课程链接如下:
2.1认识Transformer架构-part1_哔哩哔哩_bilibili
因为网上可以找到源代码,但是呢,代码似乎有点小错误,我自己改正后,放到了GPU上运行,
代码如下:
# 来自https://www.bilibili.com/video/BV188411H71g?p=3&vd_source=3083729582baecf3ad2c3c52876b23aa
# 我已经使用GPU修改了代码,加了几处.cuda()就行了
import copy
import math
import time
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable
from pyitcast.transformer_utils import Batch
from pyitcast.transformer_utils import get_std_opt
from pyitcast.transformer_utils import LabelSmoothing
from pyitcast.transformer_utils import SimpleLossCompute
from pyitcast.transformer_utils import run_epoch
from pyitcast.transformer_utils import greedy_decode
# 文本嵌入层
class Embeddings(nn.Module):
def __init__(self, d_model, vocab):
super(Embeddings, self).__init__()
self.lut = nn.Embedding(vocab, d_model)
self.d_model = d_model
def forward(self, x):
return self.lut(x) * math.sqrt(self.d_model)
# 定义位置编码器,即也是一个层
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len).unsqueeze(1)
div_term = torch.exp(
torch.arange(0, d_model, 2) * (-(math.log(10000.0) / d_model)))
# 这意味着每个位置的频率随着位置的增加而减小。这使得模型能够学习序列中每个位置的重要性。
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:, :x.size(1)]
return self.dropout(x)
# 构建掩码张量
def subsequent_maxk(size):
attn_shape = (1, size, size)
subsequent_maxk = np.triu(np.ones(attn_shape), k=1).astype('uint8')
return torch.from_numpy(1 - subsequent_maxk)
# # 2.3.2注意力机制
# 为下面函数重写了注意力机制,否则代码会报错
# 注意力机制代码实现
def attention(query, key, value, mask=None, dropout=None):
d_k = query.size(-1)
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
if mask is not None:
# mask = torch.zeros(1, 1, 1, 1).cuda()
# print("mask.shape:", mask.shape)
scores = scores.masked_fill(mask == 0, -1e9)
p_attn = F.softmax(scores, dim=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return torch.matmul(p_attn, value), p_attn
# # 2.3.3多头注意力机制
# 实现克隆函数
def clones(module, N):
return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])
# 实现多头注意力机制
class MultiHeadAttention(nn.Module):
def __init__(self, head, embedding_dim, dropout=0.1):
super(MultiHeadAttention, self).__init__()
assert embedding_dim % head == 0
self.d_k = embedding_dim // head
self.head = head
self.embedding_dim = embedding_dim
self.linears = clones(nn.Linear(embedding_dim, embedding_dim), 4)
self.attn = None
self.dropout = nn.Dropout(p=dropout)
def forward(self, query, key, value, mask=None):
if mask is not None:
mask = mask.unsqueeze(0)
batch_size = query.size(0)
# 三个张量分别是三个输入,分别用三个线性层进行处理并重塑维度
query, key, value = \
[model(x).view(batch_size, -1, self.head, self.d_k).transpose(1, 2)
for model, x in zip(self.linears, (query, key, value))]
x, self.attn = attention(query, key, value, mask=mask,
dropout=self.dropout)
x = (x.transpose(1, 2).contiguous()
.view(batch_size, -1, self.head * self.d_k))
# 拷贝的四个层还有一个就是这个对输入进行线性变换得到输出
return self.linears[-1](x)
# # 2.3.4前馈全连接层
# 构建前馈全连接网络类
class PositionWiseFeedForward(nn.Module):
def __init__(self, d_model, d_ff, dropout=0.1):
super(PositionWiseFeedForward, self).__init__()
self.w1 = nn.Linear(d_model, d_ff)
self.w2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(p=dropout)
def forward(self, x):
return self.w2(self.dropout(F.dropout(F.relu(self.w1(x)))))
# # 2.3.5规范化层
class LayerNorm(nn.Module):
def __init__(self, features, eps=1e-6):
super(LayerNorm, self).__init__()
self.a2 = nn.Parameter(torch.ones(features))
self.b2 = nn.Parameter(torch.zeros(features))
self.eps = eps
def forward(self, x):
mean = x.mean(-1, keepdim=True)
