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

import tensorflow_datasets as tfds
import tensorflow as tf

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

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/pt_to_en/1.0.0 (download: 124.94 MiB, generated: Unknown size, total: 124.94 MiB) to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0...
Shuffling and writing examples to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0.incomplete8YZA1B/ted_hrlr_translate-train.tfrecord
Shuffling and writing examples to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0.incomplete8YZA1B/ted_hrlr_translate-validation.tfrecord
Shuffling and writing examples to /home/kbuilder/tensorflow_datasets/ted_hrlr_translate/pt_to_en/1.0.0.incomplete8YZA1B/ted_hrlr_translate-test.tfrecord
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 3.8794 Accuracy 0.0000
Epoch 1 Batch 50 Loss 4.1838 Accuracy 0.0036
Epoch 1 Batch 100 Loss 4.1683 Accuracy 0.0137
Epoch 1 Batch 150 Loss 4.1365 Accuracy 0.0176
Epoch 1 Batch 200 Loss 4.0704 Accuracy 0.0202
Epoch 1 Batch 250 Loss 3.9970 Accuracy 0.0236
Epoch 1 Batch 300 Loss 3.9164 Accuracy 0.0277
Epoch 1 Batch 350 Loss 3.8260 Accuracy 0.0309
Epoch 1 Batch 400 Loss 3.7435 Accuracy 0.0333
Epoch 1 Batch 450 Loss 3.6661 Accuracy 0.0360
Epoch 1 Batch 500 Loss 3.5969 Accuracy 0.0390
Epoch 1 Batch 550 Loss 3.5352 Accuracy 0.0423
Epoch 1 Batch 600 Loss 3.4772 Accuracy 0.0459
Epoch 1 Batch 650 Loss 3.4221 Accuracy 0.0493
Epoch 1 Batch 700 Loss 3.3676 Accuracy 0.0527
Epoch 1 Loss 3.3652 Accuracy 0.0528
Time taken for 1 epoch: 53.67050647735596 secs

Epoch 2 Batch 0 Loss 2.7213 Accuracy 0.1039
Epoch 2 Batch 50 Loss 2.5758 Accuracy 0.1030
Epoch 2 Batch 100 Loss 2.5404 Accuracy 0.1052
Epoch 2 Batch 150 Loss 2.5294 Accuracy 0.1077
Epoch 2 Batch 200 Loss 2.5210 Accuracy 0.1098
Epoch 2 Batch 250 Loss 2.4973 Accuracy 0.1116
Epoch 2 Batch 300 Loss 2.4833 Accuracy 0.1136
Epoch 2 Batch 350 Loss 2.4689 Accuracy 0.1154
Epoch 2 Batch 400 Loss 2.4522 Accuracy 0.1168
Epoch 2 Batch 450 Loss 2.4335 Accuracy 0.1182
Epoch 2 Batch 500 Loss 2.4186 Accuracy 0.1198
Epoch 2 Batch 550 Loss 2.4053 Accuracy 0.1210
Epoch 2 Batch 600 Loss 2.3930 Accuracy 0.1222
Epoch 2 Batch 650 Loss 2.3769 Accuracy 0.1235
Epoch 2 Batch 700 Loss 2.3685 Accuracy 0.1248
Epoch 2 Loss 2.3680 Accuracy 0.1248
Time taken for 1 epoch: 28.824520587921143 secs

Epoch 3 Batch 0 Loss 2.2588 Accuracy 0.1619
Epoch 3 Batch 50 Loss 2.2238 Accuracy 0.1429
Epoch 3 Batch 100 Loss 2.1873 Accuracy 0.1437
Epoch 3 Batch 150 Loss 2.1784 Accuracy 0.1444
Epoch 3 Batch 200 Loss 2.1659 Accuracy 0.1447
Epoch 3 Batch 250 Loss 2.1606 Accuracy 0.1453
Epoch 3 Batch 300 Loss 2.1509 Accuracy 0.1457
Epoch 3 Batch 350 Loss 2.1462 Accuracy 0.1464
Epoch 3 Batch 400 Loss 2.1421 Accuracy 0.1472
Epoch 3 Batch 450 Loss 2.1359 Accuracy 0.1480
Epoch 3 Batch 500 Loss 2.1285 Accuracy 0.1485
Epoch 3 Batch 550 Loss 2.1218 Accuracy 0.1494
Epoch 3 Batch 600 Loss 2.1144 Accuracy 0.1503
Epoch 3 Batch 650 Loss 2.1074 Accuracy 0.1511
Epoch 3 Batch 700 Loss 2.0998 Accuracy 0.1518
Epoch 3 Loss 2.1001 Accuracy 0.1519
Time taken for 1 epoch: 28.887129068374634 secs

