Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge

Transformer model for language understanding

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

This tutorial trains a Transformer model to translate a Portuguese to English dataset. 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 makes 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

Setup

pip install tensorflow_datasets
pip install -U tensorflow-text
import collections
import logging
import os
import pathlib
import re
import string
import sys
import time

import numpy as np
import matplotlib.pyplot as plt

import tensorflow_datasets as tfds
import tensorflow_text as text
import tensorflow as tf
logging.getLogger('tensorflow').setLevel(logging.ERROR)  # suppress warnings

Download the Dataset

Use TensorFlow datasets to load the Portuguese-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']

The tf.data.Dataset object returned by TensorFlow datasets yields pairs of text examples:

for pt_examples, en_examples in train_examples.batch(3).take(1):
  for pt in pt_examples.numpy():
    print(pt.decode('utf-8'))

  print()

  for en in en_examples.numpy():
    print(en.decode('utf-8'))
e quando melhoramos a procura , tiramos a única vantagem da impressão , que é a serendipidade .
mas e se estes fatores fossem ativos ?
mas eles não tinham a curiosidade de me testar .

and when you improve searchability , you actually take away the one advantage of print , which is serendipity .
but what if it were active ?
but they did n't test for curiosity .

Text tokenization & detokenization

You can't train a model directly on text. The text needs to be converted to some numeric representation first. Typically, you convert the text to sequences of token IDs, which are used as indices into an embedding.

One popular implementation is demonstrated in the Subword tokenizer tutorial builds subword tokenizers (text.BertTokenizer) optimized for this dataset and exports them in a saved_model.

Download and unzip and import the saved_model:

model_name = "ted_hrlr_translate_pt_en_converter"
tf.keras.utils.get_file(
    f"{model_name}.zip",
    f"https://storage.googleapis.com/download.tensorflow.org/models/{model_name}.zip",
    cache_dir='.', cache_subdir='', extract=True
)
Downloading data from https://storage.googleapis.com/download.tensorflow.org/models/ted_hrlr_translate_pt_en_converter.zip
188416/184801 [==============================] - 0s 0us/step
196608/184801 [===============================] - 0s 0us/step
'./ted_hrlr_translate_pt_en_converter.zip'
tokenizers = tf.saved_model.load(model_name)

The tf.saved_model contains two text tokenizers, one for English and one for Portuguese. Both have the same methods:

[item for item in dir(tokenizers.en) if not item.startswith('_')]
['detokenize',
 'get_reserved_tokens',
 'get_vocab_path',
 'get_vocab_size',
 'lookup',
 'tokenize',
 'tokenizer',
 'vocab']

The tokenize method converts a batch of strings to a padded-batch of token IDs. This method splits punctuation, lowercases and unicode-normalizes the input before tokenizing. That standardization is not visible here because the input data is already standardized.

for en in en_examples.numpy():
  print(en.decode('utf-8'))
and when you improve searchability , you actually take away the one advantage of print , which is serendipity .
but what if it were active ?
but they did n't test for curiosity .
encoded = tokenizers.en.tokenize(en_examples)

for row in encoded.to_list():
  print(row)
[2, 72, 117, 79, 1259, 1491, 2362, 13, 79, 150, 184, 311, 71, 103, 2308, 74, 2679, 13, 148, 80, 55, 4840, 1434, 2423, 540, 15, 3]
[2, 87, 90, 107, 76, 129, 1852, 30, 3]
[2, 87, 83, 149, 50, 9, 56, 664, 85, 2512, 15, 3]

The detokenize method attempts to convert these token IDs back to human readable text:

round_trip = tokenizers.en.detokenize(encoded)
for line in round_trip.numpy():
  print(line.decode('utf-8'))
and when you improve searchability , you actually take away the one advantage of print , which is serendipity .
but what if it were active ?
but they did n ' t test for curiosity .

The lower level lookup method converts from token-IDs to token text:

tokens = tokenizers.en.lookup(encoded)
tokens
<tf.RaggedTensor [[b'[START]', b'and', b'when', b'you', b'improve', b'search', b'##ability', b',', b'you', b'actually', b'take', b'away', b'the', b'one', b'advantage', b'of', b'print', b',', b'which', b'is', b's', b'##ere', b'##nd', b'##ip', b'##ity', b'.', b'[END]'], [b'[START]', b'but', b'what', b'if', b'it', b'were', b'active', b'?', b'[END]'], [b'[START]', b'but', b'they', b'did', b'n', b"'", b't', b'test', b'for', b'curiosity', b'.', b'[END]']]>

Here you can see the "subword" aspect of the tokenizers. The word "searchability" is decomposed into "search ##ability" and the word "serendipity" into "s ##ere ##nd ##ip ##ity"

Setup input pipeline

To build an input pipeline suitable for training you'll apply some transformations to the dataset.

This function will be used to encode the batches of raw text:

def tokenize_pairs(pt, en):
    pt = tokenizers.pt.tokenize(pt)
    # Convert from ragged to dense, padding with zeros.
    pt = pt.to_tensor()

    en = tokenizers.en.tokenize(en)
    # Convert from ragged to dense, padding with zeros.
    en = en.to_tensor()
    return pt, en

Here's a simple input pipeline that processes, shuffles and batches the data:

BUFFER_SIZE = 20000
BATCH_SIZE = 64
def make_batches(ds):
  return (
      ds
      .cache()
      .shuffle(BUFFER_SIZE)
      .batch(BATCH_SIZE)
      .map(tokenize_pairs, num_parallel_calls=tf.data.AUTOTUNE)
      .prefetch(tf.data.AUTOTUNE))


train_batches = make_batches(train_examples)
val_batches = make_batches(val_examples)

Positional encoding

Attention layers see their input as a set of vectors, with no sequential order. This model also doesn't contain any recurrent or convolutional layers. Because of this a "positional encoding" is added to give the model some information about the relative position of the tokens 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 tokens in a sentence. So after adding the positional encoding, tokens will be closer to each other based on the similarity of their meaning and their position in the sentence, in the d-dimensional space.

