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...

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.incompleteR1I9OZ/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.incompleteR1I9OZ/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.incompleteR1I9OZ/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 4.3162 Accuracy 0.0004
Epoch 1 Batch 50 Loss 4.1944 Accuracy 0.0056
Epoch 1 Batch 100 Loss 4.1358 Accuracy 0.0162
Epoch 1 Batch 150 Loss 4.0919 Accuracy 0.0197
Epoch 1 Batch 200 Loss 4.0397 Accuracy 0.0215
Epoch 1 Batch 250 Loss 3.9647 Accuracy 0.0226
Epoch 1 Batch 300 Loss 3.8941 Accuracy 0.0234
Epoch 1 Batch 350 Loss 3.8098 Accuracy 0.0258
Epoch 1 Batch 400 Loss 3.7242 Accuracy 0.0296
Epoch 1 Batch 450 Loss 3.6491 Accuracy 0.0331
Epoch 1 Batch 500 Loss 3.5812 Accuracy 0.0362
Epoch 1 Batch 550 Loss 3.5255 Accuracy 0.0394
Epoch 1 Batch 600 Loss 3.4650 Accuracy 0.0431
Epoch 1 Batch 650 Loss 3.4120 Accuracy 0.0467
Epoch 1 Batch 700 Loss 3.3598 Accuracy 0.0502
Epoch 1 Loss 3.3575 Accuracy 0.0503
Time taken for 1 epoch: 60.18010640144348 secs

Epoch 2 Batch 0 Loss 2.6558 Accuracy 0.0891
Epoch 2 Batch 50 Loss 2.6002 Accuracy 0.1018
Epoch 2 Batch 100 Loss 2.5725 Accuracy 0.1050
Epoch 2 Batch 150 Loss 2.5387 Accuracy 0.1074
Epoch 2 Batch 200 Loss 2.5049 Accuracy 0.1095
Epoch 2 Batch 250 Loss 2.4916 Accuracy 0.1120
Epoch 2 Batch 300 Loss 2.4726 Accuracy 0.1139
Epoch 2 Batch 350 Loss 2.4599 Accuracy 0.1159
Epoch 2 Batch 400 Loss 2.4484 Accuracy 0.1176
Epoch 2 Batch 450 Loss 2.4314 Accuracy 0.1188
Epoch 2 Batch 500 Loss 2.4190 Accuracy 0.1202
Epoch 2 Batch 550 Loss 2.4056 Accuracy 0.1216
Epoch 2 Batch 600 Loss 2.3924 Accuracy 0.1228
Epoch 2 Batch 650 Loss 2.3792 Accuracy 0.1239
Epoch 2 Batch 700 Loss 2.3644 Accuracy 0.1250
Epoch 2 Loss 2.3643 Accuracy 0.1250
Time taken for 1 epoch: 32.23134970664978 secs

Epoch 3 Batch 0 Loss 2.0306 Accuracy 0.1398
Epoch 3 Batch 50 Loss 2.1126 Accuracy 0.1411
Epoch 3 Batch 100 Loss 2.1321 Accuracy 0.1428
Epoch 3 Batch 150 Loss 2.1516 Accuracy 0.1450
Epoch 3 Batch 200 Loss 2.1414 Accuracy 0.1452
Epoch 3 Batch 250 Loss 2.1353 Accuracy 0.1458
Epoch 3 Batch 300 Loss 2.1314 Accuracy 0.1461
Epoch 3 Batch 350 Loss 2.1312 Accuracy 0.1468
Epoch 3 Batch 400 Loss 2.1301 Accuracy 0.1475
Epoch 3 Batch 450 Loss 2.1256 Accuracy 0.1482
Epoch 3 Batch 500 Loss 2.1224 Accuracy 0.1489
Epoch 3 Batch 550 Loss 2.1196 Accuracy 0.1497
Epoch 3 Batch 600 Loss 2.1132 Accuracy 0.1505
Epoch 3 Batch 650 Loss 2.1064 Accuracy 0.1513
Epoch 3 Batch 700 Loss 2.0995 Accuracy 0.1520
Epoch 3 Loss 2.0988 Accuracy 0.1520
Time taken for 1 epoch: 32.252318143844604 secs