Epoch 4 Batch 0 Loss 1.6940 Accuracy 0.1550
Epoch 4 Batch 50 Loss 1.9399 Accuracy 0.1647
Epoch 4 Batch 100 Loss 1.9478 Accuracy 0.1670
Epoch 4 Batch 150 Loss 1.9509 Accuracy 0.1687
Epoch 4 Batch 200 Loss 1.9421 Accuracy 0.1694
Epoch 4 Batch 250 Loss 1.9322 Accuracy 0.1697
Epoch 4 Batch 300 Loss 1.9261 Accuracy 0.1710
Epoch 4 Batch 350 Loss 1.9191 Accuracy 0.1722
Epoch 4 Batch 400 Loss 1.9112 Accuracy 0.1734
Epoch 4 Batch 450 Loss 1.9015 Accuracy 0.1741
Epoch 4 Batch 500 Loss 1.8907 Accuracy 0.1751
Epoch 4 Batch 550 Loss 1.8844 Accuracy 0.1762
Epoch 4 Batch 600 Loss 1.8764 Accuracy 0.1773
Epoch 4 Batch 650 Loss 1.8675 Accuracy 0.1782
Epoch 4 Batch 700 Loss 1.8600 Accuracy 0.1790
Epoch 4 Loss 1.8596 Accuracy 0.1790
Time taken for 1 epoch: 28.815260648727417 secs

Epoch 5 Batch 0 Loss 1.5844 Accuracy 0.1799
Epoch 5 Batch 50 Loss 1.6949 Accuracy 0.1966
Epoch 5 Batch 100 Loss 1.7072 Accuracy 0.1989
Epoch 5 Batch 150 Loss 1.7011 Accuracy 0.1986
Epoch 5 Batch 200 Loss 1.6926 Accuracy 0.1996
Epoch 5 Batch 250 Loss 1.6899 Accuracy 0.2005
Epoch 5 Batch 300 Loss 1.6822 Accuracy 0.2009
Epoch 5 Batch 350 Loss 1.6765 Accuracy 0.2016
Epoch 5 Batch 400 Loss 1.6720 Accuracy 0.2021
Epoch 5 Batch 450 Loss 1.6636 Accuracy 0.2026
Epoch 5 Batch 500 Loss 1.6612 Accuracy 0.2035
Epoch 5 Batch 550 Loss 1.6616 Accuracy 0.2041
Epoch 5 Batch 600 Loss 1.6562 Accuracy 0.2047
Epoch 5 Batch 650 Loss 1.6525 Accuracy 0.2052
Epoch 5 Batch 700 Loss 1.6494 Accuracy 0.2060
Saving checkpoint for epoch 5 at ./checkpoints/train/ckpt-1
Epoch 5 Loss 1.6488 Accuracy 0.2060
Time taken for 1 epoch: 29.175358772277832 secs

