寫在前面
本文首發(fā)于公眾號：NewBeeNLP
update@2020.02.10
最近在看paddle相關(guān)，于是就打算仔細過一遍百度ERNIE的源碼。之前粗看的時候還沒有ERNIE2.0、ERNIE-tiny,整體感覺跟BERT也挺類似的，不知道更新了之后會是啥樣~看完也會整理跟下面類似的總結(jié)，剛好也在研究paddle或ERNIE的同學(xué)可以加我一起討論哈哈哈
原內(nèi)容@2019.05.16
BERT 模型也出來很久了， 之前有看過論文和一些博客對其做了解讀：NLP 大殺器 BERT 模型解讀[1]，但是一直沒有細致地去看源碼具體實現(xiàn)。最近有用到就抽時間來仔細看看記錄下來，和大家一起討論。
注意，源碼閱讀系列需要提前對 NLP 相關(guān)知識有所了解，比如 attention 機制、transformer 框架以及 python 和 tensorflow 基礎(chǔ)等，關(guān)于 BERT 的原理不是本文的重點。
附上關(guān)于 BERT 的資料匯總：BERT 相關(guān)論文、文章和代碼資源匯總[2]
今天要介紹的是 BERT 最主要的模型實現(xiàn)部分-----BertModel，代碼位于
如有解讀不正確，請務(wù)必指出~
1、配置類（BertConfig）
這部分代碼主要定義了 BERT 模型的一些默認參數(shù)，另外包括了一些文件處理函數(shù)。
class BertConfig(object):
"""BERT模型的配置類."""
def __init__(self,
vocab_size,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
initializer_range=0.02):
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.hidden_act = hidden_act
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
@classmethod
def from_dict(cls, json_object):
"""Constructs a `BertConfig` from a Python dictionary of parameters."""
config = BertConfig(vocab_size=None)
for (key, value) in six.iteritems(json_object):
config.__dict__[key] = value
return config
@classmethod
def from_json_file(cls, json_file):
"""Constructs a `BertConfig` from a json file of parameters."""
with tf.gfile.GFile(json_file, "r") as reader:
text = reader.read()
return cls.from_dict(json.loads(text))
def to_dict(self):
"""Serializes this instance to a Python dictionary."""
output = copy.deepcopy(self.__dict__)
return output
def to_json_string(self):
"""Serializes this instance to a JSON string."""
return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
「參數(shù)具體含義」
這里要注意一點，可能剛看的時候?qū)?code>type_vocab_size這個參數(shù)會有點不理解，其實就是在next sentence prediction任務(wù)里的Segment A和 Segment B。在下載的bert_config.json文件里也有說明，默認值應(yīng)該為 2。參考這個 Issue[4]
2、獲取詞向量（Embedding_lookup）
對于輸入 word_ids，返回 embedding table?？梢赃x用 one-hot 或者 tf.gather()
def embedding_lookup(input_ids, # word_id：【batch_size, seq_length】
vocab_size,
embedding_size=128,
initializer_range=0.02,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=False):
# 該函數(shù)默認輸入的形狀為【batch_size, seq_length, input_num】
# 如果輸入為2D的【batch_size, seq_length】，則擴展到【batch_size, seq_length, 1】
if input_ids.shape.ndims == 2:
input_ids = tf.expand_dims(input_ids, axis=[-1])
embedding_table = tf.get_variable(
name=word_embedding_name,
shape=[vocab_size, embedding_size],
initializer=create_initializer(initializer_range))
flat_input_ids = tf.reshape(input_ids, [-1]) #【batch_size*seq_length*input_num】
if use_one_hot_embeddings:
one_hot_input_ids = tf.one_hot(flat_input_ids, depth=vocab_size)
output = tf.matmul(one_hot_input_ids, embedding_table)
else: # 按索引取值
output = tf.gather(embedding_table, flat_input_ids)
input_shape = get_shape_list(input_ids)
# output：[batch_size, seq_length, num_inputs]
# 轉(zhuǎn)成:[batch_size, seq_length, num_inputs*embedding_size]
output = tf.reshape(output,
input_shape[0:-1] + [input_shape[-1] * embedding_size])
return (output, embedding_table)
「參數(shù)具體含義」
3、詞向量的后續(xù)處理（embedding_postprocessor）
def embedding_postprocessor(input_tensor, # [batch_size, seq_length, embedding_size]
use_token_type=False,
token_type_ids=None,
token_type_vocab_size=16, # 一般是2
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=0.02,
max_position_embeddings=512, #最大位置編碼，必須大于等于max_seq_len
dropout_prob=0.1):
input_shape = get_shape_list(input_tensor, expected_rank=3) #【batch_size,seq_length,embedding_size】
batch_size = input_shape[0]
seq_length = input_shape[1]
width = input_shape[2]
output = input_tensor
# Segment position信息
if use_token_type:
if token_type_ids is None:
raise ValueError("`token_type_ids` must be specified if"
"`use_token_type` is True.")
