基于tensorflow2的seq2seq英语葡萄牙语翻译
Copyright 2019 The TensorFlow Authors.
#@title Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
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 |
Note: 我们的 TensorFlow 社区翻译了这些文档。因为社区翻译是尽力而为， 所以无法保证它们是最准确的，并且反映了最新的
官方英文文档。如果您有改进此翻译的建议， 请提交 pull request 到
tensorflow/docs GitHub 仓库。要志愿地撰写或者审核译文，请加入
[email protected] Google Group
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）和 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）和多头注意力（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 RNNs 或 CNNs 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。
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']
[1mDownloading 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...[0m
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.incompleteLD5ZFM/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.incompleteLD5ZFM/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.incompleteLD5ZFM/ted_hrlr_translate-test.tfrecord
HBox(children=(FloatProgress(value=0.0, max=1803.0), HTML(value='')))
[1mDataset 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.[0m
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
(,
)
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：
P E ( p o s , 2 i ) = s i n ( p o s / 1000 0 2 i / d m o d e l ) \Large{PE_{(pos, 2i)} = sin(pos / 10000^{2i / d_{model}})} PE(pos,2i)=sin(pos/100002i/dmodel)
P E ( p o s , 2 i + 1 ) = c o s ( p o s / 1000 0 2 i / d m o d e l ) \Large{PE_{(pos, 2i+1)} = cos(pos / 10000^{2i / d_{model}})} PE(pos,2i+1)=cos(pos/100002i/dmodel)
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)
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)
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
按比缩放的点积注意力（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：
A t t e n t i o n ( Q , K , V ) = s o f t m a x k ( Q K T d k ) V \Large{Attention(Q, K, V) = softmax_k(\frac{QK^T}{\sqrt{d_k}}) V} Attention(Q,K,V)=softmaxk(dkQKT)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 Q 和 K 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，Q 和 K 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）
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.transpose 和 tf.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 y，MultiHeadAttention 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 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：
- 多头注意力（有填充遮挡）
- 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：
- 遮挡的多头注意力（前瞻遮挡和填充遮挡）
- 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。
- 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：
- 输入嵌入（Input Embedding）
- 位置编码（Positional Encoding）
- 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：
- 输出嵌入（Output Embedding）
- 位置编码（Positional Encoding）
