[transformer]论文实现:Attention Is All You Need(上)https://developer.aliyun.com/article/1504069?spm=a2c6h.13148508.setting.44.36834f0eMJOehx
2.4 注意力机制
从图中可以发现,Multi-Head Attention
和Add & Norm
同样一起遍布了整个模型;
注意力机制是怎么工作的:
图中有四个东西,Query
,Key
,Value
,Attention
,图中把Key
和Value
放在了一起是因为其产生的output
是与Key
,Value
的维度是无关的,只与Query
的维度有关;文中把Key
,Value
称作为Context sequence
,Query
还是称作为Query sequence
;为什么要这么做?可以看模型中右上方的Multi-Head Attention
和左下角的Multi-Head Attention
的区别进行分析;
Query
,Key
,Value
这三个东西可以用python
中的字典来解释,Key
,Value
表示字典中的键值对,而Query
表示我们需要查询的键,Query
与Key
,Value
匹配其得到的结果就是我们需要的信息;但是在这里并不要求Query
与Key
严格匹配,只需要模糊匹配就可以;Query
对每一个Key
进行一次模糊的匹配,并给出匹配程度,越适配权重就越大,然后根据权重再与每一个Value
进行组合,得到最后的结果;其匹配程度的权重就代表了注意力机制的权重;
多头注意力机制就是把Query
,Key
,Value
多个维度的向量分为几个少数维度的向量组合,再在Query_i
,Key_i
,Value_i
上进行操作,最后把结果合并;
模型中Multi-Head Attention
有三个, 这三个分别对应三种Multi-Head Attention Layer
:the cross attention layer
,the global self attention layer
, the causal self attention layer
,从图中也可以发现每一层都有各自的不同,下面来一一介绍;
the cross attention layer
:模型右上角(解码器)的Multi-Head Attention
是注意力机制最直接的使用,其将编码器的信息和解码器的信息充分的结合了起来,文中把这一层叫做the cross attention layer
;其context sequence
是Encoder
中得到的;
在这里要注意的是,每一次的查询是可以看得到所有的Key
,Value
的,但是查询与查询相互之间是看不到的,即独立的;
the global self attention layer
:模型左下角(编码器)的Multi-Head Attention
,这一层负责处理上下文序列,并沿着他的长度去传播信息即Query
与Query
之间的信息;
Query与Query之间的信息传播有很多种方式,例如在Transformer没出来之间我们普遍采用Bidirectional RNNs
和 CNNs
的方式来处理;
但是为什么这里不使用RNN
和CNN
的方法呢?
RNN
和CNN
的限制:
- RNN 允许信息沿着序列一路流动,但是它要经过许多处理步骤才能到达那里(限制梯度流动)。这些 RNN 步骤必须按顺序运行,因此 RNN 不太能够利用现代并行设备的优势。
- 在 CNN 中,每个位置都可以并行处理,但它只提供有限的接收场。接收场只随着 CNN 层数的增加而线性增长,需要叠加许多卷积层来跨序列传输信息(小波网通过使用扩张卷积来减少这个问题)。
the global self attention layer
允许每个序列元素直接访问每个其他序列元素,只需少量操作,并且所有输出都可以并行计算。 就像下图这样:
虽然图像类似于线性层,其本质好像也是线性层,但是其信息传播能力要比普通的线性层要强;
the causal self attention layer
:因果自注意层,这一层与the global self attention layer
类似,其不同的特点是需要mask
,Masked Multi-Head Attention
;这里要注意的是Transformer
是一个自回归模型,每次产生一个输出并且把输出当作输入用来继续产生输出,因此这些模型确保每个序列元素的输出仅依赖于先前的序列元素,所以需要对Attention weights
进行处理;
根据论文公式:
可以发现 Q 和 K是做矩阵运算的,Attention中的单个元素计算是Q中的某行和 K中的某行做点积运算;论文中是计算 Attention后在相应掩码位置乘一个无穷小,这里其实可以优化一下,只计算 Attention有效的位置,可以加快速度;
这一层根据训练和推理有一点不同,如上文所说,在训练时,我们不需要每一次新产生的输出做为输入,我们直接用真值(shift)代替新产生的输出进行训练就好,这样可以在缺失一点稳定性的情况下,加快训练速度,并得到每一个位置上的损失大小;
在推理时,我们并没有真值,我们只能以每一次新产生的输出作为输入进行计算,这里有两种自回归的处理方式;一个是RNN,一个是CNN:Fast Wavenet;
以下是该层的简要表示,即每个序列元素的输出仅依赖于先前的序列元素;
2.5 前馈神经网络
在encoder和decoder中,都包含了Feed Forward
网络,如图所示:
该网络由两个线性层组成。中间有一个relu
激活函数,还有一个dropout
层;这里面维度变化是把d_model
维先提升到dff
维度,然后把dff
维度降低到d_model
维度;
三、过程实现
3.1 安装包和导包
import tensorflow as tf import keras_nlp import matplotlib.pyplot as plt import numpy as np plt.rcParams['font.sans-serif']=['SimHei'] plt.rcParams['axes.unicode_minus']=False
3.2 数据准备
训练的过程如图所示:
从图中我们可以看出,我们需要准备好源sequence和两个目标sequence,其中源sequence前后应该有\
这里需要提醒的一点是,右侧训练过程是不依赖每一步的输出结果的:这种设置被称为“teacher forcing”,因为不管模型在每个时间步的输出如何,它都会获得下一个时间步的真值作为输入。这是训练文本生成模型的一种简单而有效的方法。这是有效的,因为不需要顺序运行模型,在不同的序列位置的输出可以并行计算。
You might have expected the `input, output`, pairs to simply be the `Portuguese, English` sequences. Given the Portuguese sequence, the model would try to generate the English sequence. It's possible to train a model that way. You'd need to write out the inference loop and pass the model's output back to the input. It's slower (time steps can't run in parallel), and a harder task to learn (the model can't get the end of a sentence right until it gets the beginning right), but it can give a more stable model because the model has to learn to correct its own errors during training.
