理解语言的 Transformer 模型

在 tensorflow.google.cn 上查看 在 Google Colab 运行 在 Github 上查看源代码 下载此 notebook

本教程训练了一个 Transformer 模型 用于将葡萄牙语翻译成英语。这是一个高级示例,假定您具备文本生成(text generation)注意力机制(attention) 的知识。

Transformer 模型的核心思想是自注意力机制(self-attention)——能注意输入序列的不同位置以计算该序列的表示的能力。Transformer 创建了多层自注意力层(self-attetion layers)组成的堆栈,下文的按比缩放的点积注意力(Scaled dot product attention)多头注意力(Multi-head attention)部分对此进行了说明。

一个 transformer 模型用自注意力层而非 RNNsCNNs 来处理变长的输入。这种通用架构有一系列的优势:

  • 它不对数据间的时间/空间关系做任何假设。这是处理一组对象(objects)的理想选择(例如,星际争霸单位(StarCraft units))。
  • 层输出可以并行计算,而非像 RNN 这样的序列计算。
  • 远距离项可以影响彼此的输出,而无需经过许多 RNN 步骤或卷积层(例如,参见场景记忆 Transformer(Scene Memory Transformer)
  • 它能学习长距离的依赖。在许多序列任务中,这是一项挑战。

该架构的缺点是:

  • 对于时间序列,一个单位时间的输出是从整个历史记录计算的,而非仅从输入和当前的隐含状态计算得到。这可能效率较低。
  • 如果输入确实有时间/空间的关系,像文本,则必须加入一些位置编码,否则模型将有效地看到一堆单词。

在此 notebook 中训练完模型后,您将能输入葡萄牙语句子,得到其英文翻译。

Attention heatmap

from __future__ import absolute_import, division, print_function, unicode_literals

try:
  !pip install -q tf-nightly 
except Exception:
  pass
import tensorflow_datasets as tfds
import tensorflow as tf

import time
import numpy as np
import matplotlib.pyplot as plt
ERROR: tensorflow 2.1.0 has requirement gast==0.2.2, but you'll have gast 0.3.3 which is incompatible.

设置输入流水线(input pipeline)

使用 TFDS 来导入 葡萄牙语-英语翻译数据集,该数据集来自于 TED 演讲开放翻译项目.

该数据集包含来约 50000 条训练样本,1100 条验证样本,以及 2000 条测试样本。

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/1.0.0...

HBox(children=(FloatProgress(value=1.0, bar_style='info', description='Dl Completed...', max=1.0, style=Progre…
HBox(children=(FloatProgress(value=1.0, bar_style='info', description='Dl Size...', max=1.0, style=ProgressSty…
HBox(children=(FloatProgress(value=1.0, bar_style='info', description='Extraction completed...', max=1.0, styl…







HBox(children=(FloatProgress(value=1.0, bar_style='info', max=1.0), HTML(value='')))
Shuffling and writing examples to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0.incompleteMQS2HL/ted_hrlr_translate-train.tfrecord

HBox(children=(FloatProgress(value=0.0, max=51785.0), HTML(value='')))


HBox(children=(FloatProgress(value=1.0, bar_style='info', max=1.0), HTML(value='')))
Shuffling and writing examples to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0.incompleteMQS2HL/ted_hrlr_translate-validation.tfrecord

HBox(children=(FloatProgress(value=0.0, max=1193.0), HTML(value='')))


HBox(children=(FloatProgress(value=1.0, bar_style='info', max=1.0), HTML(value='')))
Shuffling and writing examples to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0.incompleteMQS2HL/ted_hrlr_translate-test.tfrecord

HBox(children=(FloatProgress(value=0.0, max=1803.0), HTML(value='')))
Dataset ted_hrlr_translate downloaded and prepared to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0. Subsequent calls will reuse this data.

从训练数据集创建自定义子词分词器(subwords tokenizer)。

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.

如果单词不在词典中,则分词器(tokenizer)通过将单词分解为子词来对字符串进行编码。

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

将开始和结束标记(token)添加到输入和目标。

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

Note:为了使本示例较小且相对较快,删除长度大于40个标记的样本。

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)

.map() 内部的操作以图模式(graph mode)运行,.map() 接收一个不具有 numpy 属性的图张量(graph tensor)。该分词器(tokenizer)需要将一个字符串或 Unicode 符号,编码成整数。因此,您需要在 tf.py_function 内部运行编码过程,tf.py_function 接收一个 eager 张量,该 eager 张量有一个包含字符串值的 numpy 属性。

def tf_encode(pt, en):
  result_pt, result_en = tf.py_function(encode, [pt, en], [tf.int64, tf.int64])
  result_pt.set_shape([None])
  result_en.set_shape([None])

  return result_pt, result_en
train_dataset = train_examples.map(tf_encode)
train_dataset = train_dataset.filter(filter_max_length)
# 将数据集缓存到内存中以加快读取速度。
train_dataset = train_dataset.cache()
train_dataset = train_dataset.shuffle(BUFFER_SIZE).padded_batch(BATCH_SIZE)
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)
pt_batch, en_batch = next(iter(val_dataset))
pt_batch, en_batch
(<tf.Tensor: shape=(64, 38), dtype=int64, numpy=
 array([[8214,  342, 3032, ...,    0,    0,    0],
        [8214,   95,  198, ...,    0,    0,    0],
        [8214, 4479, 7990, ...,    0,    0,    0],
        ...,
        [8214,  584,   12, ...,    0,    0,    0],
        [8214,   59, 1548, ...,    0,    0,    0],
        [8214,  118,   34, ...,    0,    0,    0]])>,
 <tf.Tensor: shape=(64, 40), dtype=int64, numpy=
 array([[8087,   98,   25, ...,    0,    0,    0],
        [8087,   12,   20, ...,    0,    0,    0],
        [8087,   12, 5453, ...,    0,    0,    0],
        ...,
        [8087,   18, 2059, ...,    0,    0,    0],
        [8087,   16, 1436, ...,    0,    0,    0],
        [8087,   15,   57, ...,    0,    0,    0]])>)