token_type_table = tf.get_variable(
name=token_type_embedding_name,
shape=[token_type_vocab_size, width],
initializer=create_initializer(initializer_range))
# 由于token-type-table比較小，所以這里采用one-hot的embedding方式加速
flat_token_type_ids = tf.reshape(token_type_ids, [-1])
one_hot_ids = tf.one_hot(flat_token_type_ids, depth=token_type_vocab_size)
token_type_embeddings = tf.matmul(one_hot_ids, token_type_table)
token_type_embeddings = tf.reshape(token_type_embeddings,
[batch_size, seq_length, width])
output += token_type_embeddings
# Position embedding信息
if use_position_embeddings:
# 確保seq_length小于等于max_position_embeddings
assert_op = tf.assert_less_equal(seq_length, max_position_embeddings)
with tf.control_dependencies([assert_op]):
full_position_embeddings = tf.get_variable(
name=position_embedding_name,
shape=[max_position_embeddings, width],
initializer=create_initializer(initializer_range))
# 這里position embedding是可學(xué)習的參數(shù)，[max_position_embeddings, width]
# 但是通常實際輸入序列沒有達到max_position_embeddings
# 所以為了提高訓(xùn)練速度，使用tf.slice取出句子長度的embedding
position_embeddings = tf.slice(full_position_embeddings, [0, 0],
[seq_length, -1])
num_dims = len(output.shape.as_list())
# word embedding之后的tensor是[batch_size, seq_length, width]
# 因為位置編碼是與輸入內(nèi)容無關(guān)，它的shape總是[seq_length, width]
# 我們無法把位置Embedding加到word embedding上
# 因此我們需要擴展位置編碼為[1, seq_length, width]
# 然后就能通過broadcasting加上去了。
position_broadcast_shape = []
for _ in range(num_dims - 2):
position_broadcast_shape.append(1)
position_broadcast_shape.extend([seq_length, width])
position_embeddings = tf.reshape(position_embeddings,
position_broadcast_shape)
output += position_embeddings
output = layer_norm_and_dropout(output, dropout_prob)
return output
4、構(gòu)造 attention_mask
該部分代碼的作用是構(gòu)造 attention 可視域的 attention_mask， 因為每個樣本都經(jīng)過 padding 過程，在做self-attention的是padding的部分不能attend到其他部分上。輸入為形狀為 [batch_size, from_seq_length,...] 的 padding 好的 input_ids 和形狀為 [batch_size, to_seq_length] 的 mask 標記向量。
def create_attention_mask_from_input_mask(from_tensor, to_mask):
from_shape = get_shape_list(from_tensor, expected_rank=[2, 3])
batch_size = from_shape[0]
from_seq_length = from_shape[1]
to_shape = get_shape_list(to_mask, expected_rank=2)
to_seq_length = to_shape[1]
to_mask = tf.cast(
tf.reshape(to_mask, [batch_size, 1, to_seq_length]), tf.float32)
broadcast_ones = tf.ones(
shape=[batch_size, from_seq_length, 1], dtype=tf.float32)
mask = broadcast_ones * to_mask
return mask
5、注意力層（attention layer）
這部分代碼是「multi-head attention」的實現(xiàn)，主要來自《Attention is all you need》這篇論文?？紤]key-query-value形式的 attention，輸入的from_tensor當做是 query， to_tensor當做是 key 和 value，當兩者相同的時候即為 self-attention。關(guān)于 attention 更詳細的介紹可以轉(zhuǎn)到【理解 Attention 機制原理及模型[5]】。
def attention_layer(from_tensor, # 【batch_size, from_seq_length, from_width】
to_tensor, #【batch_size, to_seq_length, to_width】
attention_mask=None, #【batch_size,from_seq_length, to_seq_length】
num_attention_heads=1, # attention head numbers
size_per_head=512, # 每個head的大小