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
  angle_rads[:, 0::2] = np.sin(angle_rads[:, 0::2])

  # apply cos to odd indices in the array; 2i+1
  angle_rads[:, 1::2] = np.cos(angle_rads[:, 1::2])

  pos_encoding = angle_rads[np.newaxis, ...]

  return tf.cast(pos_encoding, dtype=tf.float32)
n, d = 2048, 512
pos_encoding = positional_encoding(n, d)
print(pos_encoding.shape)
pos_encoding = pos_encoding[0]

# Juggle the dimensions for the plot
pos_encoding = tf.reshape(pos_encoding, (n, d//2, 2))
pos_encoding = tf.transpose(pos_encoding, (2, 1, 0))
pos_encoding = tf.reshape(pos_encoding, (d, n))

plt.pcolormesh(pos_encoding, cmap='RdBu')
plt.ylabel('Depth')
plt.xlabel('Position')
plt.colorbar()
plt.show()
(1, 2048, 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 to 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: 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 token, only the first and second token will be used. Similarly to predict the fourth token, only the first, second and the third tokens 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: 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:

\[\Large{Attention(Q, K, V) = softmax_k\left(\frac{QK^T}{\sqrt{d_k} }\right) V} \]

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. So the square root of dk is used for scaling, so you get a consistent variance regardless of the value of dk. If the variance is too low the output may be too flat to optimize effectively. If the variance is too high the softmax may saturate at initialization making it difficult to learn.

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_q, 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 tokens you want to focus on are kept as-is and the irrelevant tokens 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.
  • Scaled dot-product attention.
  • Final linear layer.

Each multi-head attention block gets three inputs; Q (query), K (key), V (value). These are put through linear (Dense) layers before the multi-head attention function.

In the diagram above (K,Q,V) are passed through sepearte linear (Dense) layers for each attention head. For simplicity/efficiency the code below implements this using a single dense layer with num_heads times as many outputs. The output is rearranged to a shape of (batch, num_heads, ...) before applying the attention function.

The scaled_dot_product_attention function defined above is applied in a single call, 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 from different representation subspaces at different positions. 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_q, 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_q, num_heads, depth)

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

    output = self.dense(concat_attention)  # (batch_size, seq_len_q, 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 token in the sequence.
  • The decoder attends to 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 token 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,
               maximum_position_encoding, 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(maximum_position_encoding,
                                            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,
                         maximum_position_encoding=10000)
temp_input = tf.random.uniform((64, 62), dtype=tf.int64, minval=0, maxval=200)

sample_encoder_output = sample_encoder(temp_input, 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,
               maximum_position_encoding, 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(maximum_position_encoding, 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[f'decoder_layer{i+1}_block1'] = block1
      attention_weights[f'decoder_layer{i+1}_block2'] = 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,
                         maximum_position_encoding=5000)
temp_input = tf.random.uniform((64, 26), dtype=tf.int64, minval=0, maxval=200)

output, attn = sample_decoder(temp_input,
                              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, pe_input, pe_target, rate=0.1):
    super().__init__()
    self.encoder = Encoder(num_layers, d_model, num_heads, dff,
                             input_vocab_size, pe_input, rate)

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

    self.final_layer = tf.keras.layers.Dense(target_vocab_size)

  def call(self, inputs, training):
    # Keras models prefer if you pass all your inputs in the first argument
    inp, tar = inputs

    enc_padding_mask, look_ahead_mask, dec_padding_mask = self.create_masks(inp, tar)

    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

  def create_masks(self, 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)
    look_ahead_mask = tf.maximum(dec_target_padding_mask, look_ahead_mask)

    return enc_padding_mask, look_ahead_mask, dec_padding_mask
sample_transformer = Transformer(
    num_layers=2, d_model=512, num_heads=8, dff=2048,
    input_vocab_size=8500, target_vocab_size=8000,
    pe_input=10000, pe_target=6000)

temp_input = tf.random.uniform((64, 38), dtype=tf.int64, minval=0, maxval=200)
temp_target = tf.random.uniform((64, 36), dtype=tf.int64, minval=0, maxval=200)

fn_out, _ = sample_transformer([temp_input, temp_target], training=False)

fn_out.shape  # (batch_size, tar_seq_len, target_vocab_size)
TensorShape([64, 36, 8000])

Set hyperparameters

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

The base model described in the paper used: num_layers=6, d_model=512, dff=2048.

num_layers = 4
d_model = 128
dff = 512
num_heads = 8
dropout_rate = 0.1

Optimizer

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

\[\Large{lrate = d_{model}^{-0.5} * \min(step{\_}num^{-0.5}, step{\_}num \cdot warmup{\_}steps^{-1.5})}\]

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_sum(loss_)/tf.reduce_sum(mask)


def accuracy_function(real, pred):
  accuracies = tf.equal(real, tf.argmax(pred, axis=2))

  mask = tf.math.logical_not(tf.math.equal(real, 0))
  accuracies = tf.math.logical_and(mask, accuracies)

  accuracies = tf.cast(accuracies, dtype=tf.float32)
  mask = tf.cast(mask, dtype=tf.float32)
  return tf.reduce_sum(accuracies)/tf.reduce_sum(mask)
train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy = tf.keras.metrics.Mean(name='train_accuracy')

Training and checkpointing

transformer = Transformer(
    num_layers=num_layers,
    d_model=d_model,
    num_heads=num_heads,
    dff=dff,
    input_vocab_size=tokenizers.pt.get_vocab_size().numpy(),
    target_vocab_size=tokenizers.en.get_vocab_size().numpy(),
    pe_input=1000,
    pe_target=1000,
    rate=dropout_rate)

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 token, self-attention allows it to look at the previous tokens in the input sequence to better predict the next token.

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

EPOCHS = 20
# The @tf.function trace-compiles train_step into a TF graph for faster
# execution. The function specializes to the precise shape of the argument
# tensors. To avoid re-tracing due to the variable sequence lengths or variable
# batch sizes (the last batch is smaller), use input_signature to specify
# more generic shapes.

train_step_signature = [
    tf.TensorSpec(shape=(None, None), dtype=tf.int64),
    tf.TensorSpec(shape=(None, None), dtype=tf.int64),
]


@tf.function(input_signature=train_step_signature)
def train_step(inp, tar):
  tar_inp = tar[:, :-1]
  tar_real = tar[:, 1:]

  with tf.GradientTape() as tape:
    predictions, _ = transformer([inp, tar_inp],
                                 training = True)
    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(accuracy_function(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_batches):
    train_step(inp, tar)

    if batch % 50 == 0:
      print(f'Epoch {epoch + 1} Batch {batch} Loss {train_loss.result():.4f} Accuracy {train_accuracy.result():.4f}')

  if (epoch + 1) % 5 == 0:
    ckpt_save_path = ckpt_manager.save()
    print(f'Saving checkpoint for epoch {epoch+1} at {ckpt_save_path}')

  print(f'Epoch {epoch + 1} Loss {train_loss.result():.4f} Accuracy {train_accuracy.result():.4f}')