位置编码(Positional encoding)

因为该模型并不包括任何的循环(recurrence)或卷积,所以模型添加了位置编码,为模型提供一些关于单词在句子中相对位置的信息。

位置编码向量被加到嵌入(embedding)向量中。嵌入表示一个 d 维空间的标记,在 d 维空间中有着相似含义的标记会离彼此更近。但是,嵌入并没有对在一句话中的词的相对位置进行编码。因此,当加上位置编码后,词将基于它们含义的相似度以及它们在句子中的位置,在 d 维空间中离彼此更近。

参看 位置编码 的 notebook 了解更多信息。计算位置编码的公式如下:

$$\Large{PE_{(pos, 2i)} = sin(pos / 10000^{2i / d_{model}})} $$
$$\Large{PE_{(pos, 2i+1)} = cos(pos / 10000^{2i / d_{model}})} $$
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)
  
  # 将 sin 应用于数组中的偶数索引(indices);2i
  angle_rads[:, 0::2] = np.sin(angle_rads[:, 0::2])
  
  # 将 cos 应用于数组中的奇数索引;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)
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

遮挡(Masking)

遮挡一批序列中所有的填充标记(pad tokens)。这确保了模型不会将填充作为输入。该 mask 表明填充值 0 出现的位置:在这些位置 mask 输出 1,否则输出 0

def create_padding_mask(seq):
  seq = tf.cast(tf.math.equal(seq, 0), tf.float32)
  
  # 添加额外的维度来将填充加到
  # 注意力对数(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)>

前瞻遮挡(look-ahead mask)用于遮挡一个序列中的后续标记(future tokens)。换句话说,该 mask 表明了不应该使用的条目。

这意味着要预测第三个词,将仅使用第一个和第二个词。与此类似,预测第四个词,仅使用第一个,第二个和第三个词,依此类推。

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)>

按比缩放的点积注意力(Scaled dot product attention)

scaled_dot_product_attention

Transformer 使用的注意力函数有三个输入:Q(请求(query))、K(主键(key))、V(数值(value))。用于计算注意力权重的等式为:

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

点积注意力被缩小了深度的平方根倍。这样做是因为对于较大的深度值,点积的大小会增大,从而推动 softmax 函数往仅有很小的梯度的方向靠拢,导致了一种很硬的(hard)softmax。

例如,假设 QK 的均值为0,方差为1。它们的矩阵乘积将有均值为0,方差为 dk。因此,dk 的平方根被用于缩放(而非其他数值),因为,QK 的矩阵乘积的均值本应该为 0,方差本应该为1,这样会获得一个更平缓的 softmax。

遮挡(mask)与 -1e9(接近于负无穷)相乘。这样做是因为遮挡与缩放的 Q 和 K 的矩阵乘积相加,并在 softmax 之前立即应用。目标是将这些单元归零,因为 softmax 的较大负数输入在输出中接近于零。

def scaled_dot_product_attention(q, k, v, mask):
  """计算注意力权重。
  q, k, v 必须具有匹配的前置维度。
  k, v 必须有匹配的倒数第二个维度,例如:seq_len_k = seq_len_v。
  虽然 mask 根据其类型(填充或前瞻)有不同的形状,
  但是 mask 必须能进行广播转换以便求和。
  
  参数:
    q: 请求的形状 == (..., seq_len_q, depth)
    k: 主键的形状 == (..., seq_len_k, depth)
    v: 数值的形状 == (..., seq_len_v, depth_v)
    mask: Float 张量,其形状能转换成
          (..., seq_len_q, seq_len_k)。默认为None。
    
  返回值:
    输出,注意力权重
  """

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

  # 将 mask 加入到缩放的张量上。
  if mask is not None:
    scaled_attention_logits += (mask * -1e9)  

  # softmax 在最后一个轴(seq_len_k)上归一化,因此分数
  # 相加等于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

当 softmax 在 K 上进行归一化后,它的值决定了分配到 Q 的重要程度。

输出表示注意力权重和 V(数值)向量的乘积。这确保了要关注的词保持原样,而无关的词将被清除掉。

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)

# 这条 `请求(query)符合第二个`主键(key)`,
# 因此返回了第二个`数值(value)`。
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)
# 这条请求符合重复出现的主键(第三第四个),
# 因此,对所有的相关数值取了平均。
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)
# 这条请求符合第一和第二条主键,
# 因此,对它们的数值去了平均。
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)

将所有请求一起传递

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)

多头注意力(Multi-head attention)

multi-head attention

多头注意力由四部分组成:

  • 线性层并分拆成多头。
  • 按比缩放的点积注意力。
  • 多头及联。
  • 最后一层线性层。

每个多头注意力块有三个输入:Q(请求)、K(主键)、V(数值)。这些输入经过线性(Dense)层,并分拆成多头。

将上面定义的 scaled_dot_product_attention 函数应用于每个头(进行了广播(broadcasted)以提高效率)。注意力这步必须使用一个恰当的 mask。然后将每个头的注意力输出连接起来(用tf.transposetf.reshape),并放入最后的 Dense 层。

Q、K、和 V 被拆分到了多个头,而非单个的注意力头,因为多头允许模型共同注意来自不同表示空间的不同位置的信息。在分拆后,每个头部的维度减少,因此总的计算成本与有着全部维度的单个注意力头相同。

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):
    """分拆最后一个维度到 (num_heads, depth).
    转置结果使得形状为 (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

创建一个 MultiHeadAttention 层进行尝试。在序列中的每个位置 yMultiHeadAttention 在序列中的所有其他位置运行所有8个注意力头,在每个位置y,返回一个新的同样长度的向量。

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]))

点式前馈网络(Point wise feed forward network)

点式前馈网络由两层全联接层组成,两层之间有一个 ReLU 激活函数。

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])

编码与解码(Encoder and decoder)

transformer

Transformer 模型与标准的具有注意力机制的序列到序列模型(sequence to sequence with attention model),遵循相同的一般模式。