std = x.std(-1, keepdim=True)
return self.a2 * (x - mean) / (std + self.eps) + self.b2
# # 2.3.6子层连接结构
# 构建子层连接结构
class SublayerConnection(nn.Module):
def __init__(self, size, dropout=0.1):
super(SublayerConnection, self).__init__()
self.norm = LayerNorm(size)
self.dropout = nn.Dropout(p=dropout)
self.size = size
def forward(self, x, sublayer):
return x + self.dropout(sublayer(self.norm(x)))
# # 2.3.7编码器层
# 编码器层
class EncoderLayer(nn.Module):
def __init__(self, size, self_attn, feed_forward, dropout):
super(EncoderLayer, self).__init__()
self.self_attn = self_attn
self.feed_forward = feed_forward
self.size = size
self.sublayer = clones(SublayerConnection(size, dropout), 2)
def forward(self, x, mask):
x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
return self.sublayer[1](x, self.feed_forward)
# # 2.3.8编码器
# 构建编码器类
class Encoder(nn.Module):
def __init__(self, layer, N):
super(Encoder, self).__init__()
self.layers = clones(layer, N)
self.norm = LayerNorm(layer.size)
def forward(self, x, mask):
for layer in self.layers:
x = layer(x, mask)
return self.norm(x)
# # 2.4解码器
# # 2.4.1解码器层
# 构建解码器层类
class DecoderLayer(nn.Module):
def __init__(self, size, self_attn, src_attn, feed_forward, dropout):
super(DecoderLayer, self).__init__()
self.size = size
self.self_attn = self_attn
self.src_attn = src_attn
self.feed_forward = feed_forward
self.dropout = dropout
self.sublayer = clones(SublayerConnection(size, dropout), 3)
def forward(self, x, memory, source_mask, target_mask):
m = memory
x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, target_mask))
x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, source_mask))
return self.sublayer[2](x, self.feed_forward)
# # 2.4.2 解码器
# 构建解码器类
class Decoder(nn.Module):
def __init__(self, layer, N):
super(Decoder, self).__init__()
self.layers = clones(layer, N)
self.norm = LayerNorm(layer.size)
def forward(self, x, memory, source_mask, target_mask):
for layer in self.layers:
x = layer(x, memory, source_mask, target_mask)
return self.norm(x)
# # 2.5输出部分实现
m = nn.Linear(20, 30)
input = torch.randn(128, 20)
output = m(input)
print(output.shape)
# 构建Generator类
import torch.nn.functional as F
class Generator(nn.Module):
def __init__(self, d_model, vocal_size):
super(Generator, self).__init__()
self.project = nn.Linear(d_model, vocal_size)
def forward(self, x):
return F.log_softmax(self.project(x), dim=1)
# # 2.6 Transformer模型构建
# 实现编码解码结构
class EncoderDecoder(nn.Module):
def __init__(self, encoder, decoder, source_embed, target_embed, generator):
super(EncoderDecoder, self).__init__()
self.encoder = encoder
self.decoder = decoder
self.src_embed = source_embed
self.tgt_embed = target_embed
self.generator = generator
def forward(self, source, target, source_mask, target_mask):
return self.decode(self.encode(source, source_mask), source_mask,
target, target_mask)
def encode(self, source, source_mask):
return self.encoder(self.src_embed(source), source_mask)
def decode(self, memory, source_mask, target, target_mask):
return self.decoder(self.tgt_embed(target), memory, source_mask,
target_mask)
# Transformer模型构建过程的代码分析
def make_model(source_vocab, target_vocab, N=6, d_model=512, d_ff=2048, head=8,
dropout=0.1):
c = copy.deepcopy
attn = MultiHeadAttention(head, d_model)
ff = PositionWiseFeedForward(d_model, d_ff, dropout)
position = PositionalEncoding(d_model, dropout)
model = EncoderDecoder(
Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),
Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),
