PS：要转载请注明出处，本人版权所有。
PS: 这个只是基于《我自己》的理解，
如果和你的原则及想法相冲突，请谅解，勿喷。
环境说明
无
前言
本文是这个系列第七篇，它们是：
- 《大模型基础补全计划(一)---重温一些深度学习相关的数学知识》 https://www.cnblogs.com/Iflyinsky/p/18717317
- 《大模型基础补全计划(二)---词嵌入(word embedding) 》 https://www.cnblogs.com/Iflyinsky/p/18775451
- 《大模型基础补全计划(三)---RNN实例与测试》 https://www.cnblogs.com/Iflyinsky/p/18967569
- 《大模型基础补全计划(四)---LSTM的实例与测试(RNN的改进)》 https://www.cnblogs.com/Iflyinsky/p/19091089
- 《大模型基础补全计划(五)---seq2seq实例与测试(编码器、解码器架构)》 https://www.cnblogs.com/Iflyinsky/p/19150535
- 《大模型基础补全计划(六)---带注意力机制的seq2seq实例与测试(Bahdanau Attention)》 https://www.cnblogs.com/Iflyinsky/p/19184558
本文的核心是介绍transformer模型结构，下面是transformer的网络结构示意图（图来源：见参考文献部分）。
从上面的架构图可以知道，在开始介绍之前，需要提前介绍多头注意力、自注意力、位置编码等前置知识。
点积注意力与自注意力
首先我们来介绍一种新的注意力评分方式，点积注意力，其计算公式是：$$\text{Attention}(Q, K, V) = \text{Softmax}\left(\frac{Q K^T}{\sqrt{d_k}}\right) V$$。
回到前面文章中的seq2seq中的注意力机制（一种加法注意力评分方式），其KV来自于encoder的output，Q来自于decoder的隐藏态。这个时候，我们假设一下，如果QKV都是同一种数据，那么每一次Q，都会输出对整个KV（也就是Q本身）的注意力，这种特殊的注意力被称为自注意力。
下面是点积注意力的代码，当QKV都是同一个输入时，下面的注意力就是自注意力。
python
class DotProductAttention(nn.Module): #@save
"""Scaled dot product attention."""
def __init__(self, dropout):
super().__init__()
self.dropout = nn.Dropout(dropout)
# Shape of queries: (batch_size, no. of queries, d)
# Shape of keys: (batch_size, no. of key-value pairs, d)
# Shape of values: (batch_size, no. of key-value pairs, value dimension)
# Shape of valid_lens: (batch_size,) or (batch_size, no. of queries)
def forward(self, queries, keys, values, valid_lens=None):
d = queries.shape[-1]
# Swap the last two dimensions of keys with keys.transpose(1, 2)
scores = torch.bmm(queries, keys.transpose(1, 2)) / math.sqrt(d)
self.attention_weights = masked_softmax(scores, valid_lens)
return torch.bmm(self.dropout(self.attention_weights), values)位置编码
我们知道，我们的序列数据中的每个数据都是在序列中有位置信息的，根据点积注意力的并行计算的实现，我们知道每个序列数据在同一时间进行了运算，没有序列之间的顺序信息。为了让我们的并行计算过程中，让模型感受到序列的顺序信息，因此我们需要在输入数据中含有位置信息，因此有人设计了位置编码。其代码实现如下：
python
class PositionalEncoding(nn.Module): #@save
"""Positional encoding."""
def __init__(self, num_hiddens, dropout, max_len=1000):
super().__init__()
self.dropout = nn.Dropout(dropout)
# Create a long enough P
self.P = torch.zeros((1, max_len, num_hiddens))
X = torch.arange(max_len, dtype=torch.float32).reshape(
-1, 1) / torch.pow(10000, torch.arange(
0, num_hiddens, 2, dtype=torch.float32) / num_hiddens)
self.P[:, :, 0::2] = torch.sin(X)
self.P[:, :, 1::2] = torch.cos(X)
def forward(self, X):
X = X + self.P[:, :X.shape[1], :].to(X.device)
return self.dropout(X)当我们的序列数据经过了位置编码后，在进行点积注意力计算时，我们的输入数据有了顺序信息，会让我们的模型学习到序列顺序相关的信息。
多头注意力
注意力机制已经可以对一个数据进行有侧重的关注。但是我们希望的是，注意力机制可以对数据的多个维度的侧重关注，因为我们的数据有很多的不同维度的属性信息。例如：一句英文，其有语法信息、有语境信息、有单词之间的信息等等。
基于这里提到的问题，有人提出了多头注意力机制。从上面的介绍来看，很好理解这个机制，就是每个头单独分析数据的属性，这样我们可以同时关注数据的多个维度的属性，提升我们的模型的理解能力。
其代码实现如下：
python
class MultiHeadAttention(nn.Module): #@save
"""Multi-head attention."""
def __init__(self, num_hiddens, num_heads, dropout, bias=False, **kwargs):
super().__init__()
self.num_heads = num_heads
self.attention = DotProductAttention(dropout)
self.W_q = nn.LazyLinear(num_hiddens, bias=bias)
self.W_k = nn.LazyLinear(num_hiddens, bias=bias)
self.W_v = nn.LazyLinear(num_hiddens, bias=bias)
self.W_o = nn.LazyLinear(num_hiddens, bias=bias)
def transpose_qkv(self, X):
"""Transposition for parallel computation of multiple attention heads."""
