp)
  
  # Used in the 2nd attention block in the decoder.
  # This padding mask is used to mask the encoder outputs.
  dec_padding_mask = create_padding_mask(inp)
  
  # Used in the 1st attention block in the decoder.
  # It is used to pad and mask future tokens in the input received by 
  # the decoder.
  look_ahead_mask = create_look_ahead_mask(tf.shape(tar)[1])
  dec_target_padding_mask = create_padding_mask(tar)
  combined_mask = tf.maximum(dec_target_padding_mask, look_ahead_mask)
  
  return enc_padding_mask, combined_mask, dec_padding_mask

Create the checkpoint path and the checkpoint manager. This will be used to save checkpoints every n epochs.

checkpoint_path = "./checkpoints/train"

ckpt = tf.train.Checkpoint(transformer=transformer,
                           optimizer=optimizer)

ckpt_manager = tf.train.CheckpointManager(ckpt, checkpoint_path, max_to_keep=5)

# if a checkpoint exists, restore the latest checkpoint.
if ckpt_manager.latest_checkpoint:
  ckpt.restore(ckpt_manager.latest_checkpoint)
  print ('Latest checkpoint restored!!')

The target is divided into tar_inp and tar_real. tar_inp is passed as an input to the decoder. tar_real is that same input shifted by 1: At each location in tar_input, tar_real contains the next token that should be predicted.

For example, sentence = "SOS A lion in the jungle is sleeping EOS"

tar_inp = "SOS A lion in the jungle is sleeping"

tar_real = "A lion in the jungle is sleeping EOS"

The transformer is an auto-regressive model: it makes predictions one part at a time, and uses its output so far to decide what to do next.

During training this example uses teacher-forcing (like in the text generation tutorial). Teacher forcing is passing the true output to the next time step regardless of what the model predicts at the current time step.

As the transformer predicts each word, self-attention allows it to look at the previous words in the input sequence to better predict the next word.

To prevent the model from peaking at the expected output the model uses a look-ahead mask.

EPOCHS = 20
@tf.function
def train_step(inp, tar):
  tar_inp = tar[:, :-1]
  tar_real = tar[:, 1:]
  
  enc_padding_mask, combined_mask, dec_padding_mask = create_masks(inp, tar_inp)
  
  with tf.GradientTape() as tape:
    predictions, _ = transformer(inp, tar_inp, 
                                 True, 
                                 enc_padding_mask, 
                                 combined_mask, 
                                 dec_padding_mask)
    loss = loss_function(tar_real, predictions)

  gradients = tape.gradient(loss, transformer.trainable_variables)    
  optimizer.apply_gradients(zip(gradients, transformer.trainable_variables))
  
  train_loss(loss)
  train_accuracy(tar_real, predictions)

Portuguese is used as the input language and English is the target language.

for epoch in range(EPOCHS):
  start = time.time()
  
  train_loss.reset_states()
  train_accuracy.reset_states()
  
  # inp -> portuguese, tar -> english
  for (batch, (inp, tar)) in enumerate(train_dataset):
    train_step(inp, tar)
    
    if batch % 500 == 0:
      print ('Epoch {} Batch {} Loss {:.4f} Accuracy {:.4f}'.format(
          epoch + 1, batch, train_loss.result(), train_accuracy.result()))
      
  if (epoch + 1) % 5 == 0:
    ckpt_save_path = ckpt_manager.save()
    print ('Saving checkpoint for epoch {} at {}'.format(epoch+1,
                                                         ckpt_save_path))
    
  print ('Epoch {} Loss {:.4f} Accuracy {:.4f}'.format(epoch + 1, 
                                                train_loss.result(), 
                                                train_accuracy.result()))

  print ('Time taken for 1 epoch: {} secs\n'.format(time.time() - start))
Epoch 1 Batch 0 Loss 3.9991 Accuracy 0.0000
Epoch 1 Batch 500 Loss 3.6032 Accuracy 0.0401
Epoch 1 Loss 3.3757 Accuracy 0.0522
Time taken for 1 epoch: 910.9895820617676 secs

Epoch 2 Batch 0 Loss 2.4494 Accuracy 0.1020
Epoch 2 Batch 500 Loss 2.4283 Accuracy 0.1183
Epoch 2 Loss 2.3771 Accuracy 0.1235
Time taken for 1 epoch: 832.7596199512482 secs

Epoch 3 Batch 0 Loss 2.0613 Accuracy 0.1357
Epoch 3 Batch 500 Loss 2.1266 Accuracy 0.1490
Epoch 3 Loss 2.0910 Accuracy 0.1535
Time taken for 1 epoch: 64.02932643890381 secs

Epoch 4 Batch 0 Loss 1.8104 Accuracy 0.1637
Epoch 4 Batch 500 Loss 1.8686 Accuracy 0.1805
Epoch 4 Loss 1.8373 Accuracy 0.1847
Time taken for 1 epoch: 64.04561138153076 secs

Epoch 5 Batch 0 Loss 1.5748 Accuracy 0.1974
Epoch 5 Batch 500 Loss 1.6517 Accuracy 0.2062
Saving checkpoint for epoch 5 at ./checkpoints/train/ckpt-1
Epoch 5 Loss 1.6252 Accuracy 0.2095
Time taken for 1 epoch: 66.38142800331116 secs