  print(f'Time taken for 1 epoch: {time.time() - start:.2f} secs\n')
Epoch 1 Batch 0 Loss 8.8389 Accuracy 0.0000
Epoch 1 Batch 50 Loss 8.7850 Accuracy 0.0015
Epoch 1 Batch 100 Loss 8.6954 Accuracy 0.0258
Epoch 1 Batch 150 Loss 8.5875 Accuracy 0.0361
Epoch 1 Batch 200 Loss 8.4497 Accuracy 0.0396
Epoch 1 Batch 250 Loss 8.2832 Accuracy 0.0412
Epoch 1 Batch 300 Loss 8.0957 Accuracy 0.0424
Epoch 1 Batch 350 Loss 7.9001 Accuracy 0.0462
Epoch 1 Batch 400 Loss 7.7106 Accuracy 0.0530
Epoch 1 Batch 450 Loss 7.5449 Accuracy 0.0608
Epoch 1 Batch 500 Loss 7.3965 Accuracy 0.0686
Epoch 1 Batch 550 Loss 7.2629 Accuracy 0.0764
Epoch 1 Batch 600 Loss 7.1374 Accuracy 0.0841
Epoch 1 Batch 650 Loss 7.0225 Accuracy 0.0915
Epoch 1 Batch 700 Loss 6.9148 Accuracy 0.0987
Epoch 1 Batch 750 Loss 6.8137 Accuracy 0.1055
Epoch 1 Batch 800 Loss 6.7228 Accuracy 0.1115
Epoch 1 Loss 6.7072 Accuracy 0.1126
Time taken for 1 epoch: 62.48 secs

Epoch 2 Batch 0 Loss 5.3320 Accuracy 0.2099
Epoch 2 Batch 50 Loss 5.2248 Accuracy 0.2119
Epoch 2 Batch 100 Loss 5.1947 Accuracy 0.2152
Epoch 2 Batch 150 Loss 5.1643 Accuracy 0.2191
Epoch 2 Batch 200 Loss 5.1383 Accuracy 0.2224
Epoch 2 Batch 250 Loss 5.1123 Accuracy 0.2254
Epoch 2 Batch 300 Loss 5.0864 Accuracy 0.2277
Epoch 2 Batch 350 Loss 5.0697 Accuracy 0.2291
Epoch 2 Batch 400 Loss 5.0487 Accuracy 0.2313
Epoch 2 Batch 450 Loss 5.0272 Accuracy 0.2331
Epoch 2 Batch 500 Loss 5.0060 Accuracy 0.2349
Epoch 2 Batch 550 Loss 4.9882 Accuracy 0.2365
Epoch 2 Batch 600 Loss 4.9677 Accuracy 0.2382
Epoch 2 Batch 650 Loss 4.9496 Accuracy 0.2400
Epoch 2 Batch 700 Loss 4.9314 Accuracy 0.2417
Epoch 2 Batch 750 Loss 4.9172 Accuracy 0.2429
Epoch 2 Batch 800 Loss 4.9005 Accuracy 0.2444
Epoch 2 Loss 4.8976 Accuracy 0.2447
Time taken for 1 epoch: 47.81 secs

Epoch 3 Batch 0 Loss 4.7853 Accuracy 0.2318
Epoch 3 Batch 50 Loss 4.6041 Accuracy 0.2673
Epoch 3 Batch 100 Loss 4.5869 Accuracy 0.2693
Epoch 3 Batch 150 Loss 4.5681 Accuracy 0.2723
Epoch 3 Batch 200 Loss 4.5502 Accuracy 0.2748
Epoch 3 Batch 250 Loss 4.5433 Accuracy 0.2755
Epoch 3 Batch 300 Loss 4.5279 Accuracy 0.2771
Epoch 3 Batch 350 Loss 4.5123 Accuracy 0.2792
Epoch 3 Batch 400 Loss 4.5001 Accuracy 0.2807
Epoch 3 Batch 450 Loss 4.4849 Accuracy 0.2825
Epoch 3 Batch 500 Loss 4.4699 Accuracy 0.2842
Epoch 3 Batch 550 Loss 4.4543 Accuracy 0.2861
Epoch 3 Batch 600 Loss 4.4397 Accuracy 0.2879
Epoch 3 Batch 650 Loss 4.4260 Accuracy 0.2896
Epoch 3 Batch 700 Loss 4.4098 Accuracy 0.2918
Epoch 3 Batch 750 Loss 4.3934 Accuracy 0.2938
Epoch 3 Batch 800 Loss 4.3774 Accuracy 0.2958
Epoch 3 Loss 4.3743 Accuracy 0.2961
Time taken for 1 epoch: 47.78 secs

Epoch 4 Batch 0 Loss 4.1607 Accuracy 0.3198
Epoch 4 Batch 50 Loss 4.0206 Accuracy 0.3366
Epoch 4 Batch 100 Loss 4.0181 Accuracy 0.3379
Epoch 4 Batch 150 Loss 4.0005 Accuracy 0.3390
Epoch 4 Batch 200 Loss 3.9809 Accuracy 0.3420
Epoch 4 Batch 250 Loss 3.9716 Accuracy 0.3436
Epoch 4 Batch 300 Loss 3.9561 Accuracy 0.3453
Epoch 4 Batch 350 Loss 3.9396 Accuracy 0.3476
Epoch 4 Batch 400 Loss 3.9287 Accuracy 0.3492
Epoch 4 Batch 450 Loss 3.9099 Accuracy 0.3516
Epoch 4 Batch 500 Loss 3.8948 Accuracy 0.3534
Epoch 4 Batch 550 Loss 3.8796 Accuracy 0.3554
Epoch 4 Batch 600 Loss 3.8647 Accuracy 0.3573
Epoch 4 Batch 650 Loss 3.8496 Accuracy 0.3593
Epoch 4 Batch 700 Loss 3.8345 Accuracy 0.3613
Epoch 4 Batch 750 Loss 3.8216 Accuracy 0.3631
Epoch 4 Batch 800 Loss 3.8113 Accuracy 0.3645
Epoch 4 Loss 3.8084 Accuracy 0.3649
Time taken for 1 epoch: 47.74 secs