- 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=6，d_model = 512，dff = 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。
l r a t e = d m o d e l − 0.5 ∗ m i n ( s t e p _ n u m − 0.5 , s t e p _ n u m ∗ w a r m u p _ s t e p s − 1.5 ) \Large{lrate = d_{model}^{-0.5} * min(step{\_}num^{-0.5}, step{\_}num * warmup{\_}steps^{-1.5})} lrate=dmodel−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')
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 3.9447 Accuracy 0.0000
Epoch 1 Batch 50 Loss 4.1703 Accuracy 0.0002
Epoch 1 Batch 100 Loss 4.1667 Accuracy 0.0098
Epoch 1 Batch 150 Loss 4.1467 Accuracy 0.0155
Epoch 1 Batch 200 Loss 4.0751 Accuracy 0.0184
Epoch 1 Batch 250 Loss 4.0015 Accuracy 0.0202
Epoch 1 Batch 300 Loss 3.9277 Accuracy 0.0229
Epoch 1 Batch 350 Loss 3.8337 Accuracy 0.0269
Epoch 1 Batch 400 Loss 3.7421 Accuracy 0.0303
Epoch 1 Batch 450 Loss 3.6652 Accuracy 0.0337
Epoch 1 Batch 500 Loss 3.5967 Accuracy 0.0370
Epoch 1 Batch 550 Loss 3.5371 Accuracy 0.0407
Epoch 1 Batch 600 Loss 3.4794 Accuracy 0.0442
Epoch 1 Batch 650 Loss 3.4204 Accuracy 0.0476
Epoch 1 Batch 700 Loss 3.3677 Accuracy 0.0511
Epoch 1 Loss 3.3660 Accuracy 0.0512
Time taken for 1 epoch: 53.15570688247681 secs
Epoch 2 Batch 0 Loss 2.4368 Accuracy 0.1022
Epoch 2 Batch 50 Loss 2.5361 Accuracy 0.1028
Epoch 2 Batch 100 Loss 2.5307 Accuracy 0.1056
Epoch 2 Batch 150 Loss 2.5100 Accuracy 0.1077
Epoch 2 Batch 200 Loss 2.4854 Accuracy 0.1096
Epoch 2 Batch 250 Loss 2.4685 Accuracy 0.1113
Epoch 2 Batch 300 Loss 2.4608 Accuracy 0.1132
Epoch 2 Batch 350 Loss 2.4484 Accuracy 0.1151
Epoch 2 Batch 400 Loss 2.4382 Accuracy 0.1169
Epoch 2 Batch 450 Loss 2.4220 Accuracy 0.1184
Epoch 2 Batch 500 Loss 2.4061 Accuracy 0.1197
Epoch 2 Batch 550 Loss 2.3948 Accuracy 0.1210
Epoch 2 Batch 600 Loss 2.3810 Accuracy 0.1221
Epoch 2 Batch 650 Loss 2.3723 Accuracy 0.1234
Epoch 2 Batch 700 Loss 2.3616 Accuracy 0.1245
Epoch 2 Loss 2.3614 Accuracy 0.1245
Time taken for 1 epoch: 29.333844900131226 secs
Epoch 3 Batch 0 Loss 2.1795 Accuracy 0.1313
Epoch 3 Batch 50 Loss 2.1398 Accuracy 0.1399
Epoch 3 Batch 100 Loss 2.1483 Accuracy 0.1412
Epoch 3 Batch 150 Loss 2.1504 Accuracy 0.1423
Epoch 3 Batch 200 Loss 2.1522 Accuracy 0.1437
Epoch 3 Batch 250 Loss 2.1452 Accuracy 0.1440
Epoch 3 Batch 300 Loss 2.1435 Accuracy 0.1446
Epoch 3 Batch 350 Loss 2.1466 Accuracy 0.1455
Epoch 3 Batch 400 Loss 2.1421 Accuracy 0.1462
Epoch 3 Batch 450 Loss 2.1412 Accuracy 0.1469
Epoch 3 Batch 500 Loss 2.1341 Accuracy 0.1475
Epoch 3 Batch 550 Loss 2.1280 Accuracy 0.1482
Epoch 3 Batch 600 Loss 2.1212 Accuracy 0.1487
Epoch 3 Batch 650 Loss 2.1137 Accuracy 0.1494
Epoch 3 Batch 700 Loss 2.1060 Accuracy 0.1502
Epoch 3 Loss 2.1058 Accuracy 0.1502
Time taken for 1 epoch: 29.96753430366516 secs
Epoch 4 Batch 0 Loss 1.8913 Accuracy 0.1663