# Shape of input X: (batch_size, no. of queries or key-value pairs,
# num_hiddens). Shape of output X: (batch_size, no. of queries or
# key-value pairs, num_heads, num_hiddens / num_heads)
X = X.reshape(X.shape[0], X.shape[1], self.num_heads, -1)
# Shape of output X: (batch_size, num_heads, no. of queries or key-value
# pairs, num_hiddens / num_heads)
X = X.permute(0, 2, 1, 3)
# Shape of output: (batch_size * num_heads, no. of queries or key-value
# pairs, num_hiddens / num_heads)
return X.reshape(-1, X.shape[2], X.shape[3])
def transpose_output(self, X):
"""Reverse the operation of transpose_qkv."""
X = X.reshape(-1, self.num_heads, X.shape[1], X.shape[2])
X = X.permute(0, 2, 1, 3)
return X.reshape(X.shape[0], X.shape[1], -1)
def forward(self, queries, keys, values, valid_lens):
# Shape of queries, keys, or values:
# (batch_size, no. of queries or key-value pairs, num_hiddens)
# Shape of valid_lens: (batch_size,) or (batch_size, no. of queries)
# After transposing, shape of output queries, keys, or values:
# (batch_size * num_heads, no. of queries or key-value pairs,
# num_hiddens / num_heads)
queries = self.transpose_qkv(self.W_q(queries))
keys = self.transpose_qkv(self.W_k(keys))
values = self.transpose_qkv(self.W_v(values))
if valid_lens is not None:
# On axis 0, copy the first item (scalar or vector) for num_heads
# times, then copy the next item, and so on
valid_lens = torch.repeat_interleave(
valid_lens, repeats=self.num_heads, dim=0)
# Shape of output: (batch_size * num_heads, no. of queries,
# num_hiddens / num_heads)
output = self.attention(queries, keys, values, valid_lens)
# Shape of output_concat: (batch_size, no. of queries, num_hiddens)
output_concat = self.transpose_output(output)
return self.W_o(output_concat)上面的代码透露了一个问题，多头注意力并不是简单的创建N个相同的注意力进行运算，而是通过nn.LazyLinear投影后，在num_hiddens维度进行num_heads个数的划分，注意经过nn.LazyLinear后，num_hiddens维度的每一个数据其实都和输入的数据有关联，因此这个时候进行num_heads个数的划分是有效的，因为这个时候每个num_heads的组都携带了输入数据的全部信息。
位置前馈网络
引入非线性计算，加强网络认知能力。代码如下：
python
class PositionWiseFFN(nn.Module): #@save
"""The positionwise feed-forward network."""
def __init__(self, ffn_num_hiddens, ffn_num_outputs):
super().__init__()
self.dense1 = nn.LazyLinear(ffn_num_hiddens)
self.relu = nn.ReLU()
self.dense2 = nn.LazyLinear(ffn_num_outputs)
def forward(self, X):
return self.dense2(self.relu(self.dense1(X)))残差连接和层归一化
这个结构主要将原始输入叠加到一个其他计算（例如注意力）的输出上面，这样可以保证输出不会丢失原始输入信息，这个在网络层数大的情况下有奇效。代码如下：
python
class AddNorm(nn.Module): #@save
"""The residual connection followed by layer normalization."""
def __init__(self, norm_shape, dropout):
super().__init__()
self.dropout = nn.Dropout(dropout)
self.ln = nn.LayerNorm(norm_shape)
def forward(self, X, Y):
return self.ln(self.dropout(Y) + X)Transformer Encoder结构
下面是transformer-Encoder部分的代码
python
class TransformerEncoderBlock(nn.Module): #@save
"""The Transformer encoder block."""