Epoch 5 Batch 0 Loss 3.7892 Accuracy 0.3515
Epoch 5 Batch 50 Loss 3.5277 Accuracy 0.3971
Epoch 5 Batch 100 Loss 3.4909 Accuracy 0.4026
Epoch 5 Batch 150 Loss 3.4709 Accuracy 0.4059
Epoch 5 Batch 200 Loss 3.4637 Accuracy 0.4068
Epoch 5 Batch 250 Loss 3.4614 Accuracy 0.4071
Epoch 5 Batch 300 Loss 3.4471 Accuracy 0.4093
Epoch 5 Batch 350 Loss 3.4360 Accuracy 0.4102
Epoch 5 Batch 400 Loss 3.4277 Accuracy 0.4112
Epoch 5 Batch 450 Loss 3.4186 Accuracy 0.4124
Epoch 5 Batch 500 Loss 3.4127 Accuracy 0.4133
Epoch 5 Batch 550 Loss 3.4023 Accuracy 0.4145
Epoch 5 Batch 600 Loss 3.3931 Accuracy 0.4154
Epoch 5 Batch 650 Loss 3.3821 Accuracy 0.4169
Epoch 5 Batch 700 Loss 3.3719 Accuracy 0.4183
Epoch 5 Batch 750 Loss 3.3632 Accuracy 0.4195
Epoch 5 Batch 800 Loss 3.3550 Accuracy 0.4205
Saving checkpoint for epoch 5 at ./checkpoints/train/ckpt-1
Epoch 5 Loss 3.3540 Accuracy 0.4206
Time taken for 1 epoch: 47.47 secs

Epoch 6 Batch 0 Loss 3.5078 Accuracy 0.3912
Epoch 6 Batch 50 Loss 3.1028 Accuracy 0.4487
Epoch 6 Batch 100 Loss 3.0855 Accuracy 0.4503
Epoch 6 Batch 150 Loss 3.0784 Accuracy 0.4515
Epoch 6 Batch 200 Loss 3.0664 Accuracy 0.4534
Epoch 6 Batch 250 Loss 3.0621 Accuracy 0.4537
Epoch 6 Batch 300 Loss 3.0512 Accuracy 0.4556
Epoch 6 Batch 350 Loss 3.0368 Accuracy 0.4580
Epoch 6 Batch 400 Loss 3.0283 Accuracy 0.4592
Epoch 6 Batch 450 Loss 3.0176 Accuracy 0.4608
Epoch 6 Batch 500 Loss 3.0083 Accuracy 0.4623
Epoch 6 Batch 550 Loss 2.9966 Accuracy 0.4639
Epoch 6 Batch 600 Loss 2.9871 Accuracy 0.4652
Epoch 6 Batch 650 Loss 2.9777 Accuracy 0.4664
Epoch 6 Batch 700 Loss 2.9722 Accuracy 0.4673
Epoch 6 Batch 750 Loss 2.9650 Accuracy 0.4684
Epoch 6 Batch 800 Loss 2.9562 Accuracy 0.4697
Epoch 6 Loss 2.9548 Accuracy 0.4698
Time taken for 1 epoch: 47.10 secs

Epoch 7 Batch 0 Loss 2.7935 Accuracy 0.4985
Epoch 7 Batch 50 Loss 2.6880 Accuracy 0.5017
Epoch 7 Batch 100 Loss 2.6676 Accuracy 0.5053
Epoch 7 Batch 150 Loss 2.6658 Accuracy 0.5062
Epoch 7 Batch 200 Loss 2.6631 Accuracy 0.5068
Epoch 7 Batch 250 Loss 2.6634 Accuracy 0.5069
Epoch 7 Batch 300 Loss 2.6575 Accuracy 0.5078
Epoch 7 Batch 350 Loss 2.6485 Accuracy 0.5088
Epoch 7 Batch 400 Loss 2.6472 Accuracy 0.5092
Epoch 7 Batch 450 Loss 2.6432 Accuracy 0.5097
Epoch 7 Batch 500 Loss 2.6398 Accuracy 0.5103
Epoch 7 Batch 550 Loss 2.6372 Accuracy 0.5111
Epoch 7 Batch 600 Loss 2.6340 Accuracy 0.5115
Epoch 7 Batch 650 Loss 2.6292 Accuracy 0.5123
Epoch 7 Batch 700 Loss 2.6266 Accuracy 0.5127
Epoch 7 Batch 750 Loss 2.6253 Accuracy 0.5129
Epoch 7 Batch 800 Loss 2.6217 Accuracy 0.5134
Epoch 7 Loss 2.6215 Accuracy 0.5134
Time taken for 1 epoch: 47.31 secs

Epoch 8 Batch 0 Loss 2.5891 Accuracy 0.5090
Epoch 8 Batch 50 Loss 2.4158 Accuracy 0.5395
Epoch 8 Batch 100 Loss 2.4099 Accuracy 0.5399
Epoch 8 Batch 150 Loss 2.4026 Accuracy 0.5408
Epoch 8 Batch 200 Loss 2.4023 Accuracy 0.5409
Epoch 8 Batch 250 Loss 2.4053 Accuracy 0.5404
Epoch 8 Batch 300 Loss 2.4085 Accuracy 0.5395
Epoch 8 Batch 350 Loss 2.4022 Accuracy 0.5406
Epoch 8 Batch 400 Loss 2.4031 Accuracy 0.5409
Epoch 8 Batch 450 Loss 2.3993 Accuracy 0.5416
Epoch 8 Batch 500 Loss 2.3970 Accuracy 0.5421
Epoch 8 Batch 550 Loss 2.3945 Accuracy 0.5429
Epoch 8 Batch 600 Loss 2.3908 Accuracy 0.5433
Epoch 8 Batch 650 Loss 2.3884 Accuracy 0.5439
Epoch 8 Batch 700 Loss 2.3864 Accuracy 0.5443
Epoch 8 Batch 750 Loss 2.3860 Accuracy 0.5443
Epoch 8 Batch 800 Loss 2.3829 Accuracy 0.5448
Epoch 8 Loss 2.3815 Accuracy 0.5450
Time taken for 1 epoch: 47.15 secs

Epoch 9 Batch 0 Loss 2.4162 Accuracy 0.5382
Epoch 9 Batch 50 Loss 2.2300 Accuracy 0.5635
Epoch 9 Batch 100 Loss 2.2048 Accuracy 0.5665
Epoch 9 Batch 150 Loss 2.2009 Accuracy 0.5681
Epoch 9 Batch 200 Loss 2.2006 Accuracy 0.5687
Epoch 9 Batch 250 Loss 2.2034 Accuracy 0.5682
Epoch 9 Batch 300 Loss 2.1974 Accuracy 0.5690
Epoch 9 Batch 350 Loss 2.2068 Accuracy 0.5678
Epoch 9 Batch 400 Loss 2.2058 Accuracy 0.5682
Epoch 9 Batch 450 Loss 2.2109 Accuracy 0.5677
Epoch 9 Batch 500 Loss 2.2112 Accuracy 0.5676
Epoch 9 Batch 550 Loss 2.2080 Accuracy 0.5685
Epoch 9 Batch 600 Loss 2.2106 Accuracy 0.5680
Epoch 9 Batch 650 Loss 2.2084 Accuracy 0.5684
Epoch 9 Batch 700 Loss 2.2082 Accuracy 0.5686
Epoch 9 Batch 750 Loss 2.2089 Accuracy 0.5685
Epoch 9 Batch 800 Loss 2.2084 Accuracy 0.5687
Epoch 9 Loss 2.2083 Accuracy 0.5686
Time taken for 1 epoch: 46.95 secs