Epoch 6 Batch 0 Loss 1.3306 Accuracy 0.2052
Epoch 6 Batch 50 Loss 1.4646 Accuracy 0.2208
Epoch 6 Batch 100 Loss 1.4786 Accuracy 0.2198
Epoch 6 Batch 150 Loss 1.4877 Accuracy 0.2201
Epoch 6 Batch 200 Loss 1.4891 Accuracy 0.2204
Epoch 6 Batch 250 Loss 1.4940 Accuracy 0.2212
Epoch 6 Batch 300 Loss 1.4899 Accuracy 0.2213
Epoch 6 Batch 350 Loss 1.4855 Accuracy 0.2215
Epoch 6 Batch 400 Loss 1.4846 Accuracy 0.2222
Epoch 6 Batch 450 Loss 1.4840 Accuracy 0.2230
Epoch 6 Batch 500 Loss 1.4829 Accuracy 0.2234
Epoch 6 Batch 550 Loss 1.4795 Accuracy 0.2239
Epoch 6 Batch 600 Loss 1.4773 Accuracy 0.2243
Epoch 6 Batch 650 Loss 1.4716 Accuracy 0.2249
Epoch 6 Batch 700 Loss 1.4686 Accuracy 0.2253
Epoch 6 Loss 1.4681 Accuracy 0.2253
Time taken for 1 epoch: 29.272542238235474 secs

Epoch 7 Batch 0 Loss 1.4585 Accuracy 0.2732
Epoch 7 Batch 50 Loss 1.3170 Accuracy 0.2431
Epoch 7 Batch 100 Loss 1.3139 Accuracy 0.2415
Epoch 7 Batch 150 Loss 1.3080 Accuracy 0.2411
Epoch 7 Batch 200 Loss 1.3091 Accuracy 0.2420
Epoch 7 Batch 250 Loss 1.3087 Accuracy 0.2428
Epoch 7 Batch 300 Loss 1.3036 Accuracy 0.2437
Epoch 7 Batch 350 Loss 1.3029 Accuracy 0.2443
Epoch 7 Batch 400 Loss 1.2994 Accuracy 0.2446
Epoch 7 Batch 450 Loss 1.2927 Accuracy 0.2445
Epoch 7 Batch 500 Loss 1.2904 Accuracy 0.2450
Epoch 7 Batch 550 Loss 1.2885 Accuracy 0.2453
Epoch 7 Batch 600 Loss 1.2859 Accuracy 0.2456
Epoch 7 Batch 650 Loss 1.2836 Accuracy 0.2462
Epoch 7 Batch 700 Loss 1.2821 Accuracy 0.2469
Epoch 7 Loss 1.2827 Accuracy 0.2469
Time taken for 1 epoch: 28.96388077735901 secs

Epoch 8 Batch 0 Loss 1.1522 Accuracy 0.2546
Epoch 8 Batch 50 Loss 1.1273 Accuracy 0.2609
Epoch 8 Batch 100 Loss 1.1276 Accuracy 0.2630
Epoch 8 Batch 150 Loss 1.1298 Accuracy 0.2622
Epoch 8 Batch 200 Loss 1.1315 Accuracy 0.2621
Epoch 8 Batch 250 Loss 1.1355 Accuracy 0.2629
Epoch 8 Batch 300 Loss 1.1353 Accuracy 0.2628
Epoch 8 Batch 350 Loss 1.1347 Accuracy 0.2630
Epoch 8 Batch 400 Loss 1.1351 Accuracy 0.2637
Epoch 8 Batch 450 Loss 1.1350 Accuracy 0.2643
Epoch 8 Batch 500 Loss 1.1334 Accuracy 0.2648
Epoch 8 Batch 550 Loss 1.1348 Accuracy 0.2655
Epoch 8 Batch 600 Loss 1.1334 Accuracy 0.2656
Epoch 8 Batch 650 Loss 1.1313 Accuracy 0.2654
Epoch 8 Batch 700 Loss 1.1291 Accuracy 0.2653
Epoch 8 Loss 1.1293 Accuracy 0.2654
Time taken for 1 epoch: 28.96521830558777 secs