def __init__(self, num_hiddens, ffn_num_hiddens, num_heads, dropout,
use_bias=False):
super().__init__()
self.attention = MultiHeadAttention(num_hiddens, num_heads,
dropout, use_bias)
self.addnorm1 = AddNorm(num_hiddens, dropout)
self.ffn = PositionWiseFFN(ffn_num_hiddens, num_hiddens)
self.addnorm2 = AddNorm(num_hiddens, dropout)
def forward(self, X, valid_lens):
Y = self.addnorm1(X, self.attention(X, X, X, valid_lens))
return self.addnorm2(Y, self.ffn(Y))从代码中可以知道，其计算过程就是多头注意力、残差连接及层归一化、位置前馈网络、残差连接及层归一化的过程。
Transformer Decoder结构
下面是transformer-Decoder部分的代码
python
class TransformerDecoderBlock(nn.Module):
# The i-th block in the Transformer decoder
def __init__(self, num_hiddens, ffn_num_hiddens, num_heads, dropout, i):
super().__init__()
self.i = i
self.attention1 = MultiHeadAttention(num_hiddens, num_heads,
dropout)
self.addnorm1 = AddNorm(num_hiddens, dropout)
self.attention2 = MultiHeadAttention(num_hiddens, num_heads,
dropout)
self.addnorm2 = AddNorm(num_hiddens, dropout)
self.ffn = PositionWiseFFN(ffn_num_hiddens, num_hiddens)
self.addnorm3 = AddNorm(num_hiddens, dropout)
def forward(self, X, state):
enc_outputs, enc_valid_lens = state[0], state[1]
# During training, all the tokens of any output sequence are processed
# at the same time, so state[2][self.i] is None as initialized. When
# decoding any output sequence token by token during prediction,
# state[2][self.i] contains representations of the decoded output at
# the i-th block up to the current time step
if state[2][self.i] is None:
key_values = X
else:
key_values = torch.cat((state[2][self.i], X), dim=1)
state[2][self.i] = key_values
if self.training:
batch_size, num_steps, _ = X.shape
# Shape of dec_valid_lens: (batch_size, num_steps), where every
# row is [1, 2, ..., num_steps]
dec_valid_lens = torch.arange(
1, num_steps + 1, device=X.device).repeat(batch_size, 1)
else:
dec_valid_lens = None
# Self-attention
X2 = self.attention1(X, key_values, key_values, dec_valid_lens)
Y = self.addnorm1(X, X2)
# Encoder-decoder attention. Shape of enc_outputs:
# (batch_size, num_steps, num_hiddens)
Y2 = self.attention2(Y, enc_outputs, enc_outputs, enc_valid_lens)
Z = self.addnorm2(Y, Y2)
return self.addnorm3(Z, self.ffn(Z)), state从代码中可以知道，其计算过程就是多头注意力、残差连接及层归一化、多头注意力、残差连接及层归一化、位置前馈网络、残差连接及层归一化的过程。
基于transformer的类似seq2seq 英文翻译中文 的实例
关于dataset部分的内容，请参考前面seq2seq相关文章。
完整代码如下
python
import os
import random
import torch
import math
from torch import nn
from torch.nn import functional as F
import numpy as np
import time
import visdom
import collections
import dataset
class Accumulator:
"""在n个变量上累加"""
def __init__(self, n):
"""Defined in :numref:`sec_softmax_scratch`"""
self.data = [0.0] * n
def add(self, *args):
self.data = [a + float(b) for a, b in zip(self.data, args)]
def reset(self):
self.data = [0.0] * len(self.data)
def __getitem__(self, idx):
return self.data[idx]
class Timer:
"""记录多次运行时间"""
def __init__(self):
"""Defined in :numref:`subsec_linear_model`"""
self.times = []
self.start()
def start(self):
"""启动计时器"""
self.tik = time.time()
def stop(self):
"""停止计时器并将时间记录在列表中"""
self.times.append(time.time() - self.tik)
return self.times[-1]
def avg(self):
"""返回平均时间"""
return sum(self.times) / len(self.times)
def sum(self):
"""返回时间总和"""
return sum(self.times)
def cumsum(self):
"""返回累计时间"""
return np.array(self.times).cumsum().tolist()
class Encoder(nn.Module):
"""编码器-解码器架构的基本编码器接口"""
def __init__(self, **kwargs):
# 调用父类nn.Module的构造函数，确保正确初始化
super(Encoder, self).__init__(**kwargs)
def forward(self, X, *args):
# 抛出未实现错误，意味着该方法需要在子类中具体实现
raise NotImplementedError
class Decoder(nn.Module):
"""编码器-解码器架构的基本解码器接口
Defined in :numref:`sec_encoder-decoder`"""
def __init__(self, **kwargs):
# 调用父类nn.Module的构造函数，确保正确初始化
super(Decoder, self).__init__(**kwargs)
def init_state(self, enc_outputs, *args):
# 抛出未实现错误，意味着该方法需要在子类中具体实现
raise NotImplementedError
def forward(self, X, state):
# 抛出未实现错误，意味着该方法需要在子类中具体实现
raise NotImplementedError
class EncoderDecoder(nn.Module):
"""编码器-解码器架构的基类
Defined in :numref:`sec_encoder-decoder`"""
def __init__(self, encoder, decoder, **kwargs):
# 调用父类nn.Module的构造函数，确保正确初始化
super(EncoderDecoder, self).__init__(**kwargs)
# 将传入的编码器实例赋值给类的属性
self.encoder = encoder
# 将传入的解码器实例赋值给类的属性
self.decoder = decoder
def forward(self, enc_X, dec_X, enc_X_valid_len, *args):
# 调用编码器的前向传播方法，处理输入的编码器输入数据enc_X
enc_outputs = self.encoder(enc_X, enc_X_valid_len, *args)
# 调用解码器的init_state方法，根据编码器的输出初始化解码器的状态
dec_state = self.decoder.init_state(enc_outputs, enc_X_valid_len)
# 调用解码器的前向传播方法，处理输入的解码器输入数据dec_X和初始化后的状态
return self.decoder(dec_X, dec_state)
def masked_softmax(X, valid_lens): #@save
"""Perform softmax operation by masking elements on the last axis."""