Epoch 4 Batch 50 Loss 1.9372 Accuracy 0.1635
Epoch 4 Batch 100 Loss 1.9416 Accuracy 0.1641
Epoch 4 Batch 150 Loss 1.9398 Accuracy 0.1656
Epoch 4 Batch 200 Loss 1.9356 Accuracy 0.1665
Epoch 4 Batch 250 Loss 1.9355 Accuracy 0.1672
Epoch 4 Batch 300 Loss 1.9296 Accuracy 0.1686
Epoch 4 Batch 350 Loss 1.9230 Accuracy 0.1699
Epoch 4 Batch 400 Loss 1.9139 Accuracy 0.1708
Epoch 4 Batch 450 Loss 1.9070 Accuracy 0.1720
Epoch 4 Batch 500 Loss 1.8999 Accuracy 0.1728
Epoch 4 Batch 550 Loss 1.8931 Accuracy 0.1740
Epoch 4 Batch 600 Loss 1.8871 Accuracy 0.1750
Epoch 4 Batch 650 Loss 1.8810 Accuracy 0.1761
Epoch 4 Batch 700 Loss 1.8734 Accuracy 0.1772
Epoch 4 Loss 1.8733 Accuracy 0.1772
Time taken for 1 epoch: 30.745400428771973 secs
Epoch 5 Batch 0 Loss 1.8901 Accuracy 0.2141
Epoch 5 Batch 50 Loss 1.7113 Accuracy 0.1979
Epoch 5 Batch 100 Loss 1.7037 Accuracy 0.1979
Epoch 5 Batch 150 Loss 1.6968 Accuracy 0.1981
Epoch 5 Batch 200 Loss 1.6915 Accuracy 0.1983
Epoch 5 Batch 250 Loss 1.6840 Accuracy 0.1982
Epoch 5 Batch 300 Loss 1.6846 Accuracy 0.1996
Epoch 5 Batch 350 Loss 1.6755 Accuracy 0.1996
Epoch 5 Batch 400 Loss 1.6708 Accuracy 0.2006
Epoch 5 Batch 450 Loss 1.6663 Accuracy 0.2017
Epoch 5 Batch 500 Loss 1.6641 Accuracy 0.2020
Epoch 5 Batch 550 Loss 1.6595 Accuracy 0.2024
Epoch 5 Batch 600 Loss 1.6560 Accuracy 0.2031
Epoch 5 Batch 650 Loss 1.6519 Accuracy 0.2041
Epoch 5 Batch 700 Loss 1.6484 Accuracy 0.2049
Saving checkpoint for epoch 5 at ./checkpoints/train/ckpt-1
Epoch 5 Loss 1.6480 Accuracy 0.2049
Time taken for 1 epoch: 29.267417192459106 secs
Epoch 6 Batch 0 Loss 1.4895 Accuracy 0.2146
Epoch 6 Batch 50 Loss 1.4947 Accuracy 0.2205
Epoch 6 Batch 100 Loss 1.4961 Accuracy 0.2210
Epoch 6 Batch 150 Loss 1.4943 Accuracy 0.2208
Epoch 6 Batch 200 Loss 1.4903 Accuracy 0.2208
Epoch 6 Batch 250 Loss 1.4920 Accuracy 0.2215
Epoch 6 Batch 300 Loss 1.4863 Accuracy 0.2219
Epoch 6 Batch 350 Loss 1.4844 Accuracy 0.2225
Epoch 6 Batch 400 Loss 1.4808 Accuracy 0.2229
Epoch 6 Batch 450 Loss 1.4784 Accuracy 0.2230
Epoch 6 Batch 500 Loss 1.4738 Accuracy 0.2232
Epoch 6 Batch 550 Loss 1.4718 Accuracy 0.2233
Epoch 6 Batch 600 Loss 1.4679 Accuracy 0.2239
Epoch 6 Batch 650 Loss 1.4656 Accuracy 0.2244
Epoch 6 Batch 700 Loss 1.4633 Accuracy 0.2250
Epoch 6 Loss 1.4636 Accuracy 0.2250
Time taken for 1 epoch: 29.181689739227295 secs
Epoch 7 Batch 0 Loss 1.3316 Accuracy 0.2256
Epoch 7 Batch 50 Loss 1.3221 Accuracy 0.2444
Epoch 7 Batch 100 Loss 1.3067 Accuracy 0.2428
Epoch 7 Batch 150 Loss 1.3019 Accuracy 0.2415
Epoch 7 Batch 200 Loss 1.2991 Accuracy 0.2413
Epoch 7 Batch 250 Loss 1.2978 Accuracy 0.2419
Epoch 7 Batch 300 Loss 1.2954 Accuracy 0.2425
Epoch 7 Batch 350 Loss 1.2976 Accuracy 0.2436
Epoch 7 Batch 400 Loss 1.2941 Accuracy 0.2437
Epoch 7 Batch 450 Loss 1.2935 Accuracy 0.2440
Epoch 7 Batch 500 Loss 1.2915 Accuracy 0.2449