Epoch 9 Batch 0 Loss 0.9535 Accuracy 0.2669
Epoch 9 Batch 50 Loss 0.9929 Accuracy 0.2816
Epoch 9 Batch 100 Loss 1.0012 Accuracy 0.2781
Epoch 9 Batch 150 Loss 1.0076 Accuracy 0.2785
Epoch 9 Batch 200 Loss 1.0088 Accuracy 0.2792
Epoch 9 Batch 250 Loss 1.0142 Accuracy 0.2797
Epoch 9 Batch 300 Loss 1.0171 Accuracy 0.2800
Epoch 9 Batch 350 Loss 1.0179 Accuracy 0.2797
Epoch 9 Batch 400 Loss 1.0233 Accuracy 0.2804
Epoch 9 Batch 450 Loss 1.0195 Accuracy 0.2797
Epoch 9 Batch 500 Loss 1.0191 Accuracy 0.2798
Epoch 9 Batch 550 Loss 1.0186 Accuracy 0.2797
Epoch 9 Batch 600 Loss 1.0185 Accuracy 0.2797
Epoch 9 Batch 650 Loss 1.0184 Accuracy 0.2794
Epoch 9 Batch 700 Loss 1.0196 Accuracy 0.2793
Epoch 9 Loss 1.0199 Accuracy 0.2793
Time taken for 1 epoch: 29.06449270248413 secs

Epoch 10 Batch 0 Loss 0.7134 Accuracy 0.2513
Epoch 10 Batch 50 Loss 0.9117 Accuracy 0.2895
Epoch 10 Batch 100 Loss 0.9109 Accuracy 0.2920
Epoch 10 Batch 150 Loss 0.9124 Accuracy 0.2920
Epoch 10 Batch 200 Loss 0.9152 Accuracy 0.2907
Epoch 10 Batch 250 Loss 0.9148 Accuracy 0.2906
Epoch 10 Batch 300 Loss 0.9232 Accuracy 0.2903
Epoch 10 Batch 350 Loss 0.9226 Accuracy 0.2900
Epoch 10 Batch 400 Loss 0.9287 Accuracy 0.2905
Epoch 10 Batch 450 Loss 0.9291 Accuracy 0.2907
Epoch 10 Batch 500 Loss 0.9306 Accuracy 0.2905
Epoch 10 Batch 550 Loss 0.9304 Accuracy 0.2903
Epoch 10 Batch 600 Loss 0.9327 Accuracy 0.2900
Epoch 10 Batch 650 Loss 0.9335 Accuracy 0.2898
Epoch 10 Batch 700 Loss 0.9354 Accuracy 0.2896
Saving checkpoint for epoch 10 at ./checkpoints/train/ckpt-2
Epoch 10 Loss 0.9355 Accuracy 0.2896
Time taken for 1 epoch: 29.15560221672058 secs

Epoch 11 Batch 0 Loss 0.9483 Accuracy 0.3076
Epoch 11 Batch 50 Loss 0.8396 Accuracy 0.2984
Epoch 11 Batch 100 Loss 0.8405 Accuracy 0.2997
Epoch 11 Batch 150 Loss 0.8427 Accuracy 0.2996
Epoch 11 Batch 200 Loss 0.8530 Accuracy 0.3005
Epoch 11 Batch 250 Loss 0.8553 Accuracy 0.3008
Epoch 11 Batch 300 Loss 0.8571 Accuracy 0.3008
Epoch 11 Batch 350 Loss 0.8604 Accuracy 0.3013
Epoch 11 Batch 400 Loss 0.8629 Accuracy 0.3008
Epoch 11 Batch 450 Loss 0.8619 Accuracy 0.3002
Epoch 11 Batch 500 Loss 0.8633 Accuracy 0.2998
Epoch 11 Batch 550 Loss 0.8653 Accuracy 0.2997
Epoch 11 Batch 600 Loss 0.8677 Accuracy 0.2993
Epoch 11 Batch 650 Loss 0.8687 Accuracy 0.2988
Epoch 11 Batch 700 Loss 0.8699 Accuracy 0.2986
Epoch 11 Loss 0.8698 Accuracy 0.2986
Time taken for 1 epoch: 29.991504430770874 secs