nn.Sequential(Embeddings(d_model, source_vocab), c(position)),
nn.Sequential(Embeddings(d_model, target_vocab), c(position)),
Generator(d_model, target_vocab)
)
for p in model.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
return model
# # 2.7模型基本测试运行
# 构建数据集生成器
def data_generator(V, batch_size, num_batch):
for i in range(num_batch):
data = torch.from_numpy(
np.random.randint(1, V, size=(batch_size, 10)))
data[:, 0] = 1
# source = Variable(data, requires_grad=False).long()
# target = Variable(data, requires_grad=False).long()
source = Variable(data, requires_grad=False).long().cuda()
target = Variable(data, requires_grad=False).long().cuda()
yield Batch(source, target)
V = 11
batch_size = 20
num_batch = 30
# 获得Transformer模型及其优化器和损失函数
model = make_model(V, V, N=2)
# 将模型移动到GPU上
model.cuda()
model_optimizer = get_std_opt(model)
criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0)
loss = SimpleLossCompute(model.generator, criterion, model_optimizer)
# 运行模型进行训练和评估
def run(model, loss, epochs=10):
for epoch in range(epochs):
model.train()
run_epoch(data_generator(V, 8, 20), model, loss)
model.eval()
run_epoch(data_generator(V, 8, 5), model, loss)
start = time.time()
run(model, loss)
end = time.time()
# 总时间
total_time = end - start
print(f"Total time: {total_time:.3f}s")
# 使用模型进行贪婪解码
def run(model, loss, epochs=10):
for epoch in range(epochs):
model.train()
run_epoch(data_generator(V, 8, 20), model, loss)
model.eval()
run_epoch(data_generator(V, 8, 5), model, loss)
model.eval()
source = torch.LongTensor([[1, 3, 2, 5, 4, 6, 7, 8, 9, 10]]).cuda()
source_mask = torch.ones(1, 1, 10).cuda()
result = greedy_decode(model, source, source_mask, max_len=10,
start_symbol=1)
print(result)
start = time.time()
run(model, loss)
end = time.time()
# 总时间
total_time = end - start
print(f"Total time: {total_time:.3f}s")
然后来自这个人的源代码讲解也非常好,和视频一样,我也修改了可以放在GPU运行,代码如下
【精选】Pytorch:Transformer(Encoder编码器-Decoder解码器、多头注意力机制、多头自注意力机制、掩码张量、前馈全连接层、规范化层、子层连接结构、pyitcast) part1_あずにゃん的博客-CSDN博客
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import math
import matplotlib.pyplot as plt
import numpy as np
import copy
# embedding = nn.Embedding(10, 3)
# input1 = torch.LongTensor([[1, 2, 4, 5], [4, 3, 2, 9]])
# print(embedding(input1))
# embedding = nn.Embedding(10, 3, padding_idx=0)
# input1 = torch.LongTensor([[0, 2, 0, 5]])
# print(embedding(input1))
# 构建Embedding类来实现文本嵌入层
class Embeddings(nn.Module):
def __init__(self, d_model, vocab):
# d_model: 词嵌入的维度
# vocab: 词表的大小
super(Embeddings, self).__init__()
# 定义Embedding层
self.lut = nn.Embedding(vocab, d_model)
# 将参数传入类中
self.d_model = d_model
def forward(self, x):
# x: 代表输入进模型的文本通过词汇映射后的数字张量
return self.lut(x) * math.sqrt(self.d_model)
d_model = 512
vocab = 1000
# x = Variable(torch.LongTensor([[100, 2, 421, 508], [491, 998, 1, 221]]))
# emb = Embeddings(d_model, vocab)
# embr = emb(x)
# print("embr:", embr)
# print(embr.shape)
# m = nn.Dropout(p=0.2)
# input1 = torch.randn(4, 5)
# output = m(input1)
# print(output)
# x = torch.tensor([1, 2, 3, 4])
# y = torch.unsqueeze(x, 0)
# print(y)
# z = torch.unsqueeze(x, 1)
# print(z)
# 构建位置编码器的类
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout, max_len=5000):
# d_model: 代表词嵌入的维度
# dropout: 代表Dropout层的置零比率
# max_len: 代表每隔句子的最大长度
super(PositionalEncoding, self).__init__()
# 实例化Dropout层
self.dropout = nn.Dropout(p=dropout)
# 初始化一个位置编码矩阵, 大小是max_len * d_model
pe = torch.zeros(max_len, d_model)