Epoch 10 Batch 0 Loss 2.2234 Accuracy 0.5604
Epoch 10 Batch 50 Loss 2.1030 Accuracy 0.5814
Epoch 10 Batch 100 Loss 2.0774 Accuracy 0.5866
Epoch 10 Batch 150 Loss 2.0655 Accuracy 0.5881
Epoch 10 Batch 200 Loss 2.0661 Accuracy 0.5880
Epoch 10 Batch 250 Loss 2.0635 Accuracy 0.5883
Epoch 10 Batch 300 Loss 2.0616 Accuracy 0.5888
Epoch 10 Batch 350 Loss 2.0636 Accuracy 0.5886
Epoch 10 Batch 400 Loss 2.0621 Accuracy 0.5888
Epoch 10 Batch 450 Loss 2.0670 Accuracy 0.5878
Epoch 10 Batch 500 Loss 2.0652 Accuracy 0.5880
Epoch 10 Batch 550 Loss 2.0686 Accuracy 0.5875
Epoch 10 Batch 600 Loss 2.0692 Accuracy 0.5875
Epoch 10 Batch 650 Loss 2.0691 Accuracy 0.5876
Epoch 10 Batch 700 Loss 2.0707 Accuracy 0.5876
Epoch 10 Batch 750 Loss 2.0704 Accuracy 0.5878
Epoch 10 Batch 800 Loss 2.0722 Accuracy 0.5876
Saving checkpoint for epoch 10 at ./checkpoints/train/ckpt-2
Epoch 10 Loss 2.0715 Accuracy 0.5879
Time taken for 1 epoch: 47.13 secs

Epoch 11 Batch 0 Loss 1.9391 Accuracy 0.6270
Epoch 11 Batch 50 Loss 1.9758 Accuracy 0.5977
Epoch 11 Batch 100 Loss 1.9564 Accuracy 0.6024
Epoch 11 Batch 150 Loss 1.9526 Accuracy 0.6037
Epoch 11 Batch 200 Loss 1.9573 Accuracy 0.6030
Epoch 11 Batch 250 Loss 1.9471 Accuracy 0.6048
Epoch 11 Batch 300 Loss 1.9459 Accuracy 0.6048
Epoch 11 Batch 350 Loss 1.9490 Accuracy 0.6042
Epoch 11 Batch 400 Loss 1.9518 Accuracy 0.6040
Epoch 11 Batch 450 Loss 1.9549 Accuracy 0.6038
Epoch 11 Batch 500 Loss 1.9543 Accuracy 0.6040
Epoch 11 Batch 550 Loss 1.9537 Accuracy 0.6043
Epoch 11 Batch 600 Loss 1.9546 Accuracy 0.6042
Epoch 11 Batch 650 Loss 1.9556 Accuracy 0.6040
Epoch 11 Batch 700 Loss 1.9582 Accuracy 0.6038
Epoch 11 Batch 750 Loss 1.9592 Accuracy 0.6037
Epoch 11 Batch 800 Loss 1.9606 Accuracy 0.6037
Epoch 11 Loss 1.9614 Accuracy 0.6036
Time taken for 1 epoch: 46.59 secs

Epoch 12 Batch 0 Loss 1.7234 Accuracy 0.6256
Epoch 12 Batch 50 Loss 1.8507 Accuracy 0.6169
Epoch 12 Batch 100 Loss 1.8460 Accuracy 0.6184
Epoch 12 Batch 150 Loss 1.8366 Accuracy 0.6207
Epoch 12 Batch 200 Loss 1.8454 Accuracy 0.6196
Epoch 12 Batch 250 Loss 1.8398 Accuracy 0.6205
Epoch 12 Batch 300 Loss 1.8441 Accuracy 0.6202
Epoch 12 Batch 350 Loss 1.8440 Accuracy 0.6199
Epoch 12 Batch 400 Loss 1.8443 Accuracy 0.6200
Epoch 12 Batch 450 Loss 1.8452 Accuracy 0.6198
Epoch 12 Batch 500 Loss 1.8499 Accuracy 0.6193
Epoch 12 Batch 550 Loss 1.8520 Accuracy 0.6190
Epoch 12 Batch 600 Loss 1.8554 Accuracy 0.6185
Epoch 12 Batch 650 Loss 1.8587 Accuracy 0.6179
Epoch 12 Batch 700 Loss 1.8610 Accuracy 0.6176
Epoch 12 Batch 750 Loss 1.8655 Accuracy 0.6170
Epoch 12 Batch 800 Loss 1.8684 Accuracy 0.6167
Epoch 12 Loss 1.8688 Accuracy 0.6167
Time taken for 1 epoch: 46.38 secs

Epoch 13 Batch 0 Loss 1.7697 Accuracy 0.6350
Epoch 13 Batch 50 Loss 1.7443 Accuracy 0.6337
Epoch 13 Batch 100 Loss 1.7422 Accuracy 0.6352
Epoch 13 Batch 150 Loss 1.7497 Accuracy 0.6331
Epoch 13 Batch 200 Loss 1.7603 Accuracy 0.6315
Epoch 13 Batch 250 Loss 1.7623 Accuracy 0.6316
Epoch 13 Batch 300 Loss 1.7656 Accuracy 0.6315
Epoch 13 Batch 350 Loss 1.7669 Accuracy 0.6312
Epoch 13 Batch 400 Loss 1.7683 Accuracy 0.6310
Epoch 13 Batch 450 Loss 1.7758 Accuracy 0.6298
Epoch 13 Batch 500 Loss 1.7758 Accuracy 0.6300
Epoch 13 Batch 550 Loss 1.7784 Accuracy 0.6296
Epoch 13 Batch 600 Loss 1.7804 Accuracy 0.6294
Epoch 13 Batch 650 Loss 1.7825 Accuracy 0.6291
Epoch 13 Batch 700 Loss 1.7867 Accuracy 0.6286
Epoch 13 Batch 750 Loss 1.7878 Accuracy 0.6284
Epoch 13 Batch 800 Loss 1.7913 Accuracy 0.6280
Epoch 13 Loss 1.7916 Accuracy 0.6280
Time taken for 1 epoch: 46.41 secs