Epoch 4 Batch 0 Loss 2.1852 Accuracy 0.1576
Epoch 4 Batch 50 Loss 1.9393 Accuracy 0.1662
Epoch 4 Batch 100 Loss 1.9390 Accuracy 0.1669
Epoch 4 Batch 150 Loss 1.9428 Accuracy 0.1671
Epoch 4 Batch 200 Loss 1.9428 Accuracy 0.1672
Epoch 4 Batch 250 Loss 1.9409 Accuracy 0.1677
Epoch 4 Batch 300 Loss 1.9429 Accuracy 0.1689
Epoch 4 Batch 350 Loss 1.9361 Accuracy 0.1697
Epoch 4 Batch 400 Loss 1.9285 Accuracy 0.1709
Epoch 4 Batch 450 Loss 1.9226 Accuracy 0.1715
Epoch 4 Batch 500 Loss 1.9157 Accuracy 0.1726
Epoch 4 Batch 550 Loss 1.9066 Accuracy 0.1736
Epoch 4 Batch 600 Loss 1.8998 Accuracy 0.1748
Epoch 4 Batch 650 Loss 1.8937 Accuracy 0.1757
Epoch 4 Batch 700 Loss 1.8848 Accuracy 0.1765
Epoch 4 Loss 1.8847 Accuracy 0.1765
Time taken for 1 epoch: 32.29405498504639 secs

Epoch 5 Batch 0 Loss 1.5690 Accuracy 0.1785
Epoch 5 Batch 50 Loss 1.6999 Accuracy 0.1921
Epoch 5 Batch 100 Loss 1.6955 Accuracy 0.1932
Epoch 5 Batch 150 Loss 1.7020 Accuracy 0.1956
Epoch 5 Batch 200 Loss 1.6976 Accuracy 0.1964
Epoch 5 Batch 250 Loss 1.6952 Accuracy 0.1975
Epoch 5 Batch 300 Loss 1.6910 Accuracy 0.1981
Epoch 5 Batch 350 Loss 1.6882 Accuracy 0.1989
Epoch 5 Batch 400 Loss 1.6828 Accuracy 0.1997
Epoch 5 Batch 450 Loss 1.6783 Accuracy 0.2004
Epoch 5 Batch 500 Loss 1.6742 Accuracy 0.2010
Epoch 5 Batch 550 Loss 1.6703 Accuracy 0.2017
Epoch 5 Batch 600 Loss 1.6641 Accuracy 0.2023
Epoch 5 Batch 650 Loss 1.6612 Accuracy 0.2031
Epoch 5 Batch 700 Loss 1.6577 Accuracy 0.2040
Saving checkpoint for epoch 5 at ./checkpoints/train/ckpt-1
Epoch 5 Loss 1.6574 Accuracy 0.2040
Time taken for 1 epoch: 32.73478412628174 secs

Epoch 6 Batch 0 Loss 1.3764 Accuracy 0.2124
Epoch 6 Batch 50 Loss 1.5054 Accuracy 0.2177
Epoch 6 Batch 100 Loss 1.4844 Accuracy 0.2185
Epoch 6 Batch 150 Loss 1.4878 Accuracy 0.2183
Epoch 6 Batch 200 Loss 1.4960 Accuracy 0.2192
Epoch 6 Batch 250 Loss 1.5045 Accuracy 0.2200
Epoch 6 Batch 300 Loss 1.4970 Accuracy 0.2198
Epoch 6 Batch 350 Loss 1.4967 Accuracy 0.2206
Epoch 6 Batch 400 Loss 1.4922 Accuracy 0.2214
Epoch 6 Batch 450 Loss 1.4872 Accuracy 0.2220
Epoch 6 Batch 500 Loss 1.4862 Accuracy 0.2226
Epoch 6 Batch 550 Loss 1.4839 Accuracy 0.2233
Epoch 6 Batch 600 Loss 1.4809 Accuracy 0.2235
Epoch 6 Batch 650 Loss 1.4773 Accuracy 0.2241
Epoch 6 Batch 700 Loss 1.4735 Accuracy 0.2246
Epoch 6 Loss 1.4737 Accuracy 0.2247
Time taken for 1 epoch: 32.320966720581055 secs