  • 输入语句经过 N 个编码器层,为序列中的每个词/标记生成一个输出。
  • 解码器关注编码器的输出以及它自身的输入(自注意力)来预测下一个词。

编码器层(Encoder layer)

每个编码器层包括以下子层:

  1. 多头注意力(有填充遮挡)
  2. 点式前馈网络(Point wise feed forward networks)。

每个子层在其周围有一个残差连接,然后进行层归一化。残差连接有助于避免深度网络中的梯度消失问题。

每个子层的输出是 LayerNorm(x + Sublayer(x))。归一化是在 d_model(最后一个)维度完成的。Transformer 中有 N 个编码器层。

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])

解码器层(Decoder layer)

每个解码器层包括以下子层:

  1. 遮挡的多头注意力(前瞻遮挡和填充遮挡)
  2. 多头注意力(用填充遮挡)。V(数值)和 K(主键)接收编码器输出作为输入。Q(请求)接收遮挡的多头注意力子层的输出
  3. 点式前馈网络

每个子层在其周围有一个残差连接,然后进行层归一化。每个子层的输出是 LayerNorm(x + Sublayer(x))。归一化是在 d_model(最后一个)维度完成的。

Transformer 中共有 N 个解码器层。

当 Q 接收到解码器的第一个注意力块的输出,并且 K 接收到编码器的输出时,注意力权重表示根据编码器的输出赋予解码器输入的重要性。换一种说法,解码器通过查看编码器输出和对其自身输出的自注意力,预测下一个词。参看按比缩放的点积注意力部分的演示。

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])

编码器(Encoder)

编码器 包括:

  1. 输入嵌入(Input Embedding)
  2. 位置编码(Positional Encoding)
  3. N 个编码器层(encoder layers)

输入经过嵌入(embedding)后,该嵌入与位置编码相加。该加法结果的输出是编码器层的输入。编码器的输出是解码器的输入。

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]
    
    # 将嵌入和位置编码相加。
    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)

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)

解码器(Decoder)

解码器包括:

  1. 输出嵌入(Output Embedding)
  2. 位置编码(Positional Encoding)
  3. N 个解码器层(decoder layers)

目标(target)经过一个嵌入后,该嵌入和位置编码相加。该加法结果是解码器层的输入。解码器的输出是最后的线性层的输入。

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['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,
                         maximum_position_encoding=5000)

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]))

创建 Transformer

Transformer 包括编码器,解码器和最后的线性层。解码器的输出是线性层的输入,返回线性层的输出。

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(Transformer, self).__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, 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, 
    pe_input=10000, pe_target=6000)

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])

配置超参数(hyperparameters)

为了让本示例小且相对较快,已经减小了num_layers、 d_model 和 dff 的值。

Transformer 的基础模型使用的数值为:num_layers=6d_model = 512dff = 2048。关于所有其他版本的 Transformer,请查阅论文

Note:通过改变以下数值,您可以获得在许多任务上达到最先进水平的模型。

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

优化器(Optimizer)

根据论文中的公式,将 Adam 优化器与自定义的学习速率调度程序(scheduler)配合使用。

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

损失函数与指标(Loss and metrics)

由于目标序列是填充(padded)过的,因此在计算损失函数时,应用填充遮挡非常重要。

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')

训练与检查点(Training and checkpointing)

transformer = Transformer(num_layers, d_model, num_heads, dff,
                          input_vocab_size, target_vocab_size, 
                          pe_input=input_vocab_size, 
                          pe_target=target_vocab_size,
                          rate=dropout_rate)
def create_masks(inp, tar):
  # 编码器填充遮挡
  enc_padding_mask = create_padding_mask(inp)
  
  # 在解码器的第二个注意力模块使用。
  # 该填充遮挡用于遮挡编码器的输出。
  dec_padding_mask = create_padding_mask(inp)
  
  # 在解码器的第一个注意力模块使用。
  # 用于填充(pad)和遮挡(mask)解码器获取到的输入的后续标记(future tokens)。
  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

创建检查点的路径和检查点管理器(manager)。这将用于在每 n 个周期(epochs)保存检查点。

checkpoint_path = "./checkpoints/train"

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

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

# 如果检查点存在,则恢复最新的检查点。
if ckpt_manager.latest_checkpoint:
  ckpt.restore(ckpt_manager.latest_checkpoint)
  print ('Latest checkpoint restored!!')

目标(target)被分成了 tar_inp 和 tar_real。tar_inp 作为输入传递到解码器。tar_real 是位移了 1 的同一个输入:在 tar_inp 中的每个位置,tar_real 包含了应该被预测到的下一个标记(token)。

例如,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"

Transformer 是一个自回归(auto-regressive)模型:它一次作一个部分的预测,然后使用到目前为止的自身的输出来决定下一步要做什么。

在训练过程中,本示例使用了 teacher-forcing 的方法(就像文本生成教程中一样)。无论模型在当前时间步骤下预测出什么,teacher-forcing 方法都会将真实的输出传递到下一个时间步骤上。

当 transformer 预测每个词时,自注意力(self-attention)功能使它能够查看输入序列中前面的单词,从而更好地预测下一个单词。

为了防止模型在期望的输出上达到峰值,模型使用了前瞻遮挡(look-ahead mask)。

EPOCHS = 20
# 该 @tf.function 将追踪-编译 train_step 到 TF 图中,以便更快地
# 执行。该函数专用于参数张量的精确形状。为了避免由于可变序列长度或可变
# 批次大小(最后一批次较小)导致的再追踪,使用 input_signature 指定
# 更多的通用形状。

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:]
  
  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)