Epoch 14 Batch 0 Loss 1.4785 Accuracy 0.6762
Epoch 14 Batch 50 Loss 1.7171 Accuracy 0.6378
Epoch 14 Batch 100 Loss 1.6949 Accuracy 0.6424
Epoch 14 Batch 150 Loss 1.6862 Accuracy 0.6440
Epoch 14 Batch 200 Loss 1.6893 Accuracy 0.6442
Epoch 14 Batch 250 Loss 1.6927 Accuracy 0.6432
Epoch 14 Batch 300 Loss 1.6943 Accuracy 0.6428
Epoch 14 Batch 350 Loss 1.6951 Accuracy 0.6427
Epoch 14 Batch 400 Loss 1.6961 Accuracy 0.6426
Epoch 14 Batch 450 Loss 1.7003 Accuracy 0.6420
Epoch 14 Batch 500 Loss 1.7028 Accuracy 0.6415
Epoch 14 Batch 550 Loss 1.7027 Accuracy 0.6413
Epoch 14 Batch 600 Loss 1.7066 Accuracy 0.6406
Epoch 14 Batch 650 Loss 1.7110 Accuracy 0.6399
Epoch 14 Batch 700 Loss 1.7140 Accuracy 0.6395
Epoch 14 Batch 750 Loss 1.7174 Accuracy 0.6390
Epoch 14 Batch 800 Loss 1.7198 Accuracy 0.6389
Epoch 14 Loss 1.7210 Accuracy 0.6387
Time taken for 1 epoch: 46.84 secs

Epoch 15 Batch 0 Loss 1.7209 Accuracy 0.6358
Epoch 15 Batch 50 Loss 1.6188 Accuracy 0.6545
Epoch 15 Batch 100 Loss 1.6084 Accuracy 0.6567
Epoch 15 Batch 150 Loss 1.6169 Accuracy 0.6557
Epoch 15 Batch 200 Loss 1.6215 Accuracy 0.6542
Epoch 15 Batch 250 Loss 1.6268 Accuracy 0.6534
Epoch 15 Batch 300 Loss 1.6257 Accuracy 0.6539
Epoch 15 Batch 350 Loss 1.6300 Accuracy 0.6531
Epoch 15 Batch 400 Loss 1.6318 Accuracy 0.6527
Epoch 15 Batch 450 Loss 1.6360 Accuracy 0.6521
Epoch 15 Batch 500 Loss 1.6403 Accuracy 0.6515
Epoch 15 Batch 550 Loss 1.6445 Accuracy 0.6509
Epoch 15 Batch 600 Loss 1.6467 Accuracy 0.6505
Epoch 15 Batch 650 Loss 1.6501 Accuracy 0.6499
Epoch 15 Batch 700 Loss 1.6537 Accuracy 0.6495
Epoch 15 Batch 750 Loss 1.6560 Accuracy 0.6492
Epoch 15 Batch 800 Loss 1.6602 Accuracy 0.6486
Saving checkpoint for epoch 15 at ./checkpoints/train/ckpt-3
Epoch 15 Loss 1.6598 Accuracy 0.6487
Time taken for 1 epoch: 46.54 secs

Epoch 16 Batch 0 Loss 1.5373 Accuracy 0.6739
Epoch 16 Batch 50 Loss 1.5678 Accuracy 0.6614
Epoch 16 Batch 100 Loss 1.5524 Accuracy 0.6646
Epoch 16 Batch 150 Loss 1.5550 Accuracy 0.6640
Epoch 16 Batch 200 Loss 1.5612 Accuracy 0.6627
Epoch 16 Batch 250 Loss 1.5703 Accuracy 0.6608
Epoch 16 Batch 300 Loss 1.5771 Accuracy 0.6597
Epoch 16 Batch 350 Loss 1.5812 Accuracy 0.6593
Epoch 16 Batch 400 Loss 1.5846 Accuracy 0.6588
Epoch 16 Batch 450 Loss 1.5864 Accuracy 0.6587
Epoch 16 Batch 500 Loss 1.5906 Accuracy 0.6581
Epoch 16 Batch 550 Loss 1.5940 Accuracy 0.6575
Epoch 16 Batch 600 Loss 1.5938 Accuracy 0.6576
Epoch 16 Batch 650 Loss 1.5959 Accuracy 0.6575
Epoch 16 Batch 700 Loss 1.5991 Accuracy 0.6571
Epoch 16 Batch 750 Loss 1.6017 Accuracy 0.6568
Epoch 16 Batch 800 Loss 1.6035 Accuracy 0.6566
Epoch 16 Loss 1.6043 Accuracy 0.6565
Time taken for 1 epoch: 46.25 secs

Epoch 17 Batch 0 Loss 1.6442 Accuracy 0.6463
Epoch 17 Batch 50 Loss 1.5289 Accuracy 0.6670
Epoch 17 Batch 100 Loss 1.5225 Accuracy 0.6684
Epoch 17 Batch 150 Loss 1.5237 Accuracy 0.6685
Epoch 17 Batch 200 Loss 1.5301 Accuracy 0.6674
Epoch 17 Batch 250 Loss 1.5323 Accuracy 0.6672
Epoch 17 Batch 300 Loss 1.5332 Accuracy 0.6670
Epoch 17 Batch 350 Loss 1.5349 Accuracy 0.6667
Epoch 17 Batch 400 Loss 1.5377 Accuracy 0.6662
Epoch 17 Batch 450 Loss 1.5399 Accuracy 0.6660
Epoch 17 Batch 500 Loss 1.5412 Accuracy 0.6660
Epoch 17 Batch 550 Loss 1.5468 Accuracy 0.6650
Epoch 17 Batch 600 Loss 1.5474 Accuracy 0.6651
Epoch 17 Batch 650 Loss 1.5504 Accuracy 0.6647
Epoch 17 Batch 700 Loss 1.5534 Accuracy 0.6643
Epoch 17 Batch 750 Loss 1.5562 Accuracy 0.6639
Epoch 17 Batch 800 Loss 1.5583 Accuracy 0.6635
Epoch 17 Loss 1.5593 Accuracy 0.6633
Time taken for 1 epoch: 46.66 secs