query_act=None, # query變換的激活函數(shù)
key_act=None, # key變換的激活函數(shù)
value_act=None, # value變換的激活函數(shù)
attention_probs_dropout_prob=0.0, # attention層的dropout
initializer_range=0.02, # 初始化取值范圍
do_return_2d_tensor=False, # 是否返回2d張量。
#如果True，輸出形狀【batch_size*from_seq_length,num_attention_heads*size_per_head】
#如果False，輸出形狀【batch_size, from_seq_length, num_attention_heads*size_per_head】
batch_size=None, #如果輸入是3D的，
#那么batch就是第一維，但是可能3D的壓縮成了2D的，所以需要告訴函數(shù)batch_size
from_seq_length=None, # 同上
to_seq_length=None): # 同上
def transpose_for_scores(input_tensor, batch_size, num_attention_heads,
seq_length, width):
output_tensor = tf.reshape(
input_tensor, [batch_size, seq_length, num_attention_heads, width])
output_tensor = tf.transpose(output_tensor, [0, 2, 1, 3]) #[batch_size, num_attention_heads, seq_length, width]
return output_tensor
from_shape = get_shape_list(from_tensor, expected_rank=[2, 3])
to_shape = get_shape_list(to_tensor, expected_rank=[2, 3])
if len(from_shape) != len(to_shape):
raise ValueError(
"The rank of `from_tensor` must match the rank of `to_tensor`.")
if len(from_shape) == 3:
batch_size = from_shape[0]
from_seq_length = from_shape[1]
to_seq_length = to_shape[1]
elif len(from_shape) == 2:
if (batch_size is None or from_seq_length is None or to_seq_length is None):
raise ValueError(
"When passing in rank 2 tensors to attention_layer, the values "
"for `batch_size`, `from_seq_length`, and `to_seq_length` "
"must all be specified.")
# 為了方便備注shape，采用以下簡寫:
# B = batch size (number of sequences)
# F = `from_tensor` sequence length
# T = `to_tensor` sequence length
# N = `num_attention_heads`
# H = `size_per_head`
# 把from_tensor和to_tensor壓縮成2D張量
from_tensor_2d = reshape_to_matrix(from_tensor) # 【B*F, hidden_size】
to_tensor_2d = reshape_to_matrix(to_tensor) # 【B*T, hidden_size】
# 將from_tensor輸入全連接層得到query_layer
# `query_layer` = [B*F, N*H]
query_layer = tf.layers.dense(
from_tensor_2d,
num_attention_heads * size_per_head,
activation=query_act,
name="query",
kernel_initializer=create_initializer(initializer_range))
# 將from_tensor輸入全連接層得到query_layer
# `key_layer` = [B*T, N*H]
key_layer = tf.layers.dense(
to_tensor_2d,
num_attention_heads * size_per_head,
activation=key_act,
name="key",
kernel_initializer=create_initializer(initializer_range))
# 同上
# `value_layer` = [B*T, N*H]
value_layer = tf.layers.dense(
to_tensor_2d,
num_attention_heads * size_per_head,
activation=value_act,
name="value",
kernel_initializer=create_initializer(initializer_range))
# query_layer轉(zhuǎn)成多頭：[B*F, N*H]==>[B, F, N, H]==>[B, N, F, H]
query_layer = transpose_for_scores(query_layer, batch_size,
num_attention_heads, from_seq_length,
size_per_head)
# key_layer轉(zhuǎn)成多頭：[B*T, N*H] ==> [B, T, N, H] ==> [B, N, T, H]
key_layer = transpose_for_scores(key_layer, batch_size, num_attention_heads,
to_seq_length, size_per_head)
# 將query與key做點積，然后做一個scale，公式可以參見原始論文
# `attention_scores` = [B, N, F, T]
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