Epoch 7 Batch 0 Loss 1.2604 Accuracy 0.2404
Epoch 7 Batch 50 Loss 1.3355 Accuracy 0.2428
Epoch 7 Batch 100 Loss 1.3115 Accuracy 0.2397
Epoch 7 Batch 150 Loss 1.3063 Accuracy 0.2409
Epoch 7 Batch 200 Loss 1.3027 Accuracy 0.2411
Epoch 7 Batch 250 Loss 1.3025 Accuracy 0.2413
Epoch 7 Batch 300 Loss 1.3012 Accuracy 0.2413
Epoch 7 Batch 350 Loss 1.3024 Accuracy 0.2421
Epoch 7 Batch 400 Loss 1.3001 Accuracy 0.2431
Epoch 7 Batch 450 Loss 1.2962 Accuracy 0.2428
Epoch 7 Batch 500 Loss 1.2953 Accuracy 0.2439
Epoch 7 Batch 550 Loss 1.2939 Accuracy 0.2442
Epoch 7 Batch 600 Loss 1.2921 Accuracy 0.2443
Epoch 7 Batch 650 Loss 1.2910 Accuracy 0.2448
Epoch 7 Batch 700 Loss 1.2887 Accuracy 0.2454
Epoch 7 Loss 1.2888 Accuracy 0.2455
Time taken for 1 epoch: 32.23518991470337 secs

Epoch 8 Batch 0 Loss 1.1231 Accuracy 0.2460
Epoch 8 Batch 50 Loss 1.1360 Accuracy 0.2581
Epoch 8 Batch 100 Loss 1.1280 Accuracy 0.2580
Epoch 8 Batch 150 Loss 1.1292 Accuracy 0.2595
Epoch 8 Batch 200 Loss 1.1318 Accuracy 0.2596
Epoch 8 Batch 250 Loss 1.1359 Accuracy 0.2603
Epoch 8 Batch 300 Loss 1.1391 Accuracy 0.2613
Epoch 8 Batch 350 Loss 1.1354 Accuracy 0.2617
Epoch 8 Batch 400 Loss 1.1339 Accuracy 0.2618
Epoch 8 Batch 450 Loss 1.1353 Accuracy 0.2626
Epoch 8 Batch 500 Loss 1.1350 Accuracy 0.2631
Epoch 8 Batch 550 Loss 1.1355 Accuracy 0.2635
Epoch 8 Batch 600 Loss 1.1345 Accuracy 0.2634
Epoch 8 Batch 650 Loss 1.1342 Accuracy 0.2639
Epoch 8 Batch 700 Loss 1.1322 Accuracy 0.2641
Epoch 8 Loss 1.1324 Accuracy 0.2641
Time taken for 1 epoch: 33.407273292541504 secs

Epoch 9 Batch 0 Loss 0.9370 Accuracy 0.2665
Epoch 9 Batch 50 Loss 0.9899 Accuracy 0.2819
Epoch 9 Batch 100 Loss 0.9961 Accuracy 0.2793
Epoch 9 Batch 150 Loss 1.0060 Accuracy 0.2791
Epoch 9 Batch 200 Loss 1.0125 Accuracy 0.2778
Epoch 9 Batch 250 Loss 1.0172 Accuracy 0.2790
Epoch 9 Batch 300 Loss 1.0182 Accuracy 0.2788
Epoch 9 Batch 350 Loss 1.0187 Accuracy 0.2785
Epoch 9 Batch 400 Loss 1.0169 Accuracy 0.2790
Epoch 9 Batch 450 Loss 1.0165 Accuracy 0.2788
Epoch 9 Batch 500 Loss 1.0186 Accuracy 0.2789
Epoch 9 Batch 550 Loss 1.0181 Accuracy 0.2787
Epoch 9 Batch 600 Loss 1.0189 Accuracy 0.2786
Epoch 9 Batch 650 Loss 1.0207 Accuracy 0.2787
Epoch 9 Batch 700 Loss 1.0216 Accuracy 0.2786
Epoch 9 Loss 1.0215 Accuracy 0.2786
Time taken for 1 epoch: 32.33360838890076 secs