Epoch 6 Batch 0 Loss 1.3695 Accuracy 0.2122
Epoch 6 Batch 500 Loss 1.4655 Accuracy 0.2263
Epoch 6 Loss 1.4420 Accuracy 0.2291
Time taken for 1 epoch: 64.81302809715271 secs

Epoch 7 Batch 0 Loss 1.2018 Accuracy 0.2245
Epoch 7 Batch 500 Loss 1.2865 Accuracy 0.2464
Epoch 7 Loss 1.2620 Accuracy 0.2497
Time taken for 1 epoch: 66.47986888885498 secs

Epoch 8 Batch 0 Loss 1.0526 Accuracy 0.2492
Epoch 8 Batch 500 Loss 1.1341 Accuracy 0.2647
Epoch 8 Loss 1.1174 Accuracy 0.2670
Time taken for 1 epoch: 65.33747363090515 secs

Epoch 9 Batch 0 Loss 0.9072 Accuracy 0.2644
Epoch 9 Batch 500 Loss 1.0208 Accuracy 0.2791
Epoch 9 Loss 1.0096 Accuracy 0.2807
Time taken for 1 epoch: 65.66915273666382 secs

Epoch 10 Batch 0 Loss 0.8419 Accuracy 0.2726
Epoch 10 Batch 500 Loss 0.9354 Accuracy 0.2896
Saving checkpoint for epoch 10 at ./checkpoints/train/ckpt-2
Epoch 10 Loss 0.9279 Accuracy 0.2909
Time taken for 1 epoch: 65.89093470573425 secs

Epoch 11 Batch 0 Loss 0.7948 Accuracy 0.2812
Epoch 11 Batch 500 Loss 0.8674 Accuracy 0.2986
Epoch 11 Loss 0.8616 Accuracy 0.2997
Time taken for 1 epoch: 66.40026807785034 secs

Epoch 12 Batch 0 Loss 0.7166 Accuracy 0.2936
Epoch 12 Batch 500 Loss 0.8128 Accuracy 0.3062
Epoch 12 Loss 0.8075 Accuracy 0.3072
Time taken for 1 epoch: 65.21522116661072 secs

Epoch 13 Batch 0 Loss 0.6670 Accuracy 0.2981
Epoch 13 Batch 500 Loss 0.7676 Accuracy 0.3120
Epoch 13 Loss 0.7627 Accuracy 0.3131
Time taken for 1 epoch: 65.45641231536865 secs

Epoch 14 Batch 0 Loss 0.6333 Accuracy 0.3084
Epoch 14 Batch 500 Loss 0.7283 Accuracy 0.3177
Epoch 14 Loss 0.7246 Accuracy 0.3184
Time taken for 1 epoch: 66.96323156356812 secs

Epoch 15 Batch 0 Loss 0.6053 Accuracy 0.3096
Epoch 15 Batch 500 Loss 0.6915 Accuracy 0.3231
Saving checkpoint for epoch 15 at ./checkpoints/train/ckpt-3
Epoch 15 Loss 0.6895 Accuracy 0.3235
Time taken for 1 epoch: 64.72928357124329 secs

Epoch 16 Batch 0 Loss 0.5931 Accuracy 0.3076
Epoch 16 Batch 500 Loss 0.6605 Accuracy 0.3277
Epoch 16 Loss 0.6585 Accuracy 0.3281
Time taken for 1 epoch: 66.47149324417114 secs

Epoch 17 Batch 0 Loss 0.5575 Accuracy 0.3109
Epoch 17 Batch 500 Loss 0.6301 Accuracy 0.3327
Epoch 17 Loss 0.6287 Accuracy 0.3331
Time taken for 1 epoch: 64.4195568561554 secs

Epoch 18 Batch 0 Loss 0.5340 Accuracy 0.3261
Epoch 18 Batch 500 Loss 0.6066 Accuracy 0.3360
Epoch 18 Loss 0.6048 Accuracy 0.3366
Time taken for 1 epoch: 64.8441870212555 secs

Epoch 19 Batch 0 Loss 0.5088 Accuracy 0.3215
Epoch 19 Batch 500 Loss 0.5837 Accuracy 0.3394
Epoch 19 Loss 0.5817 Accuracy 0.3400
Time taken for 1 epoch: 66.19496250152588 secs

Epoch 20 Batch 0 Loss 0.4719 Accuracy 0.3359
Epoch 20 Batch 500 Loss 0.5618 Accuracy 0.3427
Saving checkpoint for epoch 20 at ./checkpoints/train/ckpt-4
Epoch 20 Loss 0.5607 Accuracy 0.3432
Time taken for 1 epoch: 65.4521656036377 secs

Evaluate

The following steps are used for evaluation:

  • Encode the input sentence using the Portuguese tokenizer (tokenizer_pt). Moreover, add the start and end token so the input is equivalent to what the model is trained with. This is the encoder input.
  • The decoder input is the start token == tokenizer_en.vocab_size.
  • Calculate the padding masks and the look ahead masks.
  • The decoder then outputs the predictions by looking at the encoder output and its own output (self-attention).
  • Select the last word and calculate the argmax of that.
  • Concatentate the predicted word to the decoder input as pass it to the decoder.
  • In this approach, the decoder predicts the next word based on the previous words it predicted.
def evaluate(inp_sentence):
  start_token = [tokenizer_pt.vocab_size]
  end_token = [tokenizer_pt.vocab_size + 1]
  
  # inp sentence is portuguese, hence adding the start and end token
  inp_sentence = start_token + tokenizer_pt.encode(inp_sentence) + end_token
  encoder_input = tf.expand_dims(inp_sentence, 0)
  
  # as the target is english, the first word to the transformer should be the
  # english start token.
  decoder_input = [tokenizer_en.vocab_size]
  output = tf.expand_dims(decoder_input, 0)
    
  for i in range(MAX_LENGTH):
    enc_padding_mask, combined_mask, dec_padding_mask = create_masks(
        encoder_input, output)
  
    # predictions.shape == (batch_size, seq_len, vocab_size)
    predictions, attention_weights = transformer(encoder_input, 
                                                 output,
                                                 False,
                                                 enc_padding_mask,
                                                 combined_mask,
                                                 dec_padding_mask)
    
    # select the last word from the seq_len dimension
    predictions = predictions[: ,-1:, :]  # (batch_size, 1, vocab_size)

    predicted_id = tf.cast(tf.argmax(predictions, axis=-1), tf.int32)
    
    # return the result if the predicted_id is equal to the end token
    if tf.equal(predicted_id, tokenizer_en.vocab_size+1):
      return tf.squeeze(output, axis=0), attention_weights
    
    # concatentate the predicted_id to the output which is given to the decoder
    # as its input.
    output = tf.concat([output, predicted_id], axis=-1)

  return tf.squeeze(output, axis=0), attention_weights
def plot_attention_weights(attention, sentence, result, layer):
  fig = plt.figure(figsize=(16, 8))
  
  sentence = tokenizer_pt.encode(sentence)
  
  attention = tf.squeeze(attention[layer], axis=0)
  
  for head in range(attention.shape[0]):
    ax = fig.add_subplot(2, 4, head+1)
    
    # plot the attention weights
    ax.matshow(attention[head][:-1, :], cmap='viridis')

    fontdict = {'fontsize': 10}
    
    ax.set_xticks(range(len(sentence)+2))
    ax.set_yticks(range(len(result)))
    
    ax.set_ylim(len(result)-1.5, -0.5)
        
    ax.set_xticklabels(
        ['<start>']+[tokenizer_pt.decode([i]) for i in sentence]+['<end>'], 
        fontdict=fontdict, rotation=90)
    
    ax.set_yticklabels([tokenizer_en.decode([i]) for i in result 
                        if i < tokenizer_en.vocab_size], 
                       fontdict=fontdict)
    
    ax.set_xlabel('Head {}'.format(head+1))
  
  plt.tight_layout()
  plt.show()
def translate(sentence, plot=''):
  result, attention_weights = evaluate(sentence)
  
  predicted_sentence = tokenizer_en.decode([i for i in result 
                                            if i < tokenizer_en.vocab_size])  

  print('Input: {}'.format(sentence))
  print('Predicted translation: {}'.format(predicted_sentence))
  
  if plot:
    plot_attention_weights(attention_weights, sentence, result, plot)
translate("este é um problema que temos que resolver.")
print ("Real translation: this is a problem we have to solve .")
Input: este é um problema que temos que resolver.
Predicted translation: this is a problem that we have to deal with .
Real translation: this is a problem we have to solve .
translate("os meus vizinhos ouviram sobre esta ideia.")
print ("Real translation: and my neighboring homes heard about this idea .")
Input: os meus vizinhos ouviram sobre esta ideia.
Predicted translation: my neighbors heard about this idea of the mathematics .
Real translation: and my neighboring homes heard about this idea .
translate("vou então muito rapidamente partilhar convosco algumas histórias de algumas coisas mágicas que aconteceram.")
print ("Real translation: so i 'll just share with you some stories very quickly of some magical things that have happened .")
Input: vou então muito rapidamente partilhar convosco algumas histórias de algumas coisas mágicas que aconteceram.
Predicted translation: i 'm going to take a lot of share with you some stories of some magic things that happened .
Real translation: so i 'll just share with you some stories very quickly of some magical things that have happened .

You can pass different layers and attention blocks of the decoder to the plot parameter.

translate("este é o primeiro livro que eu fiz.", plot='decoder_layer4_block2')
print ("Real translation: this is the first book i've ever done.")
Input: este é o primeiro livro que eu fiz.
Predicted translation: this is the first book that i did .

png

Real translation: this is the first book i've ever done.

Summary

In this tutorial, you learned about positional encoding, multi-head attention, the importance of masking and how to create a transformer.

Try using a different dataset to train the transformer. You can also create the base transformer or transformer XL by changing the hyperparameters above. You can also use the layers defined here to create BERT and train state of the art models. Futhermore, you can implement beam search to get better predictions.