Epoch 7 Batch 550 Loss 1.2902 Accuracy 0.2455
Epoch 7 Batch 600 Loss 1.2874 Accuracy 0.2456
Epoch 7 Batch 650 Loss 1.2869 Accuracy 0.2461
Epoch 7 Batch 700 Loss 1.2831 Accuracy 0.2465
Epoch 7 Loss 1.2829 Accuracy 0.2465
Time taken for 1 epoch: 29.13695764541626 secs
Epoch 8 Batch 0 Loss 1.0755 Accuracy 0.2821
Epoch 8 Batch 50 Loss 1.1188 Accuracy 0.2618
Epoch 8 Batch 100 Loss 1.1192 Accuracy 0.2625
Epoch 8 Batch 150 Loss 1.1195 Accuracy 0.2625
Epoch 8 Batch 200 Loss 1.1232 Accuracy 0.2626
Epoch 8 Batch 250 Loss 1.1329 Accuracy 0.2629
Epoch 8 Batch 300 Loss 1.1335 Accuracy 0.2632
Epoch 8 Batch 350 Loss 1.1331 Accuracy 0.2642
Epoch 8 Batch 400 Loss 1.1332 Accuracy 0.2643
Epoch 8 Batch 450 Loss 1.1341 Accuracy 0.2650
Epoch 8 Batch 500 Loss 1.1312 Accuracy 0.2649
Epoch 8 Batch 550 Loss 1.1288 Accuracy 0.2649
Epoch 8 Batch 600 Loss 1.1315 Accuracy 0.2651
Epoch 8 Batch 650 Loss 1.1318 Accuracy 0.2651
Epoch 8 Batch 700 Loss 1.1311 Accuracy 0.2652
Epoch 8 Loss 1.1311 Accuracy 0.2652
Time taken for 1 epoch: 29.10897397994995 secs
Epoch 9 Batch 0 Loss 0.9926 Accuracy 0.2632
Epoch 9 Batch 50 Loss 0.9922 Accuracy 0.2762
Epoch 9 Batch 100 Loss 0.9996 Accuracy 0.2768
Epoch 9 Batch 150 Loss 1.0093 Accuracy 0.2789
Epoch 9 Batch 200 Loss 1.0135 Accuracy 0.2786
Epoch 9 Batch 250 Loss 1.0115 Accuracy 0.2778
Epoch 9 Batch 300 Loss 1.0129 Accuracy 0.2783
Epoch 9 Batch 350 Loss 1.0118 Accuracy 0.2779
Epoch 9 Batch 400 Loss 1.0168 Accuracy 0.2784
Epoch 9 Batch 450 Loss 1.0183 Accuracy 0.2787
Epoch 9 Batch 500 Loss 1.0196 Accuracy 0.2785
Epoch 9 Batch 550 Loss 1.0219 Accuracy 0.2786
Epoch 9 Batch 600 Loss 1.0194 Accuracy 0.2787
Epoch 9 Batch 650 Loss 1.0204 Accuracy 0.2790
Epoch 9 Batch 700 Loss 1.0198 Accuracy 0.2790
Epoch 9 Loss 1.0198 Accuracy 0.2790
Time taken for 1 epoch: 29.176567554473877 secs
Epoch 10 Batch 0 Loss 0.9880 Accuracy 0.2877
Epoch 10 Batch 50 Loss 0.9109 Accuracy 0.2908
Epoch 10 Batch 100 Loss 0.9178 Accuracy 0.2928
Epoch 10 Batch 150 Loss 0.9211 Accuracy 0.2922
Epoch 10 Batch 200 Loss 0.9259 Accuracy 0.2919
Epoch 10 Batch 250 Loss 0.9215 Accuracy 0.2908
Epoch 10 Batch 300 Loss 0.9237 Accuracy 0.2908
Epoch 10 Batch 350 Loss 0.9270 Accuracy 0.2910
Epoch 10 Batch 400 Loss 0.9290 Accuracy 0.2910
Epoch 10 Batch 450 Loss 0.9287 Accuracy 0.2912
Epoch 10 Batch 500 Loss 0.9297 Accuracy 0.2909
Epoch 10 Batch 550 Loss 0.9301 Accuracy 0.2909
Epoch 10 Batch 600 Loss 0.9338 Accuracy 0.2906
Epoch 10 Batch 650 Loss 0.9334 Accuracy 0.2900
Epoch 10 Batch 700 Loss 0.9339 Accuracy 0.2899
Saving checkpoint for epoch 10 at ./checkpoints/train/ckpt-2
Epoch 10 Loss 0.9337 Accuracy 0.2899
Time taken for 1 epoch: 29.33884572982788 secs
Epoch 11 Batch 0 Loss 0.8248 Accuracy 0.2685
Epoch 11 Batch 50 Loss 0.8496 Accuracy 0.3032