葡萄牙语作为输入语言,英语为目标语言。

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 % 50 == 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 4.5346 Accuracy 0.0000
Epoch 1 Batch 50 Loss 4.2325 Accuracy 0.0009
Epoch 1 Batch 100 Loss 4.1759 Accuracy 0.0132
Epoch 1 Batch 150 Loss 4.1361 Accuracy 0.0178
Epoch 1 Batch 200 Loss 4.0848 Accuracy 0.0201
Epoch 1 Batch 250 Loss 4.0141 Accuracy 0.0215
Epoch 1 Batch 300 Loss 3.9299 Accuracy 0.0226
Epoch 1 Batch 350 Loss 3.8423 Accuracy 0.0259
Epoch 1 Batch 400 Loss 3.7545 Accuracy 0.0296
Epoch 1 Batch 450 Loss 3.6796 Accuracy 0.0332
Epoch 1 Batch 500 Loss 3.6120 Accuracy 0.0365
Epoch 1 Batch 550 Loss 3.5470 Accuracy 0.0400
Epoch 1 Batch 600 Loss 3.4856 Accuracy 0.0437
Epoch 1 Batch 650 Loss 3.4281 Accuracy 0.0472
Epoch 1 Batch 700 Loss 3.3742 Accuracy 0.0507
Epoch 1 Loss 3.3725 Accuracy 0.0508
Time taken for 1 epoch: 57.364415407180786 secs

Epoch 2 Batch 0 Loss 2.7556 Accuracy 0.1128
Epoch 2 Batch 50 Loss 2.5457 Accuracy 0.1048
Epoch 2 Batch 100 Loss 2.5533 Accuracy 0.1061
Epoch 2 Batch 150 Loss 2.5281 Accuracy 0.1086
Epoch 2 Batch 200 Loss 2.5075 Accuracy 0.1104
Epoch 2 Batch 250 Loss 2.4885 Accuracy 0.1119
Epoch 2 Batch 300 Loss 2.4743 Accuracy 0.1135
Epoch 2 Batch 350 Loss 2.4630 Accuracy 0.1150
Epoch 2 Batch 400 Loss 2.4446 Accuracy 0.1166
Epoch 2 Batch 450 Loss 2.4273 Accuracy 0.1178
Epoch 2 Batch 500 Loss 2.4176 Accuracy 0.1192
Epoch 2 Batch 550 Loss 2.4059 Accuracy 0.1206
Epoch 2 Batch 600 Loss 2.3925 Accuracy 0.1218
Epoch 2 Batch 650 Loss 2.3809 Accuracy 0.1230
Epoch 2 Batch 700 Loss 2.3720 Accuracy 0.1241
Epoch 2 Loss 2.3722 Accuracy 0.1242
Time taken for 1 epoch: 30.23842215538025 secs

Epoch 3 Batch 0 Loss 2.0335 Accuracy 0.1310
Epoch 3 Batch 50 Loss 2.1599 Accuracy 0.1389
Epoch 3 Batch 100 Loss 2.1513 Accuracy 0.1407
Epoch 3 Batch 150 Loss 2.1390 Accuracy 0.1416
Epoch 3 Batch 200 Loss 2.1431 Accuracy 0.1429
Epoch 3 Batch 250 Loss 2.1412 Accuracy 0.1439
Epoch 3 Batch 300 Loss 2.1380 Accuracy 0.1448
Epoch 3 Batch 350 Loss 2.1423 Accuracy 0.1459
Epoch 3 Batch 400 Loss 2.1383 Accuracy 0.1465
Epoch 3 Batch 450 Loss 2.1313 Accuracy 0.1469
Epoch 3 Batch 500 Loss 2.1287 Accuracy 0.1476
Epoch 3 Batch 550 Loss 2.1245 Accuracy 0.1484
Epoch 3 Batch 600 Loss 2.1195 Accuracy 0.1492
Epoch 3 Batch 650 Loss 2.1109 Accuracy 0.1499
Epoch 3 Batch 700 Loss 2.1033 Accuracy 0.1508
Epoch 3 Loss 2.1030 Accuracy 0.1508
Time taken for 1 epoch: 31.80252456665039 secs

Epoch 4 Batch 0 Loss 1.7214 Accuracy 0.1686
Epoch 4 Batch 50 Loss 1.9639 Accuracy 0.1652
Epoch 4 Batch 100 Loss 1.9519 Accuracy 0.1650
Epoch 4 Batch 150 Loss 1.9437 Accuracy 0.1663
Epoch 4 Batch 200 Loss 1.9440 Accuracy 0.1675
Epoch 4 Batch 250 Loss 1.9403 Accuracy 0.1687
Epoch 4 Batch 300 Loss 1.9368 Accuracy 0.1697
Epoch 4 Batch 350 Loss 1.9308 Accuracy 0.1706
Epoch 4 Batch 400 Loss 1.9261 Accuracy 0.1718
Epoch 4 Batch 450 Loss 1.9186 Accuracy 0.1730
Epoch 4 Batch 500 Loss 1.9108 Accuracy 0.1738
Epoch 4 Batch 550 Loss 1.8989 Accuracy 0.1746
Epoch 4 Batch 600 Loss 1.8908 Accuracy 0.1758
Epoch 4 Batch 650 Loss 1.8817 Accuracy 0.1766
Epoch 4 Batch 700 Loss 1.8748 Accuracy 0.1775
Epoch 4 Loss 1.8751 Accuracy 0.1776
Time taken for 1 epoch: 31.94111704826355 secs

Epoch 5 Batch 0 Loss 1.7333 Accuracy 0.1997
Epoch 5 Batch 50 Loss 1.6949 Accuracy 0.1947
Epoch 5 Batch 100 Loss 1.7087 Accuracy 0.1968
Epoch 5 Batch 150 Loss 1.7051 Accuracy 0.1986
Epoch 5 Batch 200 Loss 1.6993 Accuracy 0.1985
Epoch 5 Batch 250 Loss 1.6935 Accuracy 0.1991
Epoch 5 Batch 300 Loss 1.6905 Accuracy 0.1997
Epoch 5 Batch 350 Loss 1.6902 Accuracy 0.2006
Epoch 5 Batch 400 Loss 1.6859 Accuracy 0.2014
Epoch 5 Batch 450 Loss 1.6818 Accuracy 0.2022
Epoch 5 Batch 500 Loss 1.6762 Accuracy 0.2031
Epoch 5 Batch 550 Loss 1.6687 Accuracy 0.2036
Epoch 5 Batch 600 Loss 1.6681 Accuracy 0.2046
Epoch 5 Batch 650 Loss 1.6617 Accuracy 0.2052
Epoch 5 Batch 700 Loss 1.6577 Accuracy 0.2057
Saving checkpoint for epoch 5 at ./checkpoints/train/ckpt-1
Epoch 5 Loss 1.6570 Accuracy 0.2056
Time taken for 1 epoch: 31.290344715118408 secs