# X: 3D tensor, valid_lens: 1D or 2D tensor
def _sequence_mask(X, valid_len, value=0):
maxlen = X.size(1)
mask = torch.arange((maxlen), dtype=torch.float32,
device=X.device)[None, :] < valid_len[:, None]
X[~mask] = value
return X
if valid_lens is None:
return nn.functional.softmax(X, dim=-1)
else:
shape = X.shape
if valid_lens.dim() == 1:
valid_lens = torch.repeat_interleave(valid_lens, shape[1])
else:
valid_lens = valid_lens.reshape(-1)
# On the last axis, replace masked elements with a very large negative
# value, whose exponentiation outputs 0
X = _sequence_mask(X.reshape(-1, shape[-1]), valid_lens, value=-1e6)
return nn.functional.softmax(X.reshape(shape), dim=-1)
class DotProductAttention(nn.Module): #@save
"""Scaled dot product attention."""
def __init__(self, dropout):
super().__init__()
self.dropout = nn.Dropout(dropout)
# Shape of queries: (batch_size, no. of queries, d)
# Shape of keys: (batch_size, no. of key-value pairs, d)
# Shape of values: (batch_size, no. of key-value pairs, value dimension)
# Shape of valid_lens: (batch_size,) or (batch_size, no. of queries)
def forward(self, queries, keys, values, valid_lens=None):
d = queries.shape[-1]
# Swap the last two dimensions of keys with keys.transpose(1, 2)
scores = torch.bmm(queries, keys.transpose(1, 2)) / math.sqrt(d)
self.attention_weights = masked_softmax(scores, valid_lens)
return torch.bmm(self.dropout(self.attention_weights), values)
class MultiHeadAttention(nn.Module): #@save
"""Multi-head attention."""
def __init__(self, num_hiddens, num_heads, dropout, bias=False, **kwargs):
super().__init__()
self.num_heads = num_heads
self.attention = DotProductAttention(dropout)
self.W_q = nn.LazyLinear(num_hiddens, bias=bias)
self.W_k = nn.LazyLinear(num_hiddens, bias=bias)
self.W_v = nn.LazyLinear(num_hiddens, bias=bias)
self.W_o = nn.LazyLinear(num_hiddens, bias=bias)
def transpose_qkv(self, X):
"""Transposition for parallel computation of multiple attention heads."""
# Shape of input X: (batch_size, no. of queries or key-value pairs,
# num_hiddens). Shape of output X: (batch_size, no. of queries or
# key-value pairs, num_heads, num_hiddens / num_heads)
X = X.reshape(X.shape[0], X.shape[1], self.num_heads, -1)
# Shape of output X: (batch_size, num_heads, no. of queries or key-value
# pairs, num_hiddens / num_heads)
X = X.permute(0, 2, 1, 3)
# Shape of output: (batch_size * num_heads, no. of queries or key-value
# pairs, num_hiddens / num_heads)
return X.reshape(-1, X.shape[2], X.shape[3])
def transpose_output(self, X):
"""Reverse the operation of transpose_qkv."""
X = X.reshape(-1, self.num_heads, X.shape[1], X.shape[2])
X = X.permute(0, 2, 1, 3)
return X.reshape(X.shape[0], X.shape[1], -1)
def forward(self, queries, keys, values, valid_lens):
# Shape of queries, keys, or values:
# (batch_size, no. of queries or key-value pairs, num_hiddens)
# Shape of valid_lens: (batch_size,) or (batch_size, no. of queries)
# After transposing, shape of output queries, keys, or values:
# (batch_size * num_heads, no. of queries or key-value pairs,
# num_hiddens / num_heads)
queries = self.transpose_qkv(self.W_q(queries))
keys = self.transpose_qkv(self.W_k(keys))
values = self.transpose_qkv(self.W_v(values))
if valid_lens is not None:
# On axis 0, copy the first item (scalar or vector) for num_heads
# times, then copy the next item, and so on
valid_lens = torch.repeat_interleave(
valid_lens, repeats=self.num_heads, dim=0)
# Shape of output: (batch_size * num_heads, no. of queries,
# num_hiddens / num_heads)
output = self.attention(queries, keys, values, valid_lens)
# Shape of output_concat: (batch_size, no. of queries, num_hiddens)
output_concat = self.transpose_output(output)
return self.W_o(output_concat)
class PositionWiseFFN(nn.Module): #@save
"""The positionwise feed-forward network."""
def __init__(self, ffn_num_hiddens, ffn_num_outputs):
super().__init__()
self.dense1 = nn.LazyLinear(ffn_num_hiddens)
self.relu = nn.ReLU()
self.dense2 = nn.LazyLinear(ffn_num_outputs)
def forward(self, X):
return self.dense2(self.relu(self.dense1(X)))
class AddNorm(nn.Module): #@save
"""The residual connection followed by layer normalization."""
def __init__(self, norm_shape, dropout):
super().__init__()
self.dropout = nn.Dropout(dropout)
self.ln = nn.LayerNorm(norm_shape)
def forward(self, X, Y):
return self.ln(self.dropout(Y) + X)
class TransformerEncoderBlock(nn.Module): #@save
"""The Transformer encoder block."""