Epoch 11 Batch 100 Loss 0.8424 Accuracy 0.3027
Epoch 11 Batch 150 Loss 0.8385 Accuracy 0.3010
Epoch 11 Batch 200 Loss 0.8439 Accuracy 0.3008
Epoch 11 Batch 250 Loss 0.8488 Accuracy 0.3002
Epoch 11 Batch 300 Loss 0.8520 Accuracy 0.3003
Epoch 11 Batch 350 Loss 0.8523 Accuracy 0.2995
Epoch 11 Batch 400 Loss 0.8575 Accuracy 0.3007
Epoch 11 Batch 450 Loss 0.8584 Accuracy 0.3005
Epoch 11 Batch 500 Loss 0.8608 Accuracy 0.3001
Epoch 11 Batch 550 Loss 0.8629 Accuracy 0.3004
Epoch 11 Batch 600 Loss 0.8643 Accuracy 0.2999
Epoch 11 Batch 650 Loss 0.8651 Accuracy 0.2994
Epoch 11 Batch 700 Loss 0.8675 Accuracy 0.2993
Epoch 11 Loss 0.8674 Accuracy 0.2992
Time taken for 1 epoch: 29.55175805091858 secs
Epoch 12 Batch 0 Loss 0.8606 Accuracy 0.3359
Epoch 12 Batch 50 Loss 0.7656 Accuracy 0.3072
Epoch 12 Batch 100 Loss 0.7779 Accuracy 0.3073
Epoch 12 Batch 150 Loss 0.7755 Accuracy 0.3061
Epoch 12 Batch 200 Loss 0.7779 Accuracy 0.3062
Epoch 12 Batch 250 Loss 0.7808 Accuracy 0.3056
Epoch 12 Batch 300 Loss 0.7867 Accuracy 0.3051
Epoch 12 Batch 350 Loss 0.7914 Accuracy 0.3058
Epoch 12 Batch 400 Loss 0.7944 Accuracy 0.3062
Epoch 12 Batch 450 Loss 0.7974 Accuracy 0.3061
Epoch 12 Batch 500 Loss 0.8003 Accuracy 0.3064
Epoch 12 Batch 550 Loss 0.8040 Accuracy 0.3062
Epoch 12 Batch 600 Loss 0.8061 Accuracy 0.3062
Epoch 12 Batch 650 Loss 0.8082 Accuracy 0.3061
Epoch 12 Batch 700 Loss 0.8114 Accuracy 0.3063
Epoch 12 Loss 0.8114 Accuracy 0.3063
Time taken for 1 epoch: 29.188767671585083 secs
Epoch 13 Batch 0 Loss 0.8691 Accuracy 0.3529
Epoch 13 Batch 50 Loss 0.7194 Accuracy 0.3176
Epoch 13 Batch 100 Loss 0.7199 Accuracy 0.3187
Epoch 13 Batch 150 Loss 0.7295 Accuracy 0.3148
Epoch 13 Batch 200 Loss 0.7368 Accuracy 0.3145
Epoch 13 Batch 250 Loss 0.7381 Accuracy 0.3136
Epoch 13 Batch 300 Loss 0.7414 Accuracy 0.3127
Epoch 13 Batch 350 Loss 0.7439 Accuracy 0.3137
Epoch 13 Batch 400 Loss 0.7463 Accuracy 0.3131
Epoch 13 Batch 450 Loss 0.7514 Accuracy 0.3136
Epoch 13 Batch 500 Loss 0.7536 Accuracy 0.3135
Epoch 13 Batch 550 Loss 0.7564 Accuracy 0.3134
Epoch 13 Batch 600 Loss 0.7600 Accuracy 0.3131
Epoch 13 Batch 650 Loss 0.7621 Accuracy 0.3129
Epoch 13 Batch 700 Loss 0.7648 Accuracy 0.3125
Epoch 13 Loss 0.7646 Accuracy 0.3125
Time taken for 1 epoch: 29.76813769340515 secs
Epoch 14 Batch 0 Loss 0.7053 Accuracy 0.2743
Epoch 14 Batch 50 Loss 0.6693 Accuracy 0.3176
Epoch 14 Batch 100 Loss 0.6755 Accuracy 0.3203
Epoch 14 Batch 150 Loss 0.6884 Accuracy 0.3211
Epoch 14 Batch 200 Loss 0.6917 Accuracy 0.3210
Epoch 14 Batch 250 Loss 0.6945 Accuracy 0.3190
Epoch 14 Batch 300 Loss 0.6993 Accuracy 0.3198
Epoch 14 Batch 350 Loss 0.7018 Accuracy 0.3197
Epoch 14 Batch 400 Loss 0.7073 Accuracy 0.3196
Epoch 14 Batch 450 Loss 0.7094 Accuracy 0.3193
Epoch 14 Batch 500 Loss 0.7114 Accuracy 0.3191
Epoch 14 Batch 550 Loss 0.7161 Accuracy 0.3193