Epoch 6 Batch 0 Loss 1.5459 Accuracy 0.2282
Epoch 6 Batch 50 Loss 1.4751 Accuracy 0.2198
Epoch 6 Batch 100 Loss 1.4954 Accuracy 0.2208
Epoch 6 Batch 150 Loss 1.4944 Accuracy 0.2215
Epoch 6 Batch 200 Loss 1.4976 Accuracy 0.2226
Epoch 6 Batch 250 Loss 1.4977 Accuracy 0.2229
Epoch 6 Batch 300 Loss 1.4957 Accuracy 0.2234
Epoch 6 Batch 350 Loss 1.4942 Accuracy 0.2239
Epoch 6 Batch 400 Loss 1.4912 Accuracy 0.2244
Epoch 6 Batch 450 Loss 1.4833 Accuracy 0.2246
Epoch 6 Batch 500 Loss 1.4814 Accuracy 0.2251
Epoch 6 Batch 550 Loss 1.4789 Accuracy 0.2255
Epoch 6 Batch 600 Loss 1.4767 Accuracy 0.2256
Epoch 6 Batch 650 Loss 1.4731 Accuracy 0.2260
Epoch 6 Batch 700 Loss 1.4695 Accuracy 0.2262
Epoch 6 Loss 1.4691 Accuracy 0.2262
Time taken for 1 epoch: 31.403314352035522 secs

Epoch 7 Batch 0 Loss 1.1021 Accuracy 0.2200
Epoch 7 Batch 50 Loss 1.2887 Accuracy 0.2399
Epoch 7 Batch 100 Loss 1.2878 Accuracy 0.2399
Epoch 7 Batch 150 Loss 1.2850 Accuracy 0.2411
Epoch 7 Batch 200 Loss 1.2964 Accuracy 0.2417
Epoch 7 Batch 250 Loss 1.2954 Accuracy 0.2429
Epoch 7 Batch 300 Loss 1.3004 Accuracy 0.2435
Epoch 7 Batch 350 Loss 1.2995 Accuracy 0.2440
Epoch 7 Batch 400 Loss 1.2959 Accuracy 0.2445
Epoch 7 Batch 450 Loss 1.2950 Accuracy 0.2451
Epoch 7 Batch 500 Loss 1.2908 Accuracy 0.2455
Epoch 7 Batch 550 Loss 1.2885 Accuracy 0.2459
Epoch 7 Batch 600 Loss 1.2860 Accuracy 0.2463
Epoch 7 Batch 650 Loss 1.2825 Accuracy 0.2467
Epoch 7 Batch 700 Loss 1.2774 Accuracy 0.2472
Epoch 7 Loss 1.2774 Accuracy 0.2472
Time taken for 1 epoch: 31.580658674240112 secs

Epoch 8 Batch 0 Loss 1.1012 Accuracy 0.2483
Epoch 8 Batch 50 Loss 1.1082 Accuracy 0.2613
Epoch 8 Batch 100 Loss 1.1089 Accuracy 0.2629
Epoch 8 Batch 150 Loss 1.1212 Accuracy 0.2637
Epoch 8 Batch 200 Loss 1.1195 Accuracy 0.2631
Epoch 8 Batch 250 Loss 1.1232 Accuracy 0.2632
Epoch 8 Batch 300 Loss 1.1240 Accuracy 0.2643
Epoch 8 Batch 350 Loss 1.1267 Accuracy 0.2642
Epoch 8 Batch 400 Loss 1.1283 Accuracy 0.2650
Epoch 8 Batch 450 Loss 1.1263 Accuracy 0.2653
Epoch 8 Batch 500 Loss 1.1249 Accuracy 0.2656
Epoch 8 Batch 550 Loss 1.1236 Accuracy 0.2661
Epoch 8 Batch 600 Loss 1.1245 Accuracy 0.2662
Epoch 8 Batch 650 Loss 1.1231 Accuracy 0.2664
Epoch 8 Batch 700 Loss 1.1243 Accuracy 0.2667
Epoch 8 Loss 1.1245 Accuracy 0.2667
Time taken for 1 epoch: 30.735326766967773 secs

Epoch 9 Batch 0 Loss 0.9336 Accuracy 0.2704
Epoch 9 Batch 50 Loss 0.9938 Accuracy 0.2799
Epoch 9 Batch 100 Loss 0.9929 Accuracy 0.2787
Epoch 9 Batch 150 Loss 0.9966 Accuracy 0.2799
Epoch 9 Batch 200 Loss 1.0054 Accuracy 0.2799
Epoch 9 Batch 250 Loss 1.0082 Accuracy 0.2802
Epoch 9 Batch 300 Loss 1.0084 Accuracy 0.2804
Epoch 9 Batch 350 Loss 1.0067 Accuracy 0.2804
Epoch 9 Batch 400 Loss 1.0057 Accuracy 0.2808
Epoch 9 Batch 450 Loss 1.0075 Accuracy 0.2811
Epoch 9 Batch 500 Loss 1.0072 Accuracy 0.2810
Epoch 9 Batch 550 Loss 1.0074 Accuracy 0.2811
Epoch 9 Batch 600 Loss 1.0072 Accuracy 0.2810
Epoch 9 Batch 650 Loss 1.0084 Accuracy 0.2808
Epoch 9 Batch 700 Loss 1.0098 Accuracy 0.2805
Epoch 9 Loss 1.0094 Accuracy 0.2805
Time taken for 1 epoch: 30.613189935684204 secs