Epoch 10 Batch 0 Loss 0.8580 Accuracy 0.3137
Epoch 10 Batch 50 Loss 0.8910 Accuracy 0.2901
Epoch 10 Batch 100 Loss 0.9128 Accuracy 0.2926
Epoch 10 Batch 150 Loss 0.9204 Accuracy 0.2925
Epoch 10 Batch 200 Loss 0.9194 Accuracy 0.2911
Epoch 10 Batch 250 Loss 0.9241 Accuracy 0.2912
Epoch 10 Batch 300 Loss 0.9271 Accuracy 0.2907
Epoch 10 Batch 350 Loss 0.9274 Accuracy 0.2910
Epoch 10 Batch 400 Loss 0.9283 Accuracy 0.2914
Epoch 10 Batch 450 Loss 0.9311 Accuracy 0.2913
Epoch 10 Batch 500 Loss 0.9341 Accuracy 0.2916
Epoch 10 Batch 550 Loss 0.9343 Accuracy 0.2912
Epoch 10 Batch 600 Loss 0.9367 Accuracy 0.2911
Epoch 10 Batch 650 Loss 0.9378 Accuracy 0.2908
Epoch 10 Batch 700 Loss 0.9378 Accuracy 0.2901
Saving checkpoint for epoch 10 at ./checkpoints/train/ckpt-2
Epoch 10 Loss 0.9381 Accuracy 0.2902
Time taken for 1 epoch: 32.44796705245972 secs

Epoch 11 Batch 0 Loss 0.8357 Accuracy 0.3234
Epoch 11 Batch 50 Loss 0.8494 Accuracy 0.3049
Epoch 11 Batch 100 Loss 0.8550 Accuracy 0.3052
Epoch 11 Batch 150 Loss 0.8533 Accuracy 0.3039
Epoch 11 Batch 200 Loss 0.8511 Accuracy 0.3021
Epoch 11 Batch 250 Loss 0.8516 Accuracy 0.3009
Epoch 11 Batch 300 Loss 0.8543 Accuracy 0.3004
Epoch 11 Batch 350 Loss 0.8582 Accuracy 0.3009
Epoch 11 Batch 400 Loss 0.8595 Accuracy 0.3004
Epoch 11 Batch 450 Loss 0.8607 Accuracy 0.3000
Epoch 11 Batch 500 Loss 0.8617 Accuracy 0.2997
Epoch 11 Batch 550 Loss 0.8627 Accuracy 0.2992
Epoch 11 Batch 600 Loss 0.8640 Accuracy 0.2989
Epoch 11 Batch 650 Loss 0.8655 Accuracy 0.2984
Epoch 11 Batch 700 Loss 0.8674 Accuracy 0.2983
Epoch 11 Loss 0.8676 Accuracy 0.2983
Time taken for 1 epoch: 32.30225467681885 secs