# 初始化一个绝对位置矩阵, max_len * 1
position = torch.arange(0., max_len).unsqueeze(1)
# 定义一个变化矩阵div_term, 跳跃式的初始化
div_term = torch.exp(
torch.arange(0., d_model, 2) * -(math.log(10000.0) / d_model))
# 将前面定义的变化矩阵进行奇数, 偶数的分别赋值
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
# 将二维张量扩充成三维张量
pe = pe.unsqueeze(0)
# 将位置编码矩阵注册成模型的buffer, 这个buffer不是模型中的参数, 不跟随优化器同步更新
# 注册成buffer后我们就可以在模型保存后重新加载的时候, 将这个位置编码器和模型参数一同加载进来
self.register_buffer('pe', pe)
def forward(self, x):
# x: 代表文本序列的词嵌入表示
# 首先明确pe的编码太长了, 将第二个维度, 也就是max_len对应的那个维度缩小成x的句子长度同等的长度
x = x + Variable(self.pe[:, :x.size(1)], requires_grad=False)
return self.dropout(x)
d_model = 512
dropout = 0.1
max_len = 60
# x = embr
# pe = PositionalEncoding(d_model, dropout, max_len)
# pe_result = pe(x)
# print(pe_result)
# print(pe_result.shape)
# 第一步设置一个画布
# plt.figure(figsize=(15, 5))
# 实例化PositionalEncoding类对象, 词嵌入维度给20, 置零比率设置为0
# pe = PositionalEncoding(20, 0)
# 向pe中传入一个全零初始化的x, 相当于展示pe
# y = pe(Variable(torch.zeros(1, 100, 20)))
# plt.plot(np.arange(100), y[0, :, 4:8].data.numpy())
# plt.legend(["dim %d"%p for p in [4, 5, 6, 7]])
# print(np.triu([[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]], k=-1))
# print(np.triu([[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]], k=0))
# print(np.triu([[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]], k=1))
# 构建掩码张量的函数
def subsequent_mask(size):
# size: 代表掩码张量后两个维度, 形成一个方阵
attn_shape = (1, size, size)
# 使用np.ones()先构建一个全1的张量, 然后利用np.triu()形成上三角矩阵
subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8')
# 使得这个三角矩阵反转
return torch.from_numpy(1 - subsequent_mask)
size = 5
# sm = subsequent_mask(size)
# print("sm:", sm)
# plt.figure(figsize=(5, 5))
# plt.imshow(subsequent_mask(20)[0])
# x = Variable(torch.randn(5, 5))
# print(x)
# mask = Variable(torch.zeros(5, 5))
# print(mask)
# y = x.masked_fill(mask == 0, -1e9)
# print(y)
def attention(query, key, value, mask=None, dropout=None):
# query, key, value: 代表注意力的三个输入张量
# mask: 掩码张量
# dropout: 传入的Dropout实例化对象
# 首先将query的最后一个维度提取出来, 代表的是词嵌入的维度
d_k = query.size(-1)
# 按照注意力计算公式, 将query和key的转置进行矩阵乘法, 然后除以缩放稀疏
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
# 判断是否使用掩码张量
if mask is not None:
# 利用masked_fill方法, 将掩码张量和0进行位置的意义比较, 如果等于0, 替换成一个非常小的数值
scores = scores.masked_fill(mask == 0, -1e9)
# 对scores的最后一个维度上进行softmax操作
p_attn = F.softmax(scores, dim=-1)
# 判断是否使用dropout
if dropout is not None:
p_attn = dropout(p_attn)
# 最后一步完成p_attn和value张量的乘法, 并返回query注意力表示
return torch.matmul(p_attn, value), p_attn
# query = key = value = pe_result
# mask = Variable(torch.zeros(2, 4, 4))
# attn, p_attn = attention(query, key, value, mask=mask)
# print('attn:', attn)
# print(attn.shape)
# print('p_attn:', p_attn)
# print(p_attn.shape)
# x = torch.randn(4, 4)
# print(x.size())
# y = x.view(16)
# print(y.size())
# z = x.view(-1, 8)
# print(z.size())
# a = torch.randn(1, 2, 3, 4)
# print(a.size())
# print(a)
# b = a.transpose(1, 2)
# print(b.size())
# print(b)
# c = a.view(1, 3, 2, 4)
# print(c.size())
# print(c)
# 实现克隆函数, 因为在多头注意力机制下, 要用到多个结构相同的线性层
# 需要使用clone函数将他们一同初始化到一个网络层列表对象中
def clones(module, N):
# module: 代表要克隆的目标网络层
# N: 将module克隆几个
return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])
# 实现多头注意力机制的类
class MultiHeadedAttention(nn.Module):
def __init__(self, head, embedding_dim, dropout=0.1):
# head: 代表几个头的参数
# embedding_dim: 代表词嵌入的维度
# dropout: 进行Dropout操作时, 置零的比率
super(MultiHeadedAttention, self).__init__()
# 要确认一个事实: 多头的数量head需要整除词嵌入的维度embedding_dim
assert embedding_dim % head == 0