def __init__(self, num_hiddens, ffn_num_hiddens, num_heads, dropout,
use_bias=False):
super().__init__()
self.attention = MultiHeadAttention(num_hiddens, num_heads,
dropout, use_bias)
self.addnorm1 = AddNorm(num_hiddens, dropout)
self.ffn = PositionWiseFFN(ffn_num_hiddens, num_hiddens)
self.addnorm2 = AddNorm(num_hiddens, dropout)
def forward(self, X, valid_lens):
Y = self.addnorm1(X, self.attention(X, X, X, valid_lens))
return self.addnorm2(Y, self.ffn(Y))
class PositionalEncoding(nn.Module): #@save
"""Positional encoding."""
def __init__(self, num_hiddens, dropout, max_len=1000):
super().__init__()
self.dropout = nn.Dropout(dropout)
# Create a long enough P
self.P = torch.zeros((1, max_len, num_hiddens))
X = torch.arange(max_len, dtype=torch.float32).reshape(
-1, 1) / torch.pow(10000, torch.arange(
0, num_hiddens, 2, dtype=torch.float32) / num_hiddens)
self.P[:, :, 0::2] = torch.sin(X)
self.P[:, :, 1::2] = torch.cos(X)
def forward(self, X):
X = X + self.P[:, :X.shape[1], :].to(X.device)
return self.dropout(X)
class TransformerEncoder(Encoder): #@save
"""The Transformer encoder."""
def __init__(self, vocab_size, num_hiddens, ffn_num_hiddens,
num_heads, num_blks, dropout, use_bias=False):
super().__init__()
self.num_hiddens = num_hiddens
self.embedding = nn.Embedding(vocab_size, num_hiddens)
self.pos_encoding = PositionalEncoding(num_hiddens, dropout)
self.blks = nn.Sequential()
for i in range(num_blks):
self.blks.add_module("block"+str(i), TransformerEncoderBlock(
num_hiddens, ffn_num_hiddens, num_heads, dropout, use_bias))
def forward(self, X, valid_lens):
# Since positional encoding values are between -1 and 1, the embedding
# values are multiplied by the square root of the embedding dimension
# to rescale before they are summed up
# X[batch_size, seq_len, num_hidden]
X = self.pos_encoding(self.embedding(X) * math.sqrt(self.num_hiddens))
self.attention_weights = [None] * len(self.blks)
for i, blk in enumerate(self.blks):
X = blk(X, valid_lens)
self.attention_weights[i] = blk.attention.attention.attention_weights
# X[batch_size, seq_len, num_hidden]
return X
class TransformerDecoderBlock(nn.Module):
# The i-th block in the Transformer decoder
def __init__(self, num_hiddens, ffn_num_hiddens, num_heads, dropout, i):
super().__init__()
self.i = i
self.attention1 = MultiHeadAttention(num_hiddens, num_heads,
dropout)
self.addnorm1 = AddNorm(num_hiddens, dropout)
self.attention2 = MultiHeadAttention(num_hiddens, num_heads,
dropout)
self.addnorm2 = AddNorm(num_hiddens, dropout)
self.ffn = PositionWiseFFN(ffn_num_hiddens, num_hiddens)
self.addnorm3 = AddNorm(num_hiddens, dropout)
def forward(self, X, state):
enc_outputs, enc_valid_lens = state[0], state[1]
# During training, all the tokens of any output sequence are processed
# at the same time, so state[2][self.i] is None as initialized. When
# decoding any output sequence token by token during prediction,
# state[2][self.i] contains representations of the decoded output at
# the i-th block up to the current time step
if state[2][self.i] is None:
key_values = X
else:
key_values = torch.cat((state[2][self.i], X), dim=1)
state[2][self.i] = key_values
if self.training:
batch_size, num_steps, _ = X.shape
# Shape of dec_valid_lens: (batch_size, num_steps), where every
# row is [1, 2, ..., num_steps]
dec_valid_lens = torch.arange(
1, num_steps + 1, device=X.device).repeat(batch_size, 1)
else:
dec_valid_lens = None
# Self-attention
X2 = self.attention1(X, key_values, key_values, dec_valid_lens)
Y = self.addnorm1(X, X2)
# Encoder-decoder attention. Shape of enc_outputs:
# (batch_size, num_steps, num_hiddens)
Y2 = self.attention2(Y, enc_outputs, enc_outputs, enc_valid_lens)