attention_scores = tf.multiply(attention_scores,
1.0 / math.sqrt(float(size_per_head)))
if attention_mask is not None:
# `attention_mask` = [B, 1, F, T]
attention_mask = tf.expand_dims(attention_mask, axis=[1])
# 如果attention_mask里的元素為1，則通過下面運算有（1-1）*-10000，adder就是0
# 如果attention_mask里的元素為0，則通過下面運算有（1-0）*-10000，adder就是-10000
adder = (1.0 - tf.cast(attention_mask, tf.float32)) * -10000.0
# 我們最終得到的attention_score一般不會很大，
#所以上述操作對mask為0的地方得到的score可以認為是負無窮
attention_scores += adder
# 負無窮經(jīng)過softmax之后為0，就相當于mask為0的位置不計算attention_score
# `attention_probs` = [B, N, F, T]
attention_probs = tf.nn.softmax(attention_scores)
# 對attention_probs進行dropout，這雖然有點奇怪，但是Transforme原始論文就是這么做的
attention_probs = dropout(attention_probs, attention_probs_dropout_prob)
# `value_layer` = [B, T, N, H]
value_layer = tf.reshape(
value_layer,
[batch_size, to_seq_length, num_attention_heads, size_per_head])
# `value_layer` = [B, N, T, H]
value_layer = tf.transpose(value_layer, [0, 2, 1, 3])
# `context_layer` = [B, N, F, H]
context_layer = tf.matmul(attention_probs, value_layer)
# `context_layer` = [B, F, N, H]
context_layer = tf.transpose(context_layer, [0, 2, 1, 3])
if do_return_2d_tensor:
# `context_layer` = [B*F, N*H]
context_layer = tf.reshape(
context_layer,
[batch_size * from_seq_length, num_attention_heads * size_per_head])
else:
# `context_layer` = [B, F, N*H]
context_layer = tf.reshape(
context_layer,
[batch_size, from_seq_length, num_attention_heads * size_per_head])
return context_layer
總結(jié)一下，attention layer 的主要流程：
batch_size、from_seq_length 、to_seq_length；transpose_for_scores轉(zhuǎn)化成 multi-head；6、Transformer
接下來的代碼就是大名鼎鼎的 Transformer 的核心代碼了，可以認為是"Attention is All You Need"原始代碼重現(xiàn)。可以參見【原始論文[6]】和【原始代碼[7]】。
def transformer_model(input_tensor, # 【batch_size, seq_length, hidden_size】
attention_mask=None, # 【batch_size, seq_length, seq_length】
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
intermediate_act_fn=gelu, # feed-forward層的激活函數(shù)
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
initializer_range=0.02,
do_return_all_layers=False):
# 這里注意，因為最終要輸出hidden_size， 我們有num_attention_head個區(qū)域，
# 每個head區(qū)域有size_per_head多的隱層
# 所以有 hidden_size = num_attention_head * size_per_head
if hidden_size % num_attention_heads != 0:
raise ValueError(
"The hidden size (%d) is not a multiple of the number of attention "
"heads (%d)" % (hidden_size, num_attention_heads))
attention_head_size = int(hidden_size / num_attention_heads)
input_shape = get_shape_list(input_tensor, expected_rank=3)
batch_size = input_shape[0]
seq_length = input_shape[1]
input_width = input_shape[2]
# 因為encoder中有殘差操作，所以需要shape相同
if input_width != hidden_size:
raise ValueError("The width of the input tensor (%d) != hidden size (%d)" %
(input_width, hidden_size))
# reshape操作在CPU/GPU上很快，但是在TPU上很不友好
# 所以為了避免2D和3D之間的頻繁reshape，我們把所有的3D張量用2D矩陣表示
prev_output = reshape_to_matrix(input_tensor)
all_layer_outputs = []
for layer_idx in range(num_hidden_layers):
with tf.variable_scope("layer_%d" % layer_idx):