# 得到每个头获得的词向量的维度
self.d_k = embedding_dim // head
self.head = head
self.embedding_dim = embedding_dim
# 获得线性层, 要获得4个, 分别是Q,K,V以及最终的输出线性层
self.linears = clones(nn.Linear(embedding_dim, embedding_dim), 4)
# 初始化注意力张量
self.attn = None
# 初始化dropout对象
self.dropout = nn.Dropout(p=dropout)
def forward(self, query, key, value, mask=None):
# query, key, value是注意力机制的三个输入张量, mask代表掩码张量
# 首先判断是否使用掩码张量
if mask is not None:
# 使用unsqueeze将掩码张量进行维度扩充, 代表多头中的第n个头
mask = mask.unsqueeze(0)
# 得到batch_size
batch_size = query.size(0)
# 首先使用zip将网络层和输入数据连接在一起, 模型的输出利用view和transpose进行维度和形状的改变
query, key, value = \
[model(x).view(batch_size, -1, self.head, self.d_k).transpose(1, 2)
for model, x in zip(self.linears, (query, key, value))]
# 将每个头的输出传入到注意力层
x, self.attn = attention(query, key, value, mask=mask,
dropout=self.dropout)
# 得到每个头的计算结果是4维张量, 需要进行形状的转换
# 前面已经将1,2两个维度进行过转置, 在这里要重新转置回来
# 注意: 经历了transpose()方法后, 必须要使用contiguous方法, 不然无法使用view()方法
x = x.transpose(1, 2).contiguous().view(batch_size, -1,
self.head * self.d_k)
# 最后将x输入线性层列表中的最后一个线性层中进行处理, 得到最终的多头注意力结构输出
return self.linears[-1](x)
# 实例化若干参数
head = 8
embedding_dim = 512
dropout = 0.2
# 若干输入参数的初始化
# query = key = value = pe_result
# mask = Variable(torch.zeros(8, 4, 4))
# mha = MultiHeadedAttention(head, embedding_dim, dropout)
# mha_result = mha(query, key, value, mask)
# print(mha_result)
# print(mha_result.shape)
# 构建前馈全连接网络类
class PositionwiseFeedForward(nn.Module):
def __init__(self, d_model, d_ff, dropout=0.1):
# d_model: 代表词嵌入的维度, 同时也是两个线性层的输入维度和输出维度
# d_ff: 代表第一个线性层的输出维度, 和第二个线性层的输入维度
# dropout: 经过Dropout层处理时, 随机置零的比率
super(PositionwiseFeedForward, self).__init__()
# 定义两层全连接的线性层
self.w1 = nn.Linear(d_model, d_ff)
self.w2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(p=dropout)
def forward(self, x):
# x: 代表来自上一层的输出
# 首先将x送入第一个线性层网络, 然后经历relu函数的激活, 再经历dropout层的处理
# 最后送入第二个线性层
return self.w2(self.dropout(F.relu(self.w1(x))))
d_model = 512
d_ff = 64
dropout = 0.2
# x = mha_result
# ff = PositionwiseFeedForward(d_model, d_ff, dropout)
# ff_result = ff(x)
# print(ff_result)
# print(ff_result.shape)
# 构建规范化层的类
class LayerNorm(nn.Module):
def __init__(self, features, eps=1e-6):
# features: 代表词嵌入的维度
# eps: 一个足够小的正数, 用来在规范化计算公式的分母中, 防止除零操作
super(LayerNorm, self).__init__()
# 初始化两个参数张量a2, b2,用于对结果做规范化操作计算
# 将其用nn.Parameter进行封装, 代表他们也是模型中的参数
self.a2 = nn.Parameter(torch.ones(features))
self.b2 = nn.Parameter(torch.zeros(features))
self.eps = eps
def forward(self, x):
# x: 代表上一层网络的输出
# 首先对x进行最后一个维度上的求均值操作, 同时操持输出维度和输入维度一致
mean = x.mean(-1, keepdim=True)
# 接着对x进行字后一个维度上的求标准差的操作, 同时保持输出维度和输入维度一致
std = x.std(-1, keepdim=True)
# 按照规范化公式进行计算并返回
return self.a2 * (x - mean) / (std + self.eps) + self.b2
features = d_model = 512
eps = 1e-6
# x = ff_result
# ln = LayerNorm(features, eps)
# ln_result = ln(x)
# print(ln_result)
# print(ln_result.shape)
# 构建子层连接结构的类
class SublayerConnection(nn.Module):
def __init__(self, size, dropout=0.1):
# size: 代表词嵌入的维度
# dropout: 进行Dropout操作的置零比率
super(SublayerConnection, self).__init__()
# 实例化一个规范化层的对象
self.norm = LayerNorm(size)
# 实例化一个dropout对象
self.dropout = nn.Dropout(p=dropout)
self.size = size
def forward(self, x, sublayer):
# x: 代表上一层传入的张量
# sublayer: 该子层连接中子层函数
# 首先将x进行规范化, 然后送入子层函数中处理, 处理结果进入dropout层, 最后进行残差连接
return x + self.dropout(sublayer(self.norm(x)))
size = d_model = 512
head = 8
dropout = 0.2