Epoch 14 Batch 600 Loss 0.7184 Accuracy 0.3189
Epoch 14 Batch 650 Loss 0.7199 Accuracy 0.3187
Epoch 14 Batch 700 Loss 0.7243 Accuracy 0.3182
Epoch 14 Loss 0.7242 Accuracy 0.3182
Time taken for 1 epoch: 29.221008777618408 secs
Epoch 15 Batch 0 Loss 0.7041 Accuracy 0.3364
Epoch 15 Batch 50 Loss 0.6372 Accuracy 0.3242
Epoch 15 Batch 100 Loss 0.6471 Accuracy 0.3253
Epoch 15 Batch 150 Loss 0.6544 Accuracy 0.3246
Epoch 15 Batch 200 Loss 0.6573 Accuracy 0.3239
Epoch 15 Batch 250 Loss 0.6611 Accuracy 0.3239
Epoch 15 Batch 300 Loss 0.6647 Accuracy 0.3241
Epoch 15 Batch 350 Loss 0.6661 Accuracy 0.3246
Epoch 15 Batch 400 Loss 0.6717 Accuracy 0.3247
Epoch 15 Batch 450 Loss 0.6729 Accuracy 0.3239
Epoch 15 Batch 500 Loss 0.6758 Accuracy 0.3236
Epoch 15 Batch 550 Loss 0.6781 Accuracy 0.3236
Epoch 15 Batch 600 Loss 0.6819 Accuracy 0.3237
Epoch 15 Batch 650 Loss 0.6841 Accuracy 0.3232
Epoch 15 Batch 700 Loss 0.6869 Accuracy 0.3229
Saving checkpoint for epoch 15 at ./checkpoints/train/ckpt-3
Epoch 15 Loss 0.6870 Accuracy 0.3230
Time taken for 1 epoch: 29.415014266967773 secs
Epoch 16 Batch 0 Loss 0.6384 Accuracy 0.3366
Epoch 16 Batch 50 Loss 0.6183 Accuracy 0.3313
Epoch 16 Batch 100 Loss 0.6249 Accuracy 0.3319
Epoch 16 Batch 150 Loss 0.6259 Accuracy 0.3317
Epoch 16 Batch 200 Loss 0.6284 Accuracy 0.3306
Epoch 16 Batch 250 Loss 0.6325 Accuracy 0.3305
Epoch 16 Batch 300 Loss 0.6344 Accuracy 0.3301
Epoch 16 Batch 350 Loss 0.6384 Accuracy 0.3300
Epoch 16 Batch 400 Loss 0.6397 Accuracy 0.3298
Epoch 16 Batch 450 Loss 0.6408 Accuracy 0.3288
Epoch 16 Batch 500 Loss 0.6429 Accuracy 0.3287
Epoch 16 Batch 550 Loss 0.6450 Accuracy 0.3283
Epoch 16 Batch 600 Loss 0.6470 Accuracy 0.3282
Epoch 16 Batch 650 Loss 0.6509 Accuracy 0.3279
Epoch 16 Batch 700 Loss 0.6540 Accuracy 0.3277
Epoch 16 Loss 0.6544 Accuracy 0.3278
Time taken for 1 epoch: 29.107020616531372 secs
Epoch 17 Batch 0 Loss 0.6762 Accuracy 0.3594
Epoch 17 Batch 50 Loss 0.5904 Accuracy 0.3353
Epoch 17 Batch 100 Loss 0.5854 Accuracy 0.3353
Epoch 17 Batch 150 Loss 0.5913 Accuracy 0.3339
Epoch 17 Batch 200 Loss 0.5985 Accuracy 0.3349
Epoch 17 Batch 250 Loss 0.6008 Accuracy 0.3347
Epoch 17 Batch 300 Loss 0.6043 Accuracy 0.3343
Epoch 17 Batch 350 Loss 0.6071 Accuracy 0.3345
Epoch 17 Batch 400 Loss 0.6113 Accuracy 0.3341
Epoch 17 Batch 450 Loss 0.6151 Accuracy 0.3339
Epoch 17 Batch 500 Loss 0.6181 Accuracy 0.3337
Epoch 17 Batch 550 Loss 0.6198 Accuracy 0.3331
Epoch 17 Batch 600 Loss 0.6233 Accuracy 0.3330
Epoch 17 Batch 650 Loss 0.6253 Accuracy 0.3327
Epoch 17 Batch 700 Loss 0.6275 Accuracy 0.3322
Epoch 17 Loss 0.6275 Accuracy 0.3321
Time taken for 1 epoch: 29.163068532943726 secs
Epoch 18 Batch 0 Loss 0.5705 Accuracy 0.3641
Epoch 18 Batch 50 Loss 0.5579 Accuracy 0.3421
Epoch 18 Batch 100 Loss 0.5641 Accuracy 0.3413