Epoch 10 Batch 0 Loss 0.8017 Accuracy 0.2989
Epoch 10 Batch 50 Loss 0.9083 Accuracy 0.2953
Epoch 10 Batch 100 Loss 0.9054 Accuracy 0.2922
Epoch 10 Batch 150 Loss 0.9138 Accuracy 0.2929
Epoch 10 Batch 200 Loss 0.9182 Accuracy 0.2931
Epoch 10 Batch 250 Loss 0.9192 Accuracy 0.2926
Epoch 10 Batch 300 Loss 0.9197 Accuracy 0.2925
Epoch 10 Batch 350 Loss 0.9159 Accuracy 0.2919
Epoch 10 Batch 400 Loss 0.9186 Accuracy 0.2916
Epoch 10 Batch 450 Loss 0.9180 Accuracy 0.2917
Epoch 10 Batch 500 Loss 0.9218 Accuracy 0.2920
Epoch 10 Batch 550 Loss 0.9214 Accuracy 0.2917
Epoch 10 Batch 600 Loss 0.9215 Accuracy 0.2911
Epoch 10 Batch 650 Loss 0.9232 Accuracy 0.2909
Epoch 10 Batch 700 Loss 0.9232 Accuracy 0.2908
Saving checkpoint for epoch 10 at ./checkpoints/train/ckpt-2
Epoch 10 Loss 0.9234 Accuracy 0.2908
Time taken for 1 epoch: 30.585023641586304 secs

Epoch 11 Batch 0 Loss 0.7168 Accuracy 0.3077
Epoch 11 Batch 50 Loss 0.8103 Accuracy 0.3022
Epoch 11 Batch 100 Loss 0.8162 Accuracy 0.3027
Epoch 11 Batch 150 Loss 0.8211 Accuracy 0.3035
Epoch 11 Batch 200 Loss 0.8298 Accuracy 0.3034
Epoch 11 Batch 250 Loss 0.8404 Accuracy 0.3040
Epoch 11 Batch 300 Loss 0.8419 Accuracy 0.3033
Epoch 11 Batch 350 Loss 0.8431 Accuracy 0.3035
Epoch 11 Batch 400 Loss 0.8451 Accuracy 0.3029
Epoch 11 Batch 450 Loss 0.8474 Accuracy 0.3027
Epoch 11 Batch 500 Loss 0.8498 Accuracy 0.3021
Epoch 11 Batch 550 Loss 0.8510 Accuracy 0.3017
Epoch 11 Batch 600 Loss 0.8539 Accuracy 0.3016
Epoch 11 Batch 650 Loss 0.8569 Accuracy 0.3013
Epoch 11 Batch 700 Loss 0.8582 Accuracy 0.3009
Epoch 11 Loss 0.8584 Accuracy 0.3009
Time taken for 1 epoch: 30.520817279815674 secs

Epoch 12 Batch 0 Loss 0.6941 Accuracy 0.2752
Epoch 12 Batch 50 Loss 0.7600 Accuracy 0.3106
Epoch 12 Batch 100 Loss 0.7632 Accuracy 0.3091
Epoch 12 Batch 150 Loss 0.7732 Accuracy 0.3102
Epoch 12 Batch 200 Loss 0.7748 Accuracy 0.3101
Epoch 12 Batch 250 Loss 0.7785 Accuracy 0.3095
Epoch 12 Batch 300 Loss 0.7800 Accuracy 0.3088
Epoch 12 Batch 350 Loss 0.7842 Accuracy 0.3088
Epoch 12 Batch 400 Loss 0.7846 Accuracy 0.3082
Epoch 12 Batch 450 Loss 0.7864 Accuracy 0.3077
Epoch 12 Batch 500 Loss 0.7886 Accuracy 0.3078
Epoch 12 Batch 550 Loss 0.7933 Accuracy 0.3081
Epoch 12 Batch 600 Loss 0.7960 Accuracy 0.3080
Epoch 12 Batch 650 Loss 0.7966 Accuracy 0.3076
Epoch 12 Batch 700 Loss 0.8012 Accuracy 0.3078
Epoch 12 Loss 0.8012 Accuracy 0.3077
Time taken for 1 epoch: 30.34489870071411 secs

Epoch 13 Batch 0 Loss 0.6835 Accuracy 0.3145
Epoch 13 Batch 50 Loss 0.7190 Accuracy 0.3171
Epoch 13 Batch 100 Loss 0.7190 Accuracy 0.3170
Epoch 13 Batch 150 Loss 0.7224 Accuracy 0.3165
Epoch 13 Batch 200 Loss 0.7305 Accuracy 0.3168
Epoch 13 Batch 250 Loss 0.7288 Accuracy 0.3158
Epoch 13 Batch 300 Loss 0.7320 Accuracy 0.3157
Epoch 13 Batch 350 Loss 0.7357 Accuracy 0.3157
Epoch 13 Batch 400 Loss 0.7376 Accuracy 0.3151
Epoch 13 Batch 450 Loss 0.7403 Accuracy 0.3152
Epoch 13 Batch 500 Loss 0.7436 Accuracy 0.3150
Epoch 13 Batch 550 Loss 0.7452 Accuracy 0.3146
Epoch 13 Batch 600 Loss 0.7489 Accuracy 0.3145
Epoch 13 Batch 650 Loss 0.7509 Accuracy 0.3139
Epoch 13 Batch 700 Loss 0.7545 Accuracy 0.3139
Epoch 13 Loss 0.7548 Accuracy 0.3140
Time taken for 1 epoch: 30.844294786453247 secs