Epoch 12 Batch 0 Loss 0.7929 Accuracy 0.3203
Epoch 12 Batch 50 Loss 0.7796 Accuracy 0.3076
Epoch 12 Batch 100 Loss 0.7905 Accuracy 0.3089
Epoch 12 Batch 150 Loss 0.7917 Accuracy 0.3096
Epoch 12 Batch 200 Loss 0.7968 Accuracy 0.3089
Epoch 12 Batch 250 Loss 0.8009 Accuracy 0.3096
Epoch 12 Batch 300 Loss 0.8012 Accuracy 0.3089
Epoch 12 Batch 350 Loss 0.8031 Accuracy 0.3088
Epoch 12 Batch 400 Loss 0.8051 Accuracy 0.3083
Epoch 12 Batch 450 Loss 0.8048 Accuracy 0.3074
Epoch 12 Batch 500 Loss 0.8070 Accuracy 0.3072
Epoch 12 Batch 550 Loss 0.8074 Accuracy 0.3067
Epoch 12 Batch 600 Loss 0.8099 Accuracy 0.3064
Epoch 12 Batch 650 Loss 0.8120 Accuracy 0.3061
Epoch 12 Batch 700 Loss 0.8144 Accuracy 0.3060
Epoch 12 Loss 0.8145 Accuracy 0.3060
Time taken for 1 epoch: 28.987586498260498 secs

Epoch 13 Batch 0 Loss 0.7626 Accuracy 0.3447
Epoch 13 Batch 50 Loss 0.7175 Accuracy 0.3140
Epoch 13 Batch 100 Loss 0.7281 Accuracy 0.3151
Epoch 13 Batch 150 Loss 0.7330 Accuracy 0.3146
Epoch 13 Batch 200 Loss 0.7359 Accuracy 0.3138
Epoch 13 Batch 250 Loss 0.7416 Accuracy 0.3133
Epoch 13 Batch 300 Loss 0.7477 Accuracy 0.3144
Epoch 13 Batch 350 Loss 0.7486 Accuracy 0.3136
Epoch 13 Batch 400 Loss 0.7527 Accuracy 0.3131
Epoch 13 Batch 450 Loss 0.7539 Accuracy 0.3132
Epoch 13 Batch 500 Loss 0.7564 Accuracy 0.3132
Epoch 13 Batch 550 Loss 0.7593 Accuracy 0.3129
Epoch 13 Batch 600 Loss 0.7614 Accuracy 0.3127
Epoch 13 Batch 650 Loss 0.7629 Accuracy 0.3125
Epoch 13 Batch 700 Loss 0.7649 Accuracy 0.3120
Epoch 13 Loss 0.7652 Accuracy 0.3120
Time taken for 1 epoch: 28.90890407562256 secs

Epoch 14 Batch 0 Loss 0.6335 Accuracy 0.2993
Epoch 14 Batch 50 Loss 0.6933 Accuracy 0.3181
Epoch 14 Batch 100 Loss 0.6956 Accuracy 0.3211
Epoch 14 Batch 150 Loss 0.6893 Accuracy 0.3191
Epoch 14 Batch 200 Loss 0.6937 Accuracy 0.3195
Epoch 14 Batch 250 Loss 0.7009 Accuracy 0.3188
Epoch 14 Batch 300 Loss 0.7049 Accuracy 0.3187
Epoch 14 Batch 350 Loss 0.7082 Accuracy 0.3184
Epoch 14 Batch 400 Loss 0.7096 Accuracy 0.3184
Epoch 14 Batch 450 Loss 0.7126 Accuracy 0.3182
Epoch 14 Batch 500 Loss 0.7147 Accuracy 0.3181
Epoch 14 Batch 550 Loss 0.7180 Accuracy 0.3177
Epoch 14 Batch 600 Loss 0.7208 Accuracy 0.3181
Epoch 14 Batch 650 Loss 0.7237 Accuracy 0.3182
Epoch 14 Batch 700 Loss 0.7256 Accuracy 0.3176
Epoch 14 Loss 0.7260 Accuracy 0.3176
Time taken for 1 epoch: 29.05610466003418 secs