# x = pe_result
# mask = Variable(torch.zeros(8, 4, 4))
# self_attn = MultiHeadedAttention(head, d_model)
# sublayer = lambda x: self_attn(x, x, x, mask)
# sc = SublayerConnection(size, dropout)
# sc_result = sc(x, sublayer)
# print(sc_result)
# print(sc_result.shape)
# 构建编码器层的类
class EncoderLayer(nn.Module):
def __init__(self, size, self_attn, feed_forward, dropout):
# size: 代表词嵌入的维度
# self_attn: 代表传入的多头自注意力子层的实例化对象
# feed_forward: 代表前馈全连接层实例化对象
# dropout: 进行dropout操作时的置零比率
super(EncoderLayer, self).__init__()
# 将两个实例化对象和参数传入类中
self.self_attn = self_attn
self.feed_forward = feed_forward
self.size = size
# 编码器层中有2个子层连接结构, 使用clones函数进行操作
self.sublayer = clones(SublayerConnection(size, dropout), 2)
def forward(self, x, mask):
# x: 代表上一层的传入张量
# mask: 代表掩码张量
# 首先让x经过第一个子层连接结构,内部包含多头自注意力机制子层
# 再让张量经过第二个子层连接结构, 其中包含前馈全连接网络
x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
return self.sublayer[1](x, self.feed_forward)
# size = d_model = 512
# head = 8
# d_ff = 64
# x = pe_result
# dropout = 0.2
# self_attn = MultiHeadedAttention(head, d_model)
# ff = PositionwiseFeedForward(d_model, d_ff, dropout)
# mask = Variable(torch.zeros(8, 4, 4))
# el = EncoderLayer(size, self_attn, ff, dropout)
# el_result = el(x, mask)
# print(el_result)
# print(el_result.shape)
# 构建编码器类Encoder
class Encoder(nn.Module):
def __init__(self, layer, N):
# layer: 代表编码器层
# N: 代表编码器中有几个layer
super(Encoder, self).__init__()
# 首先使用clones函数克隆N个编码器层放置在self.layers中
self.layers = clones(layer, N)
# 初始化一个规范化层, 作用在编码器的最后面
self.norm = LayerNorm(layer.size)
def forward(self, x, mask):
# x: 代表上一层输出的张量
# mask: 代表掩码张量
# 让x依次经历N个编码器层的处理, 最后再经过规范化层就可以输出了
for layer in self.layers:
x = layer(x, mask)
return self.norm(x)
# size = d_model = 512
# d_ff = 64
# head = 8
# c = copy.deepcopy
# attn = MultiHeadedAttention(head, d_model)
# ff = PositionwiseFeedForward(d_model, d_ff, dropout)
# dropout = 0.2
# layer = EncoderLayer(size, c(attn), c(ff), dropout)
# N = 8
# mask = Variable(torch.zeros(8, 4, 4))
# en = Encoder(layer, N)
# en_result = en(x, mask)
# print(en_result)
# print(en_result.shape)
# 构建解码器层类
class DecoderLayer(nn.Module):
def __init__(self, size, self_attn, src_attn, feed_forward, dropout):
# size: 代表词嵌入的维度
# self_attn: 代表多头自注意力机制的对象
# src_attn: 代表常规的注意力机制的对象
# feed_forward: 代表前馈全连接层的对象
# dropout: 代表Dropout的置零比率
super(DecoderLayer, self).__init__()
# 将参数传入类中
self.size = size
self.self_attn = self_attn
self.src_attn = src_attn
self.feed_forward = feed_forward
self.dropout = dropout
# 按照解码器层的结构图, 使用clones函数克隆3个子层连接对象
self.sublayer = clones(SublayerConnection(size, dropout), 3)
def forward(self, x, memory, source_mask, target_mask):
# x: 代表上一层输入的张量
# memory: 代表编码器的语义存储张量
# source_mask: 源数据的掩码张量
# target_mask: 目标数据的掩码张量
m = memory
# 第一步让x经历第一个子层, 多头自注意力机制的子层
# 采用target_mask, 为了将解码时未来的信息进行遮掩, 比如模型解码第二个字符, 只能看见第一个字符信息
x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, target_mask))
# 第二步让x经历第二个子层, 常规的注意力机制的子层, Q!=K=V
# 采用source_mask, 为了遮掩掉对结果信息无用的数据
x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, source_mask))
# 第三步让x经历第三个子层, 前馈全连接层
return self.sublayer[2](x, self.feed_forward)
# size = d_model = 512
# head = 8
# d_ff = 64
# dropout = 0.2
# self_attn = src_attn = MultiHeadedAttention(head, d_model, dropout)
# ff = PositionwiseFeedForward(d_model, d_ff, dropout)
# x = pe_result
# memory = en_result
# mask = Variable(torch.zeros(8, 4, 4))
# source_mask = target_mask = mask
# dl = DecoderLayer(size, self_attn, src_attn, ff, dropout)
# dl_result = dl(x, memory, source_mask, target_mask)
# print(dl_result)
# print(dl_result.shape)
# 构建解码器类
class Decoder(nn.Module):