layer_input = prev_output
with tf.variable_scope("attention"):
# multi-head attention
attention_heads = []
with tf.variable_scope("self"):
# self-attention
attention_head = attention_layer(
from_tensor=layer_input,
to_tensor=layer_input,
attention_mask=attention_mask,
num_attention_heads=num_attention_heads,
size_per_head=attention_head_size,
attention_probs_dropout_prob=attention_probs_dropout_prob,
initializer_range=initializer_range,
do_return_2d_tensor=True,
batch_size=batch_size,
from_seq_length=seq_length,
to_seq_length=seq_length)
attention_heads.append(attention_head)
attention_output = None
if len(attention_heads) == 1:
attention_output = attention_heads[0]
else:
# 如果有多個head，將他們拼接起來
attention_output = tf.concat(attention_heads, axis=-1)
# 對attention的輸出進行線性映射, 目的是將shape變成與input一致
# 然后dropout+residual+norm
with tf.variable_scope("output"):
attention_output = tf.layers.dense(
attention_output,
hidden_size,
kernel_initializer=create_initializer(initializer_range))
attention_output = dropout(attention_output, hidden_dropout_prob)
attention_output = layer_norm(attention_output + layer_input)
# feed-forward
with tf.variable_scope("intermediate"):
intermediate_output = tf.layers.dense(
attention_output,
intermediate_size,
activation=intermediate_act_fn,
kernel_initializer=create_initializer(initializer_range))
# 對feed-forward層的輸出使用線性變換變回‘hidden_size’
# 然后dropout + residual + norm
with tf.variable_scope("output"):
layer_output = tf.layers.dense(
intermediate_output,
hidden_size,
kernel_initializer=create_initializer(initializer_range))
layer_output = dropout(layer_output, hidden_dropout_prob)
layer_output = layer_norm(layer_output + attention_output)
prev_output = layer_output
all_layer_outputs.append(layer_output)
if do_return_all_layers:
final_outputs = []
for layer_output in all_layer_outputs:
final_output = reshape_from_matrix(layer_output, input_shape)
final_outputs.append(final_output)
return final_outputs
else:
final_output = reshape_from_matrix(prev_output, input_shape)
return final_output
7、函數(shù)入口（init）
BertModel 類的構(gòu)造函數(shù)，有了上面幾節(jié)的鋪墊，我們就可以來實現(xiàn) BERT 模型了。
def __init__(self,
config, # BertConfig對象
is_training,
input_ids, # 【batch_size, seq_length】
input_mask=None, # 【batch_size, seq_length】
token_type_ids=None, # 【batch_size, seq_length】
use_one_hot_embeddings=False, # 是否使用one-hot；否則tf.gather()
scope=None):
config = copy.deepcopy(config)
if not is_training:
config.hidden_dropout_prob = 0.0
config.attention_probs_dropout_prob = 0.0
input_shape = get_shape_list(input_ids, expected_rank=2)
batch_size = input_shape[0]
seq_length = input_shape[1]
# 不做mask，即所有元素為1
if input_mask is None:
input_mask = tf.ones(shape=[batch_size, seq_length], dtype=tf.int32)
if token_type_ids is None:
token_type_ids = tf.zeros(shape=[batch_size, seq_length], dtype=tf.int32)
with tf.variable_scope(scope, default_name="bert"):
with tf.variable_scope("embeddings"):
# word embedding
(self.embedding_output, self.embedding_table) = embedding_lookup(