Epoch 18 Batch 0 Loss 1.5130 Accuracy 0.6849
Epoch 18 Batch 50 Loss 1.4575 Accuracy 0.6796
Epoch 18 Batch 100 Loss 1.4698 Accuracy 0.6770
Epoch 18 Batch 150 Loss 1.4760 Accuracy 0.6765
Epoch 18 Batch 200 Loss 1.4783 Accuracy 0.6762
Epoch 18 Batch 250 Loss 1.4811 Accuracy 0.6758
Epoch 18 Batch 300 Loss 1.4870 Accuracy 0.6749
Epoch 18 Batch 350 Loss 1.4929 Accuracy 0.6738
Epoch 18 Batch 400 Loss 1.4945 Accuracy 0.6736
Epoch 18 Batch 450 Loss 1.4979 Accuracy 0.6730
Epoch 18 Batch 500 Loss 1.4970 Accuracy 0.6731
Epoch 18 Batch 550 Loss 1.5004 Accuracy 0.6725
Epoch 18 Batch 600 Loss 1.5034 Accuracy 0.6721
Epoch 18 Batch 650 Loss 1.5073 Accuracy 0.6715
Epoch 18 Batch 700 Loss 1.5096 Accuracy 0.6712
Epoch 18 Batch 750 Loss 1.5129 Accuracy 0.6707
Epoch 18 Batch 800 Loss 1.5164 Accuracy 0.6701
Epoch 18 Loss 1.5160 Accuracy 0.6702
Time taken for 1 epoch: 46.40 secs

Epoch 19 Batch 0 Loss 1.5139 Accuracy 0.6813
Epoch 19 Batch 50 Loss 1.4341 Accuracy 0.6832
Epoch 19 Batch 100 Loss 1.4357 Accuracy 0.6828
Epoch 19 Batch 150 Loss 1.4364 Accuracy 0.6829
Epoch 19 Batch 200 Loss 1.4439 Accuracy 0.6816
Epoch 19 Batch 250 Loss 1.4448 Accuracy 0.6816
Epoch 19 Batch 300 Loss 1.4494 Accuracy 0.6812
Epoch 19 Batch 350 Loss 1.4543 Accuracy 0.6805
Epoch 19 Batch 400 Loss 1.4543 Accuracy 0.6804
Epoch 19 Batch 450 Loss 1.4583 Accuracy 0.6795
Epoch 19 Batch 500 Loss 1.4609 Accuracy 0.6789
Epoch 19 Batch 550 Loss 1.4637 Accuracy 0.6786
Epoch 19 Batch 600 Loss 1.4641 Accuracy 0.6787
Epoch 19 Batch 650 Loss 1.4674 Accuracy 0.6780
Epoch 19 Batch 700 Loss 1.4703 Accuracy 0.6776
Epoch 19 Batch 750 Loss 1.4722 Accuracy 0.6773
Epoch 19 Batch 800 Loss 1.4760 Accuracy 0.6768
Epoch 19 Loss 1.4767 Accuracy 0.6766
Time taken for 1 epoch: 46.45 secs

Epoch 20 Batch 0 Loss 1.3065 Accuracy 0.7057
Epoch 20 Batch 50 Loss 1.3843 Accuracy 0.6909
Epoch 20 Batch 100 Loss 1.3941 Accuracy 0.6891
Epoch 20 Batch 150 Loss 1.3995 Accuracy 0.6883
Epoch 20 Batch 200 Loss 1.3997 Accuracy 0.6888
Epoch 20 Batch 250 Loss 1.4033 Accuracy 0.6884
Epoch 20 Batch 300 Loss 1.4076 Accuracy 0.6873
Epoch 20 Batch 350 Loss 1.4130 Accuracy 0.6864
Epoch 20 Batch 400 Loss 1.4144 Accuracy 0.6862
Epoch 20 Batch 450 Loss 1.4186 Accuracy 0.6853
Epoch 20 Batch 500 Loss 1.4206 Accuracy 0.6851
Epoch 20 Batch 550 Loss 1.4225 Accuracy 0.6847
Epoch 20 Batch 600 Loss 1.4253 Accuracy 0.6843
Epoch 20 Batch 650 Loss 1.4299 Accuracy 0.6835
Epoch 20 Batch 700 Loss 1.4324 Accuracy 0.6831
Epoch 20 Batch 750 Loss 1.4376 Accuracy 0.6823
Epoch 20 Batch 800 Loss 1.4405 Accuracy 0.6817
Saving checkpoint for epoch 20 at ./checkpoints/train/ckpt-4
Epoch 20 Loss 1.4410 Accuracy 0.6816
Time taken for 1 epoch: 46.87 secs

Run inference

The following steps are used for inference:

  • Encode the input sentence using the Portuguese tokenizer (tokenizers.pt). This is the encoder input.
  • The decoder input is initialized to the [START] token.
  • 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).
  • Concatenate the predicted token to the decoder input and pass it to the decoder.
  • In this approach, the decoder predicts the next token based on the previous tokens it predicted.
class Translator(tf.Module):
  def __init__(self, tokenizers, transformer):
    self.tokenizers = tokenizers
    self.transformer = transformer

  def __call__(self, sentence, max_length=20):
    # input sentence is portuguese, hence adding the start and end token
    assert isinstance(sentence, tf.Tensor)
    if len(sentence.shape) == 0:
      sentence = sentence[tf.newaxis]

    sentence = self.tokenizers.pt.tokenize(sentence).to_tensor()

    encoder_input = sentence

    # as the target is english, the first token to the transformer should be the
    # english start token.
    start_end = self.tokenizers.en.tokenize([''])[0]
    start = start_end[0][tf.newaxis]
    end = start_end[1][tf.newaxis]

    # `tf.TensorArray` is required here (instead of a python list) so that the
    # dynamic-loop can be traced by `tf.function`.
    output_array = tf.TensorArray(dtype=tf.int64, size=0, dynamic_size=True)
    output_array = output_array.write(0, start)

    for i in tf.range(max_length):
      output = tf.transpose(output_array.stack())
      predictions, _ = self.transformer([encoder_input, output], training=False)

      # select the last token from the seq_len dimension
      predictions = predictions[:, -1:, :]  # (batch_size, 1, vocab_size)

      predicted_id = tf.argmax(predictions, axis=-1)

      # concatentate the predicted_id to the output which is given to the decoder
      # as its input.
      output_array = output_array.write(i+1, predicted_id[0])

      if predicted_id == end:
        break

    output = tf.transpose(output_array.stack())
    # output.shape (1, tokens)
    text = tokenizers.en.detokenize(output)[0]  # shape: ()

    tokens = tokenizers.en.lookup(output)[0]

    # `tf.function` prevents us from using the attention_weights that were
    # calculated on the last iteration of the loop. So recalculate them outside
    # the loop.
    _, attention_weights = self.transformer([encoder_input, output[:,:-1]], training=False)

    return text, tokens, attention_weights

Create an instance of this Translator class, and try it out a few times:

translator = Translator(tokenizers, transformer)
def print_translation(sentence, tokens, ground_truth):
  print(f'{"Input:":15s}: {sentence}')
  print(f'{"Prediction":15s}: {tokens.numpy().decode("utf-8")}')
  print(f'{"Ground truth":15s}: {ground_truth}')
sentence = "este é um problema que temos que resolver."
ground_truth = "this is a problem we have to solve ."