总结就是 teacher forcing 通过舍弃模型的稳定性,加快学习训练速度,相关代码如下:
# 数据处理 def process_data(x): res = tf.strings.split(x, '\t') return res[1], res[3] # 导入数据 dataset = tf.data.TextLineDataset('ch-en.tsv') dataset = dataset.map(process_data) # 建立中英文wordpiece词表 vocab_chinese = keras_nlp.tokenizers.compute_word_piece_vocabulary( dataset.map(lambda x, y: x), vocabulary_size=20_000, lowercase=True, strip_accents=True, split_on_cjk=True, reserved_tokens=["[PAD]", "[START]", "[END]", "[MASK]", "[UNK]"], ) vocab_english = keras_nlp.tokenizers.compute_word_piece_vocabulary( dataset.map(lambda x, y: y), vocabulary_size=20_000, lowercase=True, strip_accents=True, split_on_cjk=True, reserved_tokens=["[PAD]", "[START]", "[END]", "[MASK]", "[UNK]"], ) # 构建分词器 chinese_tokenizer = keras_nlp.tokenizers.WordPieceTokenizer(vocabulary=vocab_chinese, oov_token="[UNK]") english_tokenizer = keras_nlp.tokenizers.WordPieceTokenizer(vocabulary=vocab_english, oov_token="[UNK]") # 再进行一次数据处理 def process_data_(ch, en, maxtoken=128): ch = chinese_tokenizer(ch)[:,:maxtoken] en = english_tokenizer(tf.strings.lower(en))[:,:maxtoken] ch = tf.concat([tf.ones(shape=(64,1), dtype='int32'), ch, tf.ones(shape=(64,1), dtype='int32')*2], axis=-1).to_tensor() en = tf.concat([tf.ones(shape=(64,1), dtype='int32'), en, tf.ones(shape=(64,1), dtype='int32')*2], axis=-1) en_inputs = en[:, :-1].to_tensor() # Drop the [END] tokens en_labels = en[:, 1:].to_tensor() # Drop the [START] tokens return (ch, en_inputs), en_labels dataset = dataset.batch(64).map(process_data_) train_dataset = dataset.take(1000) val_dataset = dataset.skip(500).take(300) # 数据准备完毕 查看数据 # for (pt, en), en_labels in dataset.take(1): # break # print(pt.shape) # print(en.shape) # print(en_labels.shape) # (64, 33) # (64, 28) # (64, 28)
3.3 词嵌入和位置编码
词嵌入和位置编码的地方一共有两处,一处是Input,一处是Output(shifted right);两处采用位置编码的方式是一样的;
位置编码函数如下:
def positional_encoding(length, depth): depth = depth/2 positions = np.arange(length)[:, np.newaxis] # (seq, 1) depths = np.arange(depth)[np.newaxis, :]/depth # (1, depth) angle_rates = 1 / (10000**depths) # (1, depth) angle_rads = positions * angle_rates # (pos, depth) pos_encoding = np.concatenate( [np.sin(angle_rads), np.cos(angle_rads)], axis=-1) return tf.cast(pos_encoding, dtype=tf.float32)
词嵌入和位置编码层的代码如下:
class PositionEmbedding(tf.keras.layers.Layer): def __init__(self, vocabulary_size, d_model): super().__init__() self.d_model = d_model self.embedding = tf.keras.layers.Embedding(vocabulary_size, d_model, mask_zero=True) self.pos_encoding = positional_encoding(length=2048, depth=d_model) def compute_mask(self, *args, **kwargs): return self.embedding.compute_mask(*args, **kwargs) def call(self, x): length = tf.shape(x)[1] x = self.embedding(x) x *= tf.math.sqrt(tf.cast(self.d_model, tf.float32)) x = x + self.pos_encoding[tf.newaxis, :length, :] return x
3.4 注意力机制
模型中Multi-Head Attention
有三个, 这三个分别对应三种Multi-Head Attention Layer
:the cross attention layer
,the global self attention layer
, the causal self attention layer
,如图所示:
定义一个BaseAttention