input_ids=input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
# 添加position embedding和segment embedding
# layer norm + dropout
self.embedding_output = embedding_postprocessor(
input_tensor=self.embedding_output,
use_token_type=True,
token_type_ids=token_type_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob)
with tf.variable_scope("encoder"):
# input_ids是經(jīng)過padding的word_ids：[25, 120, 34, 0, 0]
# input_mask是有效詞標記：[1, 1, 1, 0, 0]
attention_mask = create_attention_mask_from_input_mask(
input_ids, input_mask)
# transformer模塊疊加
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
self.all_encoder_layers = transformer_model(
input_tensor=self.embedding_output,
attention_mask=attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
# `self.sequence_output`是最后一層的輸出，shape為【batch_size, seq_length, hidden_size】
self.sequence_output = self.all_encoder_layers[-1]
# ‘pooler’部分將encoder輸出【batch_size, seq_length, hidden_size】
# 轉(zhuǎn)成【batch_size, hidden_size】
with tf.variable_scope("pooler"):
# 取最后一層的第一個時刻[CLS]對應(yīng)的tensor， 對于分類任務(wù)很重要
# sequence_output[:, 0:1, :]得到的是[batch_size, 1, hidden_size]
# 我們需要用squeeze把第二維去掉
first_token_tensor = tf.squeeze(self.sequence_output[:, 0:1, :], axis=1)
# 然后再加一個全連接層，輸出仍然是[batch_size, hidden_size]
self.pooled_output = tf.layers.dense(
first_token_tensor,
config.hidden_size,
activation=tf.tanh,
kernel_initializer=create_initializer(config.initializer_range))
總結(jié)一哈
有了以上對源碼的深入了解之后，我們在使用 BertModel 的時候就會更加得心應(yīng)手。舉個模型使用的簡單栗子：
# 假設(shè)輸入已經(jīng)經(jīng)過分詞變成word_ids. shape=[2, 3]
input_ids = tf.constant([[31, 51, 99], [15, 5, 0]])
input_mask = tf.constant([[1, 1, 1], [1, 1, 0]])
# segment_emebdding. 表示第一個樣本前兩個詞屬于句子1，后一個詞屬于句子2.
# 第二個樣本的第一個詞屬于句子1， 第二次詞屬于句子2，第三個元素0表示padding
# 原始代碼是下面這樣的，但是感覺么必要用 2，不知道是不是我哪里沒理解
token_type_ids = tf.constant([[0, 0, 1], [0, 2, 0]])
# 創(chuàng)建BertConfig實例
config = modeling.BertConfig(vocab_size=32000, hidden_size=512,
num_hidden_layers=8, num_attention_heads=6, intermediate_size=1024)
# 創(chuàng)建BertModel實例
model = modeling.BertModel(config=config, is_training=True,
input_ids=input_ids, input_mask=input_mask, token_type_ids=token_type_ids)
label_embeddings = tf.get_variable(...)
#得到最后一層的第一個Token也就是[CLS]向量表示，可以看成是一個句子的embedding
pooled_output = model.get_pooled_output()
logits = tf.matmul(pooled_output, label_embeddings)
在 BERT 模型構(gòu)建這一塊的主要流程：
以上~?
本文參考資料
[1]NLP 大殺器 BERT 模型解讀: https://blog.csdn.net/Kaiyuan_sjtu/article/details/83991186
[2]BERT 相關(guān)論文、文章和代碼資源匯總: http://www.52nlp.cn/bert-paper-%E8%AE%BA%E6%96%87-%E6%96%87%E7%AB%A0-%E4%BB%A3%E7%A0%81%E8%B5%84%E6%BA%90%E6%B1%87%E6%80%BB
[3]modeling.py 模塊: https://github.com/google-research/bert/blob/master/modeling.py
[4]參考這個 Issue: https://github.com/google-research/bert/issues/16
[5]理解 Attention 機制原理及模型: https://blog.csdn.net/Kaiyuan_sjtu/article/details/81806123
[6]原始論文: https://arxiv.org/abs/1706.03762
[7]原始代碼: https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/models/transformer.py