Epoch 14 Batch 0 Loss 0.7545 Accuracy 0.3039
Epoch 14 Batch 50 Loss 0.6728 Accuracy 0.3245
Epoch 14 Batch 100 Loss 0.6777 Accuracy 0.3243
Epoch 14 Batch 150 Loss 0.6835 Accuracy 0.3239
Epoch 14 Batch 200 Loss 0.6864 Accuracy 0.3240
Epoch 14 Batch 250 Loss 0.6904 Accuracy 0.3231
Epoch 14 Batch 300 Loss 0.6937 Accuracy 0.3230
Epoch 14 Batch 350 Loss 0.6977 Accuracy 0.3227
Epoch 14 Batch 400 Loss 0.7005 Accuracy 0.3223
Epoch 14 Batch 450 Loss 0.7024 Accuracy 0.3217
Epoch 14 Batch 500 Loss 0.7045 Accuracy 0.3212
Epoch 14 Batch 550 Loss 0.7069 Accuracy 0.3209
Epoch 14 Batch 600 Loss 0.7085 Accuracy 0.3202
Epoch 14 Batch 650 Loss 0.7121 Accuracy 0.3201
Epoch 14 Batch 700 Loss 0.7139 Accuracy 0.3196
Epoch 14 Loss 0.7143 Accuracy 0.3196
Time taken for 1 epoch: 30.751317501068115 secs

Epoch 15 Batch 0 Loss 0.6035 Accuracy 0.3616
Epoch 15 Batch 50 Loss 0.6348 Accuracy 0.3309
Epoch 15 Batch 100 Loss 0.6415 Accuracy 0.3283
Epoch 15 Batch 150 Loss 0.6452 Accuracy 0.3272
Epoch 15 Batch 200 Loss 0.6454 Accuracy 0.3256
Epoch 15 Batch 250 Loss 0.6495 Accuracy 0.3262
Epoch 15 Batch 300 Loss 0.6536 Accuracy 0.3262
Epoch 15 Batch 350 Loss 0.6556 Accuracy 0.3263
Epoch 15 Batch 400 Loss 0.6606 Accuracy 0.3265
Epoch 15 Batch 450 Loss 0.6647 Accuracy 0.3264
Epoch 15 Batch 500 Loss 0.6672 Accuracy 0.3258
Epoch 15 Batch 550 Loss 0.6699 Accuracy 0.3258
Epoch 15 Batch 600 Loss 0.6722 Accuracy 0.3256
Epoch 15 Batch 650 Loss 0.6753 Accuracy 0.3252
Epoch 15 Batch 700 Loss 0.6780 Accuracy 0.3249
Saving checkpoint for epoch 15 at ./checkpoints/train/ckpt-3
Epoch 15 Loss 0.6782 Accuracy 0.3249
Time taken for 1 epoch: 30.690065145492554 secs

Epoch 16 Batch 0 Loss 0.5818 Accuracy 0.3558
Epoch 16 Batch 50 Loss 0.6075 Accuracy 0.3382
Epoch 16 Batch 100 Loss 0.6105 Accuracy 0.3356
Epoch 16 Batch 150 Loss 0.6157 Accuracy 0.3344
Epoch 16 Batch 200 Loss 0.6151 Accuracy 0.3337
Epoch 16 Batch 250 Loss 0.6193 Accuracy 0.3346
Epoch 16 Batch 300 Loss 0.6249 Accuracy 0.3353
Epoch 16 Batch 350 Loss 0.6290 Accuracy 0.3348
Epoch 16 Batch 400 Loss 0.6301 Accuracy 0.3340
Epoch 16 Batch 450 Loss 0.6324 Accuracy 0.3334
Epoch 16 Batch 500 Loss 0.6339 Accuracy 0.3324
Epoch 16 Batch 550 Loss 0.6381 Accuracy 0.3316
Epoch 16 Batch 600 Loss 0.6432 Accuracy 0.3312
Epoch 16 Batch 650 Loss 0.6451 Accuracy 0.3308
Epoch 16 Batch 700 Loss 0.6469 Accuracy 0.3299
Epoch 16 Loss 0.6474 Accuracy 0.3299
Time taken for 1 epoch: 30.30579686164856 secs

Epoch 17 Batch 0 Loss 0.6191 Accuracy 0.3348
Epoch 17 Batch 50 Loss 0.5801 Accuracy 0.3370
Epoch 17 Batch 100 Loss 0.5833 Accuracy 0.3353
Epoch 17 Batch 150 Loss 0.5860 Accuracy 0.3359
Epoch 17 Batch 200 Loss 0.5888 Accuracy 0.3366
Epoch 17 Batch 250 Loss 0.5913 Accuracy 0.3349
Epoch 17 Batch 300 Loss 0.5939 Accuracy 0.3346
Epoch 17 Batch 350 Loss 0.5978 Accuracy 0.3350
Epoch 17 Batch 400 Loss 0.6011 Accuracy 0.3346
Epoch 17 Batch 450 Loss 0.6039 Accuracy 0.3344
Epoch 17 Batch 500 Loss 0.6077 Accuracy 0.3349
Epoch 17 Batch 550 Loss 0.6118 Accuracy 0.3352
Epoch 17 Batch 600 Loss 0.6140 Accuracy 0.3348
Epoch 17 Batch 650 Loss 0.6164 Accuracy 0.3345
Epoch 17 Batch 700 Loss 0.6188 Accuracy 0.3339
Epoch 17 Loss 0.6192 Accuracy 0.3340
Time taken for 1 epoch: 30.403062343597412 secs

Epoch 18 Batch 0 Loss 0.6076 Accuracy 0.3612
Epoch 18 Batch 50 Loss 0.5461 Accuracy 0.3387
Epoch 18 Batch 100 Loss 0.5601 Accuracy 0.3421
Epoch 18 Batch 150 Loss 0.5619 Accuracy 0.3416
Epoch 18 Batch 200 Loss 0.5673 Accuracy 0.3419
Epoch 18 Batch 250 Loss 0.5720 Accuracy 0.3424
Epoch 18 Batch 300 Loss 0.5734 Accuracy 0.3410
Epoch 18 Batch 350 Loss 0.5748 Accuracy 0.3403
Epoch 18 Batch 400 Loss 0.5775 Accuracy 0.3401
Epoch 18 Batch 450 Loss 0.5817 Accuracy 0.3404
Epoch 18 Batch 500 Loss 0.5843 Accuracy 0.3402
Epoch 18 Batch 550 Loss 0.5873 Accuracy 0.3397
Epoch 18 Batch 600 Loss 0.5902 Accuracy 0.3393
Epoch 18 Batch 650 Loss 0.5932 Accuracy 0.3388
Epoch 18 Batch 700 Loss 0.5958 Accuracy 0.3385
Epoch 18 Loss 0.5957 Accuracy 0.3384
Time taken for 1 epoch: 30.320571422576904 secs