Epoch 12 Batch 0 Loss 0.7255 Accuracy 0.3028
Epoch 12 Batch 50 Loss 0.7672 Accuracy 0.3104
Epoch 12 Batch 100 Loss 0.7786 Accuracy 0.3116
Epoch 12 Batch 150 Loss 0.7808 Accuracy 0.3097
Epoch 12 Batch 200 Loss 0.7873 Accuracy 0.3079
Epoch 12 Batch 250 Loss 0.7896 Accuracy 0.3074
Epoch 12 Batch 300 Loss 0.7940 Accuracy 0.3077
Epoch 12 Batch 350 Loss 0.7968 Accuracy 0.3074
Epoch 12 Batch 400 Loss 0.7995 Accuracy 0.3072
Epoch 12 Batch 450 Loss 0.8032 Accuracy 0.3074
Epoch 12 Batch 500 Loss 0.8043 Accuracy 0.3069
Epoch 12 Batch 550 Loss 0.8079 Accuracy 0.3071
Epoch 12 Batch 600 Loss 0.8100 Accuracy 0.3067
Epoch 12 Batch 650 Loss 0.8110 Accuracy 0.3063
Epoch 12 Batch 700 Loss 0.8117 Accuracy 0.3061
Epoch 12 Loss 0.8118 Accuracy 0.3061
Time taken for 1 epoch: 32.21675634384155 secs

Epoch 13 Batch 0 Loss 0.7464 Accuracy 0.3542
Epoch 13 Batch 50 Loss 0.7193 Accuracy 0.3206
Epoch 13 Batch 100 Loss 0.7200 Accuracy 0.3161
Epoch 13 Batch 150 Loss 0.7290 Accuracy 0.3145
Epoch 13 Batch 200 Loss 0.7343 Accuracy 0.3149
Epoch 13 Batch 250 Loss 0.7405 Accuracy 0.3143
Epoch 13 Batch 300 Loss 0.7442 Accuracy 0.3137
Epoch 13 Batch 350 Loss 0.7494 Accuracy 0.3141
Epoch 13 Batch 400 Loss 0.7506 Accuracy 0.3136
Epoch 13 Batch 450 Loss 0.7514 Accuracy 0.3129
Epoch 13 Batch 500 Loss 0.7543 Accuracy 0.3128
Epoch 13 Batch 550 Loss 0.7556 Accuracy 0.3122
Epoch 13 Batch 600 Loss 0.7581 Accuracy 0.3123
Epoch 13 Batch 650 Loss 0.7626 Accuracy 0.3127
Epoch 13 Batch 700 Loss 0.7644 Accuracy 0.3123
Epoch 13 Loss 0.7646 Accuracy 0.3123
Time taken for 1 epoch: 32.18987679481506 secs

Epoch 14 Batch 0 Loss 0.7860 Accuracy 0.3455
Epoch 14 Batch 50 Loss 0.6748 Accuracy 0.3255
Epoch 14 Batch 100 Loss 0.6815 Accuracy 0.3232
Epoch 14 Batch 150 Loss 0.6868 Accuracy 0.3214
Epoch 14 Batch 200 Loss 0.6935 Accuracy 0.3208
Epoch 14 Batch 250 Loss 0.6978 Accuracy 0.3212
Epoch 14 Batch 300 Loss 0.7009 Accuracy 0.3207
Epoch 14 Batch 350 Loss 0.7040 Accuracy 0.3201
Epoch 14 Batch 400 Loss 0.7076 Accuracy 0.3203
Epoch 14 Batch 450 Loss 0.7104 Accuracy 0.3201
Epoch 14 Batch 500 Loss 0.7137 Accuracy 0.3194
Epoch 14 Batch 550 Loss 0.7161 Accuracy 0.3188
Epoch 14 Batch 600 Loss 0.7183 Accuracy 0.3186
Epoch 14 Batch 650 Loss 0.7219 Accuracy 0.3185
Epoch 14 Batch 700 Loss 0.7246 Accuracy 0.3184
Epoch 14 Loss 0.7244 Accuracy 0.3184
Time taken for 1 epoch: 32.628119230270386 secs