Epoch 15 Batch 0 Loss 0.6957 Accuracy 0.3268
Epoch 15 Batch 50 Loss 0.6385 Accuracy 0.3294
Epoch 15 Batch 100 Loss 0.6494 Accuracy 0.3269
Epoch 15 Batch 150 Loss 0.6562 Accuracy 0.3280
Epoch 15 Batch 200 Loss 0.6642 Accuracy 0.3276
Epoch 15 Batch 250 Loss 0.6658 Accuracy 0.3272
Epoch 15 Batch 300 Loss 0.6677 Accuracy 0.3267
Epoch 15 Batch 350 Loss 0.6719 Accuracy 0.3258
Epoch 15 Batch 400 Loss 0.6756 Accuracy 0.3259
Epoch 15 Batch 450 Loss 0.6780 Accuracy 0.3258
Epoch 15 Batch 500 Loss 0.6801 Accuracy 0.3258
Epoch 15 Batch 550 Loss 0.6827 Accuracy 0.3255
Epoch 15 Batch 600 Loss 0.6858 Accuracy 0.3249
Epoch 15 Batch 650 Loss 0.6875 Accuracy 0.3240
Epoch 15 Batch 700 Loss 0.6914 Accuracy 0.3237
Saving checkpoint for epoch 15 at ./checkpoints/train/ckpt-3
Epoch 15 Loss 0.6916 Accuracy 0.3237
Time taken for 1 epoch: 29.04577612876892 secs

Epoch 16 Batch 0 Loss 0.7174 Accuracy 0.3497
Epoch 16 Batch 50 Loss 0.6085 Accuracy 0.3328
Epoch 16 Batch 100 Loss 0.6190 Accuracy 0.3323
Epoch 16 Batch 150 Loss 0.6265 Accuracy 0.3317
Epoch 16 Batch 200 Loss 0.6285 Accuracy 0.3310
Epoch 16 Batch 250 Loss 0.6335 Accuracy 0.3306
Epoch 16 Batch 300 Loss 0.6365 Accuracy 0.3309
Epoch 16 Batch 350 Loss 0.6396 Accuracy 0.3308
Epoch 16 Batch 400 Loss 0.6414 Accuracy 0.3303
Epoch 16 Batch 450 Loss 0.6448 Accuracy 0.3298
Epoch 16 Batch 500 Loss 0.6480 Accuracy 0.3295
Epoch 16 Batch 550 Loss 0.6517 Accuracy 0.3290
Epoch 16 Batch 600 Loss 0.6553 Accuracy 0.3290
Epoch 16 Batch 650 Loss 0.6575 Accuracy 0.3289
Epoch 16 Batch 700 Loss 0.6602 Accuracy 0.3286
Epoch 16 Loss 0.6603 Accuracy 0.3285
Time taken for 1 epoch: 28.84425640106201 secs

Epoch 17 Batch 0 Loss 0.5961 Accuracy 0.3363
Epoch 17 Batch 50 Loss 0.5899 Accuracy 0.3356
Epoch 17 Batch 100 Loss 0.5933 Accuracy 0.3362
Epoch 17 Batch 150 Loss 0.5984 Accuracy 0.3347
Epoch 17 Batch 200 Loss 0.6019 Accuracy 0.3344
Epoch 17 Batch 250 Loss 0.6046 Accuracy 0.3336
Epoch 17 Batch 300 Loss 0.6079 Accuracy 0.3335
Epoch 17 Batch 350 Loss 0.6127 Accuracy 0.3340
Epoch 17 Batch 400 Loss 0.6144 Accuracy 0.3335
Epoch 17 Batch 450 Loss 0.6159 Accuracy 0.3330
Epoch 17 Batch 500 Loss 0.6185 Accuracy 0.3328
Epoch 17 Batch 550 Loss 0.6202 Accuracy 0.3323
Epoch 17 Batch 600 Loss 0.6246 Accuracy 0.3324
Epoch 17 Batch 650 Loss 0.6271 Accuracy 0.3322
Epoch 17 Batch 700 Loss 0.6291 Accuracy 0.3316
Epoch 17 Loss 0.6295 Accuracy 0.3316
Time taken for 1 epoch: 29.231011629104614 secs