Epoch 18 Batch 150 Loss 0.5724 Accuracy 0.3426
Epoch 18 Batch 200 Loss 0.5734 Accuracy 0.3415
Epoch 18 Batch 250 Loss 0.5791 Accuracy 0.3407
Epoch 18 Batch 300 Loss 0.5836 Accuracy 0.3400
Epoch 18 Batch 350 Loss 0.5851 Accuracy 0.3395
Epoch 18 Batch 400 Loss 0.5862 Accuracy 0.3397
Epoch 18 Batch 450 Loss 0.5890 Accuracy 0.3390
Epoch 18 Batch 500 Loss 0.5913 Accuracy 0.3381
Epoch 18 Batch 550 Loss 0.5933 Accuracy 0.3374
Epoch 18 Batch 600 Loss 0.5959 Accuracy 0.3374
Epoch 18 Batch 650 Loss 0.5985 Accuracy 0.3372
Epoch 18 Batch 700 Loss 0.6018 Accuracy 0.3367
Epoch 18 Loss 0.6023 Accuracy 0.3367
Time taken for 1 epoch: 29.096425533294678 secs
Epoch 19 Batch 0 Loss 0.5866 Accuracy 0.3775
Epoch 19 Batch 50 Loss 0.5498 Accuracy 0.3462
Epoch 19 Batch 100 Loss 0.5529 Accuracy 0.3466
Epoch 19 Batch 150 Loss 0.5527 Accuracy 0.3450
Epoch 19 Batch 200 Loss 0.5566 Accuracy 0.3460
Epoch 19 Batch 250 Loss 0.5593 Accuracy 0.3448
Epoch 19 Batch 300 Loss 0.5625 Accuracy 0.3443
Epoch 19 Batch 350 Loss 0.5638 Accuracy 0.3437
Epoch 19 Batch 400 Loss 0.5668 Accuracy 0.3436
Epoch 19 Batch 450 Loss 0.5687 Accuracy 0.3434
Epoch 19 Batch 500 Loss 0.5717 Accuracy 0.3428
Epoch 19 Batch 550 Loss 0.5740 Accuracy 0.3420
Epoch 19 Batch 600 Loss 0.5764 Accuracy 0.3418
Epoch 19 Batch 650 Loss 0.5782 Accuracy 0.3414
Epoch 19 Batch 700 Loss 0.5802 Accuracy 0.3407
Epoch 19 Loss 0.5803 Accuracy 0.3407
Time taken for 1 epoch: 29.182754516601562 secs
Epoch 20 Batch 0 Loss 0.4925 Accuracy 0.3074
Epoch 20 Batch 50 Loss 0.5315 Accuracy 0.3482
Epoch 20 Batch 100 Loss 0.5309 Accuracy 0.3517
Epoch 20 Batch 150 Loss 0.5291 Accuracy 0.3496
Epoch 20 Batch 200 Loss 0.5330 Accuracy 0.3498
Epoch 20 Batch 250 Loss 0.5347 Accuracy 0.3485
Epoch 20 Batch 300 Loss 0.5385 Accuracy 0.3473
Epoch 20 Batch 350 Loss 0.5398 Accuracy 0.3465
Epoch 20 Batch 400 Loss 0.5438 Accuracy 0.3471
Epoch 20 Batch 450 Loss 0.5470 Accuracy 0.3461
Epoch 20 Batch 500 Loss 0.5489 Accuracy 0.3458
Epoch 20 Batch 550 Loss 0.5506 Accuracy 0.3451
Epoch 20 Batch 600 Loss 0.5524 Accuracy 0.3444
Epoch 20 Batch 650 Loss 0.5561 Accuracy 0.3444
Epoch 20 Batch 700 Loss 0.5586 Accuracy 0.3439
Saving checkpoint for epoch 20 at ./checkpoints/train/ckpt-4
Epoch 20 Loss 0.5588 Accuracy 0.3439
Time taken for 1 epoch: 29.352638959884644 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(
['' ]+[tokenizer_pt.decode([i]) for i in sentence]+['' ],
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 address our overall issues .
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... 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 a lot of share with you some pretty magical things that have happened .
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 .
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。