Epoch 15 Batch 0 Loss 0.6015 Accuracy 0.3177
Epoch 15 Batch 50 Loss 0.6422 Accuracy 0.3268
Epoch 15 Batch 100 Loss 0.6442 Accuracy 0.3274
Epoch 15 Batch 150 Loss 0.6517 Accuracy 0.3276
Epoch 15 Batch 200 Loss 0.6585 Accuracy 0.3266
Epoch 15 Batch 250 Loss 0.6637 Accuracy 0.3269
Epoch 15 Batch 300 Loss 0.6671 Accuracy 0.3263
Epoch 15 Batch 350 Loss 0.6697 Accuracy 0.3260
Epoch 15 Batch 400 Loss 0.6710 Accuracy 0.3259
Epoch 15 Batch 450 Loss 0.6730 Accuracy 0.3254
Epoch 15 Batch 500 Loss 0.6768 Accuracy 0.3253
Epoch 15 Batch 550 Loss 0.6799 Accuracy 0.3250
Epoch 15 Batch 600 Loss 0.6822 Accuracy 0.3246
Epoch 15 Batch 650 Loss 0.6849 Accuracy 0.3243
Epoch 15 Batch 700 Loss 0.6866 Accuracy 0.3239
Saving checkpoint for epoch 15 at ./checkpoints/train/ckpt-3
Epoch 15 Loss 0.6865 Accuracy 0.3239
Time taken for 1 epoch: 32.443984508514404 secs

Epoch 16 Batch 0 Loss 0.6423 Accuracy 0.3691
Epoch 16 Batch 50 Loss 0.6121 Accuracy 0.3350
Epoch 16 Batch 100 Loss 0.6199 Accuracy 0.3349
Epoch 16 Batch 150 Loss 0.6231 Accuracy 0.3327
Epoch 16 Batch 200 Loss 0.6259 Accuracy 0.3316
Epoch 16 Batch 250 Loss 0.6290 Accuracy 0.3309
Epoch 16 Batch 300 Loss 0.6336 Accuracy 0.3300
Epoch 16 Batch 350 Loss 0.6368 Accuracy 0.3297
Epoch 16 Batch 400 Loss 0.6387 Accuracy 0.3295
Epoch 16 Batch 450 Loss 0.6412 Accuracy 0.3296
Epoch 16 Batch 500 Loss 0.6445 Accuracy 0.3297
Epoch 16 Batch 550 Loss 0.6482 Accuracy 0.3295
Epoch 16 Batch 600 Loss 0.6517 Accuracy 0.3291
Epoch 16 Batch 650 Loss 0.6538 Accuracy 0.3285
Epoch 16 Batch 700 Loss 0.6570 Accuracy 0.3285
Epoch 16 Loss 0.6573 Accuracy 0.3285
Time taken for 1 epoch: 32.34228706359863 secs

Epoch 17 Batch 0 Loss 0.6586 Accuracy 0.3391
Epoch 17 Batch 50 Loss 0.5902 Accuracy 0.3425
Epoch 17 Batch 100 Loss 0.5910 Accuracy 0.3389
Epoch 17 Batch 150 Loss 0.5877 Accuracy 0.3359
Epoch 17 Batch 200 Loss 0.5967 Accuracy 0.3354
Epoch 17 Batch 250 Loss 0.6004 Accuracy 0.3350
Epoch 17 Batch 300 Loss 0.6052 Accuracy 0.3353
Epoch 17 Batch 350 Loss 0.6072 Accuracy 0.3345
Epoch 17 Batch 400 Loss 0.6098 Accuracy 0.3340
Epoch 17 Batch 450 Loss 0.6123 Accuracy 0.3339
Epoch 17 Batch 500 Loss 0.6152 Accuracy 0.3339
Epoch 17 Batch 550 Loss 0.6187 Accuracy 0.3339
Epoch 17 Batch 600 Loss 0.6221 Accuracy 0.3337
Epoch 17 Batch 650 Loss 0.6239 Accuracy 0.3331
Epoch 17 Batch 700 Loss 0.6257 Accuracy 0.3324
Epoch 17 Loss 0.6259 Accuracy 0.3325
Time taken for 1 epoch: 32.26639461517334 secs