def __init__(self, layer, N):
# layer: 代表解码器层的对象
# N: 代表将layer进行几层的拷贝
super(Decoder, self).__init__()
# 利用clones函数克隆N个layer
self.layers = clones(layer, N)
# 实例化一个规范化层
self.norm = LayerNorm(layer.size)
def forward(self, x, memory, source_mask, target_mask):
# x: 代表目标数据的嵌入表示,
# memory: 代表编码器的输出张量
# source_mask: 源数据的掩码张量
# target_mask: 目标数据的掩码张量
# 要将x依次经历所有的编码器层处理, 最后通过规范化层
for layer in self.layers:
x = layer(x, memory, source_mask, target_mask)
return self.norm(x)
# size = d_model = 512
# head = 8
# d_ff = 64
# dropout = 0.2
# c = copy.deepcopy
# attn = MultiHeadedAttention(head, d_model)
# ff = PositionwiseFeedForward(d_model, d_ff, dropout)
# layer = DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout)
# N = 8
# x = pe_result
# memory = en_result
# mask = Variable(torch.zeros(8, 4, 4))
# source_mask = target_mask = mask
# de = Decoder(layer, N)
# de_result = de(x, memory, source_mask, target_mask)
# print(de_result)
# print(de_result.shape)
# 构建Generator类
import torch.nn.functional as F
class Generator(nn.Module):
def __init__(self, d_model, vocab_size):
# d_model: 代表词嵌入的维度
# vocab_size: 代表词表的总大小
super(Generator, self).__init__()
# 定义一个线性层, 作用是完成网络输出维度的变换
self.project = nn.Linear(d_model, vocab_size)
def forward(self, x):
# x: 代表上一层的输出张量
# 首先将x送入线性层中, 让其经历softmax的处理
return F.log_softmax(self.project(x), dim=-1)
# d_model = 512
# vocab_size = 1000
# x = de_result
# gen = Generator(d_model, vocab_size)
# gen_result = gen(x)
# print(gen_result)
# print(gen_result.shape)
# 构建编码器-解码器结构类
class EncoderDecoder(nn.Module):
def __init__(self, encoder, decoder, source_embed, target_embed, generator):
# encoder: 代表编码器对象
# decoder: 代表解码器对象
# source_embed: 代表源数据的嵌入函数
# target_embed: 代表目标数据的嵌入函数
# generator: 代表输出部分类别生成器对象
super(EncoderDecoder, self).__init__()
self.encoder = encoder
self.decoder = decoder
self.src_embed = source_embed
self.tgt_embed = target_embed
self.generator = generator
def forward(self, source, target, source_mask, target_mask):
# source: 代表源数据
# target: 代表目标数据
# source_mask: 代表源数据的掩码张量
# target_mask: 代表目标数据的掩码张量
return self.decode(self.encode(source, source_mask), source_mask,
target, target_mask)
def encode(self, source, source_mask):
return self.encoder(self.src_embed(source), source_mask)
def decode(self, memory, source_mask, target, target_mask):
# memory: 代表经历编码器编码后的输出张量
return self.decoder(self.tgt_embed(target), memory, source_mask,
target_mask)
# vocab_size = 1000
# d_model = 512
# encoder = en
# decoder = de
# source_embed = nn.Embedding(vocab_size, d_model)
# target_embed = nn.Embedding(vocab_size, d_model)
# generator = gen
#
# source = target = Variable(torch.LongTensor([[100, 2, 421, 508], [491, 998, 1, 221]]))
#
# source_mask = target_mask = Variable(torch.zeros(8, 4, 4))
#
# ed = EncoderDecoder(encoder, decoder, source_embed, target_embed, generator)
# ed_result = ed(source, target, source_mask, target_mask)
# print(ed_result)
# print(ed_result.shape)
def make_model(source_vocab, target_vocab, N=6, d_model=512, d_ff=2048, head=8,
dropout=0.1):
# source_vocab: 代表源数据的词汇总数
# target_vocab: 代表目标数据的词汇总数
# N: 代表编码器和解码器堆叠的层数
# d_model: 代表词嵌入的维度
# d_ff: 代表前馈全连接层中变换矩阵的维度
# head: 多头注意力机制中的头数
# dropout: 指置零的比率
c = copy.deepcopy
# 实例化一个多头注意力的类
attn = MultiHeadedAttention(head, d_model)
# 实例化一个前馈全连接层的网络对象
ff = PositionwiseFeedForward(d_model, d_ff, dropout)
# 实例化一个位置编码器
position = PositionalEncoding(d_model, dropout)