Z = self.addnorm2(Y, Y2)
return self.addnorm3(Z, self.ffn(Z)), state
class TransformerDecoder(Decoder):
def __init__(self, vocab_size, num_hiddens, ffn_num_hiddens, num_heads,
num_blks, dropout):
super().__init__()
self.num_hiddens = num_hiddens
self.num_blks = num_blks
self.embedding = nn.Embedding(vocab_size, num_hiddens)
self.pos_encoding = PositionalEncoding(num_hiddens, dropout)
self.blks = nn.Sequential()
for i in range(num_blks):
self.blks.add_module("block"+str(i), TransformerDecoderBlock(
num_hiddens, ffn_num_hiddens, num_heads, dropout, i))
self.dense = nn.LazyLinear(vocab_size)
def init_state(self, enc_outputs, enc_valid_lens):
return [enc_outputs, enc_valid_lens, [None] * self.num_blks]
def forward(self, X, state):
X = self.pos_encoding(self.embedding(X) * math.sqrt(self.num_hiddens))
self._attention_weights = [[None] * len(self.blks) for _ in range (2)]
for i, blk in enumerate(self.blks):
X, state = blk(X, state)
# Decoder self-attention weights
self._attention_weights[0][
i] = blk.attention1.attention.attention_weights
# Encoder-decoder attention weights
self._attention_weights[1][
i] = blk.attention2.attention.attention_weights
return self.dense(X), state
@property
def attention_weights(self):
return self._attention_weights
def sequence_mask(X, valid_len, value=0):
"""在序列中屏蔽不相关的项"""
maxlen = X.size(1)
mask = torch.arange((maxlen), dtype=torch.float32,
device=X.device)[None, :] < valid_len[:, None]
X[~mask] = value
return X
class MaskedSoftmaxCELoss(nn.CrossEntropyLoss):
"""带遮蔽的softmax交叉熵损失函数"""
# pred的形状：(batch_size,num_steps,vocab_size)
# label的形状：(batch_size,num_steps)
# valid_len的形状：(batch_size,)
def forward(self, pred, label, valid_len):
weights = torch.ones_like(label)
weights = sequence_mask(weights, valid_len)
self.reduction='none'
unweighted_loss = super(MaskedSoftmaxCELoss, self).forward(
pred.permute(0, 2, 1), label)
weighted_loss = (unweighted_loss * weights).mean(dim=1)
return weighted_loss
def grad_clipping(net, theta): #@save
"""裁剪梯度"""
if isinstance(net, nn.Module):
params = [p for p in net.parameters() if p.requires_grad]
else:
params = net.params
norm = torch.sqrt(sum(torch.sum((p.grad ** 2)) for p in params))
if norm > theta:
for param in params:
param.grad[:] *= theta / norm
def train_seq2seq(net, data_iter, lr, num_epochs, tgt_vocab, device):
"""训练序列到序列模型"""
def xavier_init_weights(m):
if type(m) == nn.Linear:
nn.init.xavier_uniform_(m.weight)
if type(m) == nn.GRU:
for param in m._flat_weights_names:
if "weight" in param:
nn.init.xavier_uniform_(m._parameters[param])
net.apply(xavier_init_weights)
net.to(device)
optimizer = torch.optim.Adam(net.parameters(), lr=lr)
loss = MaskedSoftmaxCELoss()
net.train()
vis = visdom.Visdom(env=u'test1', server="http://127.0.0.1", port=8097)
animator = vis
for epoch in range(num_epochs):
timer = Timer()
metric = Accumulator(2) # 训练损失总和，词元数量
for batch in data_iter:
#清零（reset）优化器中的梯度缓存
optimizer.zero_grad()
# x.shape = [batch_size, num_steps]
X, X_valid_len, Y, Y_valid_len = [x.to(device) for x in batch]
# bos.shape = batch_size 个 bos-id
bos = torch.tensor([tgt_vocab['<bos>']] * Y.shape[0],
device=device).reshape(-1, 1)
# dec_input.shape = (batch_size, num_steps)
# 解码器的输入通常由序列的起始标志 bos 和目标序列（去掉末尾的部分 Y[:, :-1]）组成。
dec_input = torch.cat([bos, Y[:, :-1]], 1) # 强制教学
# Y_hat的形状:(batch_size,num_steps,vocab_size)
Y_hat, _ = net(X, dec_input, X_valid_len)
l = loss(Y_hat, Y, Y_valid_len)
l.sum().backward() # 损失函数的标量进行"反向传播"
grad_clipping(net, 1)
num_tokens = Y_valid_len.sum()
optimizer.step()
with torch.no_grad():
metric.add(l.sum(), num_tokens)
if (epoch + 1) % 10 == 0:
# print(predict('你是？'))
# print(epoch)
# animator.add(epoch + 1, )
if epoch == 9:
# 清空图表：使用空数组来替换现有内容
vis.line(X=np.array([0]), Y=np.array([0]), win='train_ch8', update='replace')
# _loss_val = l
# _loss_val = _loss_val.cpu().sum().detach().numpy()
vis.line(
X=np.array([epoch + 1]),
Y=[ metric[0] / metric[1]],
win='train_ch8',
update='append',
opts={
'title': 'train_ch8',
'xlabel': 'epoch',
'ylabel': 'loss',
'linecolor': np.array([[0, 0, 255]]), # 蓝色线条
}
)
print(f'loss {metric[0] / metric[1]:.3f}, {metric[1] / timer.stop():.1f} '
f'tokens/sec on {str(device)}')
torch.save(net.cpu().state_dict(), 'model_h.pt') # [[6]]
torch.save(net.cpu(), 'model.pt') # [[6]]
def predict_seq2seq(net, src_sentence, src_vocab, tgt_vocab, num_steps,
device, save_attention_weights=False):
"""序列到序列模型的预测"""
# 在预测时将net设置为评估模式
net.eval()
src_tokens = src_vocab[src_sentence.lower().split(' ')] + [
src_vocab['<eos>']]
enc_valid_len = torch.tensor([len(src_tokens)], device=device)
src_tokens = dataset.truncate_pad(src_tokens, num_steps, src_vocab['<pad>'])
# 添加批量轴
enc_X = torch.unsqueeze(
torch.tensor(src_tokens, dtype=torch.long, device=device), dim=0)
enc_outputs = net.encoder(enc_X, enc_valid_len)
dec_state = net.decoder.init_state(enc_outputs, enc_valid_len)
# 添加批量轴
dec_X = torch.unsqueeze(torch.tensor(
[tgt_vocab['<bos>']], dtype=torch.long, device=device), dim=0)
output_seq, attention_weight_seq = [], []
for _ in range(num_steps):
Y, dec_state = net.decoder(dec_X, dec_state)
# 我们使用具有预测最高可能性的词元，作为解码器在下一时间步的输入
dec_X = Y.argmax(dim=2)
pred = dec_X.squeeze(dim=0).type(torch.int32).item()
# 保存注意力权重（稍后讨论）
if save_attention_weights:
# 2'st block&2'st attention
attention_weight_seq.append(net.decoder.attention_weights[1][1].cpu())
# 一旦序列结束词元被预测，输出序列的生成就完成了
if pred == tgt_vocab['<eos>']:
break
output_seq.append(pred)
return ' '.join(tgt_vocab.to_tokens(output_seq)), attention_weight_seq
def bleu(pred_seq, label_seq, k): #@save
"""计算BLEU"""
pred_tokens, label_tokens = pred_seq.split(' '), [i for i in label_seq]
len_pred, len_label = len(pred_tokens), len(label_tokens)
score = math.exp(min(0, 1 - len_label / len_pred))
for n in range(1, k + 1):
num_matches, label_subs = 0, collections.defaultdict(int)
for i in range(len_label - n + 1):
label_subs[' '.join(label_tokens[i: i + n])] += 1
for i in range(len_pred - n + 1):
if label_subs[' '.join(pred_tokens[i: i + n])] > 0:
num_matches += 1
label_subs[' '.join(pred_tokens[i: i + n])] -= 1
score *= math.pow(num_matches / (len_pred - n + 1), math.pow(0.5, n))
return score
def try_gpu(i=0):
"""如果存在，则返回gpu(i)，否则返回cpu()
Defined in :numref:`sec_use_gpu`"""
if torch.cuda.device_count() >= i + 1:
return torch.device(f'cuda:{i}')
return torch.device('cpu')
from matplotlib import pyplot as plt
import matplotlib
# from matplotlib_inline import backend_inline
def show_heatmaps(matrices, xlabel, ylabel, titles=None, figsize=(2.5, 2.5),
cmap='Reds'):
"""
显示矩阵的热图（Heatmaps）。
这个函数旨在以子图网格的形式绘制多个矩阵，通常用于可视化注意力权重等。
参数:
matrices (numpy.ndarray 或 torch.Tensor 数组):
一个四维数组，形状应为 (num_rows, num_cols, height, width)。
其中，num_rows 和 num_cols 决定了子图网格的布局，
height 和 width 是每个热图（即每个矩阵）的维度。
xlabel (str):
所有最底行子图的 x 轴标签。
ylabel (str):
所有最左列子图的 y 轴标签。
titles (list of str, optional):
一个包含 num_cols 个标题的列表，用于设置每一列子图的标题。默认 None。
figsize (tuple, optional):
整个图形（figure）的大小。默认 (2.5, 2.5)。
cmap (str, optional):
用于绘制热图的颜色映射（colormap）。默认 'Reds'。
"""
# 导入所需的 matplotlib 模块，确保图形在 Jupyter/IPython 环境中正确显示为 SVG 格式
# （假设在包含这个函数的环境中已经导入了 matplotlib 的 backend_inline）
# backend_inline.set_matplotlib_formats('svg')
matplotlib.use('TkAgg')
# 从输入的 matrices 形状中解构出子图网格的行数和列数
# 假设 matrices 的形状是 (num_rows, num_cols, height, width)
num_rows, num_cols, _, _ = matrices.shape
# 创建一个包含多个子图（axes）的图形（fig）
# fig: 整个图形对象
# axes: 一个 num_rows x num_cols 的子图对象数组
fig, axes = plt.subplots(