Epoch 18 Batch 0 Loss 0.5910 Accuracy 0.3516
Epoch 18 Batch 50 Loss 0.5665 Accuracy 0.3441
Epoch 18 Batch 100 Loss 0.5615 Accuracy 0.3397
Epoch 18 Batch 150 Loss 0.5688 Accuracy 0.3398
Epoch 18 Batch 200 Loss 0.5713 Accuracy 0.3394
Epoch 18 Batch 250 Loss 0.5751 Accuracy 0.3396
Epoch 18 Batch 300 Loss 0.5807 Accuracy 0.3400
Epoch 18 Batch 350 Loss 0.5833 Accuracy 0.3389
Epoch 18 Batch 400 Loss 0.5869 Accuracy 0.3390
Epoch 18 Batch 450 Loss 0.5891 Accuracy 0.3383
Epoch 18 Batch 500 Loss 0.5923 Accuracy 0.3381
Epoch 18 Batch 550 Loss 0.5938 Accuracy 0.3379
Epoch 18 Batch 600 Loss 0.5966 Accuracy 0.3375
Epoch 18 Batch 650 Loss 0.5999 Accuracy 0.3374
Epoch 18 Batch 700 Loss 0.6015 Accuracy 0.3366
Epoch 18 Loss 0.6017 Accuracy 0.3365
Time taken for 1 epoch: 33.36975908279419 secs

Epoch 19 Batch 0 Loss 0.5694 Accuracy 0.3543
Epoch 19 Batch 50 Loss 0.5347 Accuracy 0.3441
Epoch 19 Batch 100 Loss 0.5426 Accuracy 0.3432
Epoch 19 Batch 150 Loss 0.5405 Accuracy 0.3404
Epoch 19 Batch 200 Loss 0.5470 Accuracy 0.3415
Epoch 19 Batch 250 Loss 0.5504 Accuracy 0.3414
Epoch 19 Batch 300 Loss 0.5549 Accuracy 0.3416
Epoch 19 Batch 350 Loss 0.5562 Accuracy 0.3411
Epoch 19 Batch 400 Loss 0.5594 Accuracy 0.3408
Epoch 19 Batch 450 Loss 0.5613 Accuracy 0.3404
Epoch 19 Batch 500 Loss 0.5649 Accuracy 0.3397
Epoch 19 Batch 550 Loss 0.5681 Accuracy 0.3396
Epoch 19 Batch 600 Loss 0.5708 Accuracy 0.3393
Epoch 19 Batch 650 Loss 0.5735 Accuracy 0.3389
Epoch 19 Batch 700 Loss 0.5779 Accuracy 0.3394
Epoch 19 Loss 0.5782 Accuracy 0.3394
Time taken for 1 epoch: 32.306732177734375 secs

Epoch 20 Batch 0 Loss 0.5892 Accuracy 0.3889
Epoch 20 Batch 50 Loss 0.5234 Accuracy 0.3534
Epoch 20 Batch 100 Loss 0.5216 Accuracy 0.3501
Epoch 20 Batch 150 Loss 0.5226 Accuracy 0.3470
Epoch 20 Batch 200 Loss 0.5273 Accuracy 0.3463
Epoch 20 Batch 250 Loss 0.5345 Accuracy 0.3466
Epoch 20 Batch 300 Loss 0.5380 Accuracy 0.3472
Epoch 20 Batch 350 Loss 0.5419 Accuracy 0.3466
Epoch 20 Batch 400 Loss 0.5415 Accuracy 0.3461
Epoch 20 Batch 450 Loss 0.5447 Accuracy 0.3462
Epoch 20 Batch 500 Loss 0.5473 Accuracy 0.3460
Epoch 20 Batch 550 Loss 0.5485 Accuracy 0.3455
Epoch 20 Batch 600 Loss 0.5516 Accuracy 0.3451
Epoch 20 Batch 650 Loss 0.5548 Accuracy 0.3447
Epoch 20 Batch 700 Loss 0.5586 Accuracy 0.3446
Saving checkpoint for epoch 20 at ./checkpoints/train/ckpt-4
Epoch 20 Loss 0.5585 Accuracy 0.3446
Time taken for 1 epoch: 32.4241840839386 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 we have to deal with the united states .
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 about this idea here .
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 some stories of some magic things that happened to happen .
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 .

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。