translated_text, translated_tokens, attention_weights = translator(
    tf.constant(sentence))
print_translation(sentence, translated_text, ground_truth)
Input:         : este é um problema que temos que resolver.
Prediction     : this is a problem that we have to solve .
Ground truth   : this is a problem we have to solve .
sentence = "os meus vizinhos ouviram sobre esta ideia."
ground_truth = "and my neighboring homes heard about this idea ."

translated_text, translated_tokens, attention_weights = translator(
    tf.constant(sentence))
print_translation(sentence, translated_text, ground_truth)
Input:         : os meus vizinhos ouviram sobre esta ideia.
Prediction     : my neighbors heard about this idea .
Ground truth   : and my neighboring homes heard about this idea .
sentence = "vou então muito rapidamente partilhar convosco algumas histórias de algumas coisas mágicas que aconteceram."
ground_truth = "so i \'ll just share with you some stories very quickly of some magical things that have happened ."

translated_text, translated_tokens, attention_weights = translator(
    tf.constant(sentence))
print_translation(sentence, translated_text, ground_truth)
Input:         : vou então muito rapidamente partilhar convosco algumas histórias de algumas coisas mágicas que aconteceram.
Prediction     : so i ' m going to be very quickly share with you some of the magic stories that happened .
Ground truth   : so i 'll just share with you some stories very quickly of some magical things that have happened .

Attention plots

The Translator class returns a dictionary of attention maps you can use to visualize the internal working of the model:

sentence = "este é o primeiro livro que eu fiz."
ground_truth = "this is the first book i've ever done."

translated_text, translated_tokens, attention_weights = translator(
    tf.constant(sentence))
print_translation(sentence, translated_text, ground_truth)
Input:         : este é o primeiro livro que eu fiz.
Prediction     : this is the first book i did .
Ground truth   : this is the first book i've ever done.
def plot_attention_head(in_tokens, translated_tokens, attention):
  # The plot is of the attention when a token was generated.
  # The model didn't generate `<START>` in the output. Skip it.
  translated_tokens = translated_tokens[1:]

  ax = plt.gca()
  ax.matshow(attention)
  ax.set_xticks(range(len(in_tokens)))
  ax.set_yticks(range(len(translated_tokens)))

  labels = [label.decode('utf-8') for label in in_tokens.numpy()]
  ax.set_xticklabels(
      labels, rotation=90)

  labels = [label.decode('utf-8') for label in translated_tokens.numpy()]
  ax.set_yticklabels(labels)
head = 0
# shape: (batch=1, num_heads, seq_len_q, seq_len_k)
attention_heads = tf.squeeze(
  attention_weights['decoder_layer4_block2'], 0)
attention = attention_heads[head]
attention.shape
TensorShape([9, 11])
in_tokens = tf.convert_to_tensor([sentence])
in_tokens = tokenizers.pt.tokenize(in_tokens).to_tensor()
in_tokens = tokenizers.pt.lookup(in_tokens)[0]
in_tokens
<tf.Tensor: shape=(11,), dtype=string, numpy=
array([b'[START]', b'este', b'e', b'o', b'primeiro', b'livro', b'que',
       b'eu', b'fiz', b'.', b'[END]'], dtype=object)>
translated_tokens
<tf.Tensor: shape=(10,), dtype=string, numpy=
array([b'[START]', b'this', b'is', b'the', b'first', b'book', b'i',
       b'did', b'.', b'[END]'], dtype=object)>
plot_attention_head(in_tokens, translated_tokens, attention)

png

def plot_attention_weights(sentence, translated_tokens, attention_heads):
  in_tokens = tf.convert_to_tensor([sentence])
  in_tokens = tokenizers.pt.tokenize(in_tokens).to_tensor()
  in_tokens = tokenizers.pt.lookup(in_tokens)[0]
  in_tokens

  fig = plt.figure(figsize=(16, 8))

  for h, head in enumerate(attention_heads):
    ax = fig.add_subplot(2, 4, h+1)

    plot_attention_head(in_tokens, translated_tokens, head)

    ax.set_xlabel(f'Head {h+1}')

  plt.tight_layout()
  plt.show()
plot_attention_weights(sentence, translated_tokens,
                       attention_weights['decoder_layer4_block2'][0])

png

The model does okay on unfamiliar words. Neither "triceratops" or "encyclopedia" are in the input dataset and the model almost learns to transliterate them, even without a shared vocabulary:

sentence = "Eu li sobre triceratops na enciclopédia."
ground_truth = "I read about triceratops in the encyclopedia."

translated_text, translated_tokens, attention_weights = translator(
    tf.constant(sentence))
print_translation(sentence, translated_text, ground_truth)

plot_attention_weights(sentence, translated_tokens,
                       attention_weights['decoder_layer4_block2'][0])
Input:         : Eu li sobre triceratops na enciclopédia.
Prediction     : i read about trivalopat nairconcissus in the encyclo
Ground truth   : I read about triceratops in the encyclopedia.

png

Export

That inference model is working, so next you'll export it as a tf.saved_model.

To do that, wrap it in yet another tf.Module sub-class, this time with a tf.function on the __call__ method:

class ExportTranslator(tf.Module):
  def __init__(self, translator):
    self.translator = translator

  @tf.function(input_signature=[tf.TensorSpec(shape=[], dtype=tf.string)])
  def __call__(self, sentence):
    (result, 
     tokens,
     attention_weights) = self.translator(sentence, max_length=100)

    return result

In the above tf.function only the output sentence is returned. Thanks to the non-strict execution in tf.function any unnecessary values are never computed.

translator = ExportTranslator(translator)

Since the model is decoding the predictions using tf.argmax the predictions are deterministic. The original model and one reloaded from its SavedModel should give identical predictions:

translator("este é o primeiro livro que eu fiz.").numpy()
b'this is the first book i did .'
tf.saved_model.save(translator, export_dir='translator')
2021-11-02 15:48:30.232789: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
WARNING:absl:Found untraced functions such as embedding_4_layer_call_and_return_conditional_losses, embedding_4_layer_call_fn, dropout_37_layer_call_and_return_conditional_losses, dropout_37_layer_call_fn, embedding_5_layer_call_and_return_conditional_losses while saving (showing 5 of 560). These functions will not be directly callable after loading.
reloaded = tf.saved_model.load('translator')
reloaded("este é o primeiro livro que eu fiz.").numpy()
b'this is the first book i did .'

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. Furthermore, you can implement beam search to get better predictions.