num_rows, num_cols,
figsize=figsize,
sharex=True, # 所有子图共享 x 轴刻度
sharey=True, # 所有子图共享 y 轴刻度
squeeze=False # 即使只有一行或一列，也强制返回二维数组的 axes，方便后续循环
)
# 遍历子图的行和对应的矩阵行
# i 是行索引, row_axes 是当前行的子图数组, row_matrices 是当前行的矩阵数组
for i, (row_axes, row_matrices) in enumerate(zip(axes, matrices)):
# 遍历当前行中的子图和对应的矩阵
# j 是列索引, ax 是当前的子图对象, matrix 是当前的待绘矩阵
for j, (ax, matrix) in enumerate(zip(row_axes, row_matrices)):
# 使用 ax.imshow() 绘制热图
# matrix.detach().numpy()：将 PyTorch Tensor 转换为 numpy 数组，并从计算图中分离（如果它是 Tensor）
# cmap：指定颜色映射
pcm = ax.imshow(matrix.detach().numpy(), cmap=cmap)
# --- 设置轴标签和标题 ---
# 只有最底行 (i == num_rows - 1) 的子图才显示 x 轴标签
if i == num_rows - 1:
ax.set_xlabel(xlabel)
# 只有最左列 (j == 0) 的子图才显示 y 轴标签
if j == 0:
ax.set_ylabel(ylabel)
# 如果提供了标题列表，则设置当前列的子图标题（所有行共享列标题）
if titles:
ax.set_title(titles[j])
# --- 添加颜色条（Colorbar） ---
# 为整个图形添加一个颜色条，用于表示数值和颜色的对应关系
# pcm: 之前绘制的第一个热图返回的 Colormap
# ax=axes: 颜色条将参照整个子图网格进行定位和缩放
# shrink=0.6: 缩小颜色条的高度/长度，使其只占图形高度的 60%
fig.colorbar(pcm, ax=axes, shrink=0.6)
plt.show()
if __name__ == '__main__':
num_hiddens, num_blks, dropout = 256, 2, 0.2
ffn_num_hiddens, num_heads = 64, 4
batch_size = 1024
num_steps = 10
lr, num_epochs, device = 0.001, 2000, try_gpu()
train_iter, src_vocab, tgt_vocab, source, target = dataset.load_data(batch_size, num_steps)
encoder = TransformerEncoder(
len(src_vocab), num_hiddens, ffn_num_hiddens, num_heads,
num_blks, dropout)
decoder = TransformerDecoder(
len(tgt_vocab), num_hiddens, ffn_num_hiddens, num_heads,
num_blks, dropout)
net = EncoderDecoder(encoder, decoder)
is_train = False
is_show = True
if is_train:
train_seq2seq(net, train_iter, lr, num_epochs, tgt_vocab, device)
elif is_show:
state_dict = torch.load('model_h.pt')
net.load_state_dict(state_dict)
net.to(device)
src_text = "Call us."
translation, attention_weight_seq = predict_seq2seq(
net, src_text, src_vocab, tgt_vocab, num_steps, device, True)
# attention_weights = torch.eye(10).reshape((1, 1, 10, 10))
# (num_rows, num_cols, height, width)
print(f'translation={translation}')
# print(attention_weight_seq.shape)
stacked_tensor = torch.stack(attention_weight_seq, dim=0).permute(2, 1, 0, 3)
print(stacked_tensor.shape)
show_heatmaps(
stacked_tensor,
xlabel='Attention weight', ylabel='Decode Step', titles=['Head %d' % i for i in range(1, 5)])
else:
state_dict = torch.load('model_h.pt')
net.load_state_dict(state_dict)
net.to(device)
C = 0
C1 = 0
for i in range(2000):
# print(source[i])
# print(target[i])
translation, attention_weight_seq = predict_seq2seq(
net, source[i], src_vocab, tgt_vocab, num_steps, device)
score = bleu(translation, target[i], k=2)
if score > 0.0:
C = C + 1
if score > 0.8:
C1 = C1 + 1
print(f'{source[i]} => {translation}, bleu {score:.3f}')
print(f'Counter(bleu > 0) = {C}')
print(f'Valid-Counter(bleu > 0.8) = {C1}')我们先看一下TransformerEncoder做了什么：
- 和前面类似，首先输入做了embedding，然后叠加位置编码
- 然后循环计算每一个TransformerEncoderBlock
TransformerEncoderBlock中做了：
- 计算自注意力
- 残差连接和层归一化
- 位置前馈网络
- 残差连接和层归一化
然后我们来看看TransformerDecoder做了什么：
- 和TransformerEncoder类似，首先输入做了embedding，然后叠加位置编码
- 然后循环计算每一个TransformerDecoderBlock
- 最后接一个全连接，映射到词表大小
TransformerDecoderBlock做了：
- 首先准备自注意力的\(K_1 V_1\)，其更新过程是每次输入X的拼接过程
- 将输入X 作为Q，\(K_1 V_1\)作为KV开始自注意力的运算过程
- 残差连接和层归一化，得到Y
- 将enc_output作为KV, Y作为Q，计算编码器-解码器注意力
- 残差连接和层归一化
- 位置前馈网络
- 残差连接和层归一化
下面是训练和测试的一些结果
从上面的图可以看到，这个模型的效果比seq2seq原始模型、seq2seq带注意力的模型要好很多。
此外，下面是我们翻译："Call us."-> "联 系 我 们 。" 的attention weight的可视化（block=2, head=4, mask=3）
从每一个decode step的每个head的注意力权重来看，不同head关注了不一样的重点，有效的识别了特征中的多种属性，提高了模型的能力。
后记
本文介绍了transformer结构以及其示例，这里也引入了很多现在LLM的很多概念，例如：位置编码等。
参考文献
打赏、订阅、收藏、丢香蕉、硬币，请关注公众号（攻城狮的搬砖之路）
PS: 请尊重原创，不喜勿喷。
PS: 要转载请注明出处，本人版权所有。
PS: 有问题请留言，看到后我会第一时间回复。