# 实例化模型model,利用的是EncoderDecoder类
# 编码器的结构里面有2个子层, attention层和前馈全连接层
# 解码器的结构中有3个子层, 两个attention层和前馈全连接层
model = EncoderDecoder(
Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),
Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),
nn.Sequential(Embeddings(d_model, source_vocab), c(position)),
nn.Sequential(Embeddings(d_model, target_vocab), c(position)),
Generator(d_model, target_vocab))
# 初始化整个模型中的参数, 判断参数的维度大于1, 将矩阵初始化成一个服从均匀分布的矩阵
for p in model.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
return model
source_vocab = 11
target_vocab = 11
N = 6
# if __name__ == '__main__':
# res = make_model(source_vocab, target_vocab, N)
# print(res)
# ------------------------------------------------------
from pyitcast.transformer_utils import Batch
from pyitcast.transformer_utils import get_std_opt
from pyitcast.transformer_utils import LabelSmoothing
from pyitcast.transformer_utils import SimpleLossCompute
from pyitcast.transformer_utils import run_epoch
from pyitcast.transformer_utils import greedy_decode
def data_generator(V, batch_size, num_batch):
# V: 随机生成数据的最大值+1
# batch_size: 每次输送给模型的样本数量, 经历这些样本训练后进行一次参数的更新
# num_batch: 一共输送模型多少轮数据
for i in range(num_batch):
# 使用numpy中的random.randint()来随机生成[1, V)
# 分布的形状(batch, 10)
data = torch.from_numpy(np.random.randint(1, V, size=(batch_size, 10)))
# 将数据的第一列全部设置为1, 作为起始标志
data[:, 0] = 1
# 因为是copy任务, 所以源数据和目标数据完全一致
# 设置参数requires_grad=False, 样本的参数不需要参与梯度的计算
source = Variable(data, requires_grad=False).long().cuda()
target = Variable(data, requires_grad=False).long().cuda()
yield Batch(source, target)
V = 11
batch_size = 20
num_batch = 30
# if __name__ == '__main__':
# res = data_generator(V, batch_size, num_batch)
# print(res)
# 使用make_model()函数获得模型的实例化对象
model = make_model(V, V, N=2)
model.cuda()
# 使用工具包get_std_opt获得模型的优化器
model_optimizer = get_std_opt(model)
# 使用工具包LabelSmoothing获得标签平滑对象
criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0)
# 使用工具包SimpleLossCompute获得利用标签平滑的结果得到的损失计算方法
loss = SimpleLossCompute(model.generator, criterion, model_optimizer)
# crit = LabelSmoothing(size=5, padding_idx=0, smoothing=0.5)
# predict = Variable(torch.FloatTensor([[0, 0.2, 0.7, 0.1, 0],
# [0, 0.2, 0.7, 0.1, 0],
# [0, 0.2, 0.7, 0.1, 0]]))
# target = Variable(torch.LongTensor([2, 1, 0]))
# crit(predict, target)
# plt.imshow(crit.true_dist)
# def run(model, loss, epochs=10):
# model: 代表将要训练的模型
# loss: 代表使用的损失计算方法
# epochs: 代表模型训练的轮次数
# for epoch in range(epochs):
# 首先进入训练模式, 所有的参数将会被更新
# model.train()
# 训练时, 传入的batch_size是20
# run_epoch(data_generator(V, 8, 20), model, loss)
# 训练结束后, 进入评估模式, 所有的参数固定不变
# model.eval()
# 评估时, 传入的batch_size是5
# run_epoch(data_generator(V, 8, 5), model, loss)
# if __name__ == '__main__':
# run(model, loss)
def run(model, loss, epochs=10):
for epoch in range(epochs):
# 首先进入训练模式, 所有的参数将会被更新
model.train()
run_epoch(data_generator(V, 8, 20), model, loss)
# 训练结束后, 进入评估模式, 所有的参数固定不变
model.eval()
run_epoch(data_generator(V, 8, 5), model, loss)
# 跳出for循环后, 代表模型训练结束, 进入评估模式
model.eval()
# run_epoch(data_generator(V, 8, 5), model, loss)
# 初始化一个输入张量
source = torch.LongTensor([[1, 3, 2, 5, 4, 6, 7, 8, 9, 10]]).cuda()
# 初始化一个输入张量的掩码张量, 全1代表没有任何的遮掩
source_mask = torch.ones(1, 1, 10).cuda()
# 设定解码的最大长度max_len等于10, 起始数字的标志默认等于1
result = greedy_decode(model, source, source_mask, max_len=10,
start_symbol=1)
print(result)
import time
if __name__ == '__main__':
start = time.time()
run(model, loss)
end = time.time()
# 总时间
total_time = end - start
print(f"Total time: {total_time:.3f}s")