class BaseAttention(tf.keras.layers.Layer): def __init__(self, **kwargs): super().__init__() self.mha = tf.keras.layers.MultiHeadAttention(**kwargs) self.layernorm = tf.keras.layers.LayerNormalization() self.add = tf.keras.layers.Add()
the cross attention layer
定义如下:
class CrossAttention(BaseAttention): def call(self, x, context): attn_output, attn_scores = self.mha( query=x, key=context, value=context, return_attention_scores=True) # Cache the attention scores for plotting later. self.last_attn_scores = attn_scores x = self.add([x, attn_output]) x = self.layernorm(x) return x
the global self attention layer
定义如下:
class GlobalSelfAttention(BaseAttention): def call(self, x): attn_output = self.mha( query=x, value=x, key=x) x = self.add([x, attn_output]) x = self.layernorm(x) return x
the causal self attention layer
定义如下:
class CausalSelfAttention(BaseAttention): def call(self, x): attn_output = self.mha( query=x, value=x, key=x, use_causal_mask = True) x = self.add([x, attn_output]) x = self.layernorm(x) return x
3.5 前馈神经网络
该网络由两个线性层组成。中间有一个relu
激活函数,还有一个dropout
层;这里面维度变化是把d_model
维先提升到dff
维度,然后把dff
维度降低到d_model
维度;
代码如下:
class FeedForward(tf.keras.layers.Layer): def __init__(self, d_model, dff, dropout_rate=0.1): super().__init__() self.seq = tf.keras.Sequential([ tf.keras.layers.Dense(dff, activation='relu'), tf.keras.layers.Dense(d_model), tf.keras.layers.Dropout(dropout_rate) ]) self.add = tf.keras.layers.Add() self.layer_norm = tf.keras.layers.LayerNormalization() def call(self, x): x = self.add([x, self.seq(x)]) x = self.layer_norm(x) return x
3.6 编码器
模型结构如下,论文中编码器是由6个编码层组成的;
编码层代码如下:
class EncoderLayer(tf.keras.layers.Layer): def __init__(self, *, d_model, num_heads, dff, dropout=0.1): super().__init__() self.self_attention = GlobalSelfAttention( num_heads = num_heads, key_dim = d_model, dropout = dropout ) self.ffn = FeedForward(d_model, dff) def call(self, x): x = self.self_attention(x) x = self.ffn(x) return x
编码器代码如下:
class Encoder(tf.keras.layers.Layer): def __init__(self, *, vocabulary_size, d_model, nums_heads, dff, num_layers=6, dropout=0.1): super().__init__() # 给Encoder添加属性,便于辨识 self.d_model = d_model self.num_layers = num_layers self.pos_embedding = PositionEmbedding(vocabulary_size, d_model) self.encoder_layers = [EncoderLayer(d_model=d_model, num_heads=nums_heads, dff=dff, dropout=dropout) for _ in range(num_layers)] self.dropout = tf.keras.layers.Dropout(dropout) def call(self, x): x = self.pos_embedding(x) x = self.dropout(x) for encoder_layer in self.encoder_layers: x = encoder_layer(x) return x
3.7 解码器
模型结构如下,论文中解码器是由6个解码层组成的;
解码层代码如下:
class DecoderLayer(tf.keras.layers.Layer): def __init__(self, *, d_model, num_heads, dff, dropout=0.1): super().__init__() self.causal_self_attention = CausalSelfAttention(num_heads=num_heads, key_dim=d_model, dropout=dropout) self.cross_attention = CrossAttention(num_heads=num_heads, key_dim=key_dim, dropout=dropout) self.ffn = FeedForward(d_model, dff) def call(self, x, context): x = self.causal_self_attention(x) x = self.cross_attention(x, context) # 这里存储最后的注意力分数为了后面的画图 self.last_attn_scores = self.cross_attention.last_attn_scores x = self.ffn(x) return x
解码器代码如下:
class Decoder(tf.keras.layers.Layer): def __init__(self, *, vocabulary_size, d_model, num_heads, dff, num_layers=6, dropout=0.1): super(Decoder, self).__init__() self.d_model = d_model self.num_layers = num_layers self.pos_embedding = PositionEmbedding(vocabulary_size=vocabulary_size, d_model=d_model) self.decoder_layers = [DecoderLayer(d_model=d_model, num_heads=num_heads, dff=dff, dropout=dropout) for _ in range(num_layers)] self.dropout = tf.keras.layers.Dropout(rate=dropout) self.last_attn_scores = None def call(self, x, content): x = self.pos_embedding(x) x = self.dropout(x) for decoder_layer in self.decoder_layers: x = decoder_layer(x, content) self.last_attn_scores = self.decoder_layers[-1].last_attn_scores return x
3.8 Transformer
有了编码器和解码器,现在我们来构造Transformer
模型,模型的整体框架如下:
需要设置的超参如下:
num_layers = 4 d_model = 128 dff = 512 num_heads = 8 dropout = 0.1
Transformer模型代码如下:
class Transformer(tf.keras.Model): def __init__(self, *, num_layers, d_model, num_heads, dff, input_vocabulary_size, target_vocabulary_size, dropout=0.1): super().__init__() self.encoder = Encoder(vocabulary_size=input_vocabulary_size, d_model=d_model, num_layers=num_layers, num_heads=num_heads, dff=dff) self.decoder = Decoder(vocabulary_size=target_vocabulary_size, d_model=d_model, num_layers=num_layers, num_heads=num_heads, dff=dff) self.final_layer = tf.keras.layers.Dense(target_vocabulary_size, activation='softmax') def call(self, inputs): context, x = inputs context = self.encoder(context) x = self.decoder(x, context) logits = self.final_layer(x) # 不太理解 try: # Drop the keras mask, so it doesn't scale the losses/metrics. # b/250038731 del logits._keras_mask except AttributeError: pass return logits
以下是不同层数的Transformer和RNN+Attention模型的可视化模型:
模型参数大小如下:
model.summary()
如果出现:This model has not yet been built. Build the model first by calling build()
or by calling the model on a batch of data.,试着带入数据build一下模型
model = Transformer(num_layers=num_layers, d_model=d_model, num_heads=num_heads, dff=dff, input_vocabulary_size=chinese_tokenizer.vocabulary_size(), target_vocabulary_size=english_tokenizer.vocabulary_size(), dropout=dropout) output = model((pt, en)) print(en.shape) print(pt.shape) print(output.shape)
3.9 训练
根据原文的公式自定义一个学习率调度器:
代码如下:
class CustomSchedule(tf.keras.optimizers.schedules.LearningRateSchedule): def __init__(self, d_model, warmup_steps=4000): super().__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): step = tf.cast(step, dtype=tf.float32) 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)
调度曲线如下所示:
定义loss和metrics:
def masked_loss(label, pred): mask = label != 0 loss_object = tf.keras.losses.SparseCategoricalCrossentropy(reduction='none') loss = loss_object(label, pred) mask = tf.cast(mask, dtype=loss.dtype) loss *= mask loss = tf.reduce_sum(loss)/tf.reduce_sum(mask) return loss def masked_accuracy(label, pred): pred = tf.argmax(pred, axis=2) label = tf.cast(label, pred.dtype) match = label == pred mask = label != 0 match = match & mask match = tf.cast(match, dtype=tf.float32) mask = tf.cast(mask, dtype=tf.float32) return tf.reduce_sum(match)/tf.reduce_sum(mask)
开始训练:
model.compile(loss=masked_loss, optimizer=optimizer, metrics=[masked_accuracy]) model.fit(train_dataset, epochs=10, validation_data=val_dataset)
3.10 推理
如上文所说,由于推理过程于训练过程不一致,我们需要重新写一个推理的代码,代码如下:
MAX_TOKENS = 128 class Translator(tf.Module): def __init__(self, tokenizers, transformer): self.tokenizers = tokenizers self.transformer = transformer def __call__(self, sentence, max_length=MAX_TOKENS): # sentence是中文,因此需要tokenizer并且加上<start>:1和<end>:2 assert isinstance(sentence, tf.Tensor) if len(sentence.shape) == 0: sentence = sentence[tf.newaxis] sentence = self.tokenizers(sentence) sentence = tf.concat([tf.ones(shape=[sentence.shape[0], 1], dtype='int32'), sentence, tf.ones(shape=[sentence.shape[0], 1], dtype='int32')*2], axis=-1).to_tensor() encoder_input = sentence # As the output language is English, initialize the output with the start = tf.constant(1, dtype='int64')[tf.newaxis] end = tf.constant(2, dtype='int64')[tf.newaxis] # tf.TensorArray 类似于python中的列表 output_array = tf.TensorArray(dtype=tf.int64, size=0, dynamic_size=True) # 在index=0的位置写入start output_array = output_array.write(0, start) for i in tf.range(max_length): output = tf.transpose(output_array.stack()) predictions = self.transformer([encoder_input, output], training=False) # Shape `(batch_size, seq_len, vocab_size)` # 从seq_len中的最后一个维度选择last token predictions = predictions[:, -1:, :] # Shape `(batch_size, 1, vocab_size)`. predicted_id = tf.argmax(predictions, axis=-1) # `predicted_id`加入到output_array中作为一个新的输入 output_array = output_array.write(i+1, predicted_id[0]) # 如果输出end就表明停止 if predicted_id == end: break output = tf.transpose(output_array.stack()) # 重新计算一下最外面的循环,得到最后的注意力得分 self.transformer([encoder_input, output[:,:-1]], training=False) attention_weights = self.transformer.decoder.last_attn_scores lst = [] for item in output[0].numpy(): lst.append(english_tokenizer.vocabulary[item]) translated_text = ''.join(lst) translated_tokens = output[0] return translated_text, translated_tokens, attention_weights``` 推理测试: ```python def print_translation(sentence, tokens, ground_truth): print(f'{"Input:":15s}: {sentence}') print(f'{"Prediction":15s}: {tokens}') print(f'{"Ground truth":15s}: {ground_truth}') translator = Translator(chinese_tokenizer, model) sentence = '我們試試看!' ground_truth = "Let's try it." translated_text, translated_tokens, attention_weights = translator(tf.constant(sentence)) print_translation(sentence, translated_text, ground_truth) # Input: : 我們試試看! # Prediction : [START] let ' s try to see this . [END] # Ground truth : Let's try it.
3.11 注意力可视化
定义画注意力权重的函数:
def plot_attention_head(in_tokens, translated_tokens, attention): # 模型在输出中不产生<START>,我们直接忽略 translated_tokens = translated_tokens[1:] ax = plt.gca() ax.matshow(attention[0]) ax.set_xticks(range(len(in_tokens))) ax.set_yticks(range(len(translated_tokens))) labels = [vocab_chinese[label] for label in in_tokens.numpy()] ax.set_xticklabels(labels, rotation=90) labels = [vocab_english[label] for label in translated_tokens.numpy()] ax.set_yticklabels(labels)
测试代码如下:
sentence = '我們試試看!' ground_truth = "Let's try it." translated_text, translated_tokens, attention_weights = translator(tf.constant(sentence)) in_tokens = tf.concat([tf.constant(1)[tf.newaxis], chinese_tokenizer(tf.constant(sentence)), tf.constant(2)[tf.newaxis]], axis=-1) attention = tf.squeeze(attention_weights, 0) # 画第0个头 attention[0] 这里一共有8个头 plot_attention_head(in_tokens, translated_tokens, attention[0])
第0个头结果如下:
四、整体总结
代码有点长!