Epoch 19 Batch 0 Loss 0.4949 Accuracy 0.3507
Epoch 19 Batch 50 Loss 0.5292 Accuracy 0.3479
Epoch 19 Batch 100 Loss 0.5281 Accuracy 0.3462
Epoch 19 Batch 150 Loss 0.5384 Accuracy 0.3467
Epoch 19 Batch 200 Loss 0.5429 Accuracy 0.3462
Epoch 19 Batch 250 Loss 0.5469 Accuracy 0.3455
Epoch 19 Batch 300 Loss 0.5508 Accuracy 0.3455
Epoch 19 Batch 350 Loss 0.5552 Accuracy 0.3452
Epoch 19 Batch 400 Loss 0.5565 Accuracy 0.3440
Epoch 19 Batch 450 Loss 0.5588 Accuracy 0.3433
Epoch 19 Batch 500 Loss 0.5621 Accuracy 0.3430
Epoch 19 Batch 550 Loss 0.5648 Accuracy 0.3424
Epoch 19 Batch 600 Loss 0.5667 Accuracy 0.3422
Epoch 19 Batch 650 Loss 0.5686 Accuracy 0.3414
Epoch 19 Batch 700 Loss 0.5718 Accuracy 0.3411
Epoch 19 Loss 0.5718 Accuracy 0.3411
Time taken for 1 epoch: 30.32275629043579 secs

Epoch 20 Batch 0 Loss 0.5058 Accuracy 0.3450
Epoch 20 Batch 50 Loss 0.4951 Accuracy 0.3453
Epoch 20 Batch 100 Loss 0.5074 Accuracy 0.3501
Epoch 20 Batch 150 Loss 0.5169 Accuracy 0.3490
Epoch 20 Batch 200 Loss 0.5191 Accuracy 0.3489
Epoch 20 Batch 250 Loss 0.5242 Accuracy 0.3485
Epoch 20 Batch 300 Loss 0.5278 Accuracy 0.3485
Epoch 20 Batch 350 Loss 0.5325 Accuracy 0.3479
Epoch 20 Batch 400 Loss 0.5339 Accuracy 0.3475
Epoch 20 Batch 450 Loss 0.5368 Accuracy 0.3477
Epoch 20 Batch 500 Loss 0.5398 Accuracy 0.3473
Epoch 20 Batch 550 Loss 0.5427 Accuracy 0.3469
Epoch 20 Batch 600 Loss 0.5457 Accuracy 0.3462
Epoch 20 Batch 650 Loss 0.5487 Accuracy 0.3459
Epoch 20 Batch 700 Loss 0.5515 Accuracy 0.3455
Saving checkpoint for epoch 20 at ./checkpoints/train/ckpt-4
Epoch 20 Loss 0.5517 Accuracy 0.3455
Time taken for 1 epoch: 30.38394331932068 secs

评估(Evaluate)

以下步骤用于评估:

  • 用葡萄牙语分词器(tokenizer_pt)编码输入语句。此外,添加开始和结束标记,这样输入就与模型训练的内容相同。这是编码器输入。
  • 解码器输入为 start token == tokenizer_en.vocab_size
  • 计算填充遮挡和前瞻遮挡。
  • 解码器通过查看编码器输出和它自身的输出(自注意力)给出预测。
  • 选择最后一个词并计算它的 argmax。
  • 将预测的词连接到解码器输入,然后传递给解码器。
  • 在这种方法中,解码器根据它预测的之前的词预测下一个。

Note:这里使用的模型具有较小的能力以保持相对较快,因此预测可能不太正确。要复现论文中的结果,请使用全部数据集,并通过修改上述超参数来使用基础 transformer 模型或者 transformer XL。

def evaluate(inp_sentence):
  start_token = [tokenizer_pt.vocab_size]
  end_token = [tokenizer_pt.vocab_size + 1]
  
  # 输入语句是葡萄牙语,增加开始和结束标记
  inp_sentence = start_token + tokenizer_pt.encode(inp_sentence) + end_token
  encoder_input = tf.expand_dims(inp_sentence, 0)
  
  # 因为目标是英语,输入 transformer 的第一个词应该是
  # 英语的开始标记。
  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)
    
    # 从 seq_len 维度选择最后一个词
    predictions = predictions[: ,-1:, :]  # (batch_size, 1, vocab_size)

    predicted_id = tf.cast(tf.argmax(predictions, axis=-1), tf.int32)
    
    # 如果 predicted_id 等于结束标记,就返回结果
    if predicted_id == tokenizer_en.vocab_size+1:
      return tf.squeeze(output, axis=0), attention_weights
    
    # 连接 predicted_id 与输出,作为解码器的输入传递到解码器。
    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)
    
    # 画出注意力权重
    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 that .
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 .
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: so i 'm going to share with you a few really bad stories of some magic things , which has occurred .
Real translation: so i 'll just share with you some stories very quickly of some magical things that have happened .

您可以为 plot 参数传递不同的层和解码器的注意力模块。

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 today .

png

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

总结

在本教程中,您已经学习了位置编码,多头注意力,遮挡的重要性以及如何创建一个 transformer。

尝试使用一个不同的数据集来训练 transformer。您可也可以通过修改上述的超参数来创建基础 transformer 或者 transformer XL。您也可以使用这里定义的层来创建 BERT 并训练最先进的模型。此外,您可以实现 beam search 得到更好的预测。