Epoch 18 Batch 0 Loss 0.6079 Accuracy 0.3498
Epoch 18 Batch 50 Loss 0.5678 Accuracy 0.3420
Epoch 18 Batch 100 Loss 0.5657 Accuracy 0.3398
Epoch 18 Batch 150 Loss 0.5708 Accuracy 0.3389
Epoch 18 Batch 200 Loss 0.5718 Accuracy 0.3387
Epoch 18 Batch 250 Loss 0.5765 Accuracy 0.3384
Epoch 18 Batch 300 Loss 0.5786 Accuracy 0.3379
Epoch 18 Batch 350 Loss 0.5857 Accuracy 0.3388
Epoch 18 Batch 400 Loss 0.5882 Accuracy 0.3386
Epoch 18 Batch 450 Loss 0.5903 Accuracy 0.3383
Epoch 18 Batch 500 Loss 0.5937 Accuracy 0.3377
Epoch 18 Batch 550 Loss 0.5960 Accuracy 0.3372
Epoch 18 Batch 600 Loss 0.5991 Accuracy 0.3365
Epoch 18 Batch 650 Loss 0.6023 Accuracy 0.3360
Epoch 18 Batch 700 Loss 0.6047 Accuracy 0.3358
Epoch 18 Loss 0.6047 Accuracy 0.3358
Time taken for 1 epoch: 28.772484064102173 secs

Epoch 19 Batch 0 Loss 0.5326 Accuracy 0.3346
Epoch 19 Batch 50 Loss 0.5315 Accuracy 0.3395
Epoch 19 Batch 100 Loss 0.5400 Accuracy 0.3448
Epoch 19 Batch 150 Loss 0.5496 Accuracy 0.3458
Epoch 19 Batch 200 Loss 0.5535 Accuracy 0.3436
Epoch 19 Batch 250 Loss 0.5571 Accuracy 0.3430
Epoch 19 Batch 300 Loss 0.5605 Accuracy 0.3422
Epoch 19 Batch 350 Loss 0.5637 Accuracy 0.3431
Epoch 19 Batch 400 Loss 0.5661 Accuracy 0.3423
Epoch 19 Batch 450 Loss 0.5687 Accuracy 0.3420
Epoch 19 Batch 500 Loss 0.5705 Accuracy 0.3414
Epoch 19 Batch 550 Loss 0.5735 Accuracy 0.3407
Epoch 19 Batch 600 Loss 0.5763 Accuracy 0.3402
Epoch 19 Batch 650 Loss 0.5790 Accuracy 0.3398
Epoch 19 Batch 700 Loss 0.5819 Accuracy 0.3396
Epoch 19 Loss 0.5819 Accuracy 0.3396
Time taken for 1 epoch: 28.849220514297485 secs

Epoch 20 Batch 0 Loss 0.4288 Accuracy 0.3611
Epoch 20 Batch 50 Loss 0.5139 Accuracy 0.3533
Epoch 20 Batch 100 Loss 0.5171 Accuracy 0.3503
Epoch 20 Batch 150 Loss 0.5265 Accuracy 0.3479
Epoch 20 Batch 200 Loss 0.5310 Accuracy 0.3476
Epoch 20 Batch 250 Loss 0.5367 Accuracy 0.3477
Epoch 20 Batch 300 Loss 0.5386 Accuracy 0.3464
Epoch 20 Batch 350 Loss 0.5409 Accuracy 0.3464
Epoch 20 Batch 400 Loss 0.5453 Accuracy 0.3456
Epoch 20 Batch 450 Loss 0.5482 Accuracy 0.3451
Epoch 20 Batch 500 Loss 0.5500 Accuracy 0.3446
Epoch 20 Batch 550 Loss 0.5527 Accuracy 0.3442
Epoch 20 Batch 600 Loss 0.5563 Accuracy 0.3440
Epoch 20 Batch 650 Loss 0.5591 Accuracy 0.3435
Epoch 20 Batch 700 Loss 0.5613 Accuracy 0.3430
Saving checkpoint for epoch 20 at ./checkpoints/train/ckpt-4
Epoch 20 Loss 0.5615 Accuracy 0.3429
Time taken for 1 epoch: 29.063352823257446 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 solve the united states and this is actually solving it .
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 very quickly stories of some magic stuff , which came out .
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 n't have to .

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。