最近在学Transformer,学了理论的部分之后就开始学代码的实现,这里是跟着b站的up主的视频记的笔记,视频链接:19、Transformer模型Encoder原理精讲及其PyTorch逐行实现_哔哩哔哩_bilibili
首先导入所需要的包:
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
关于Word Embedding,这里以序列建模为例,考虑source sentence、target sentence,构建序列,序列的字符以其在词表中的索引的形式表示。
首先使用定义batch_size的大小,并且使用torch.randint()函数随机生成序列长度,这里的src是生成原本的序列,tgt是生成目标的序列。
以机器翻译实现英文翻译为中文来说,src就是英文句子,tgt就是中文句子,这也就是规定了要翻译的英文句子的长度和翻译出来的句子长度。(举个例子而已,不用纠结为什么翻译要限制句子的长度)
batch_size = 2
src_len=torch.randint(2,5,(batch_size,))
tgt_len=torch.randint(2,5,(batch_size,))
将生成的src_len、tgt_len输出:
tensor([2, 3]) 生成的原序列第一个句子长度为2,第二个句子长度为3 tensor([4, 4]) 生成的目标序列第一个句子长度为4,第二个句子长度为4
因为随机生成的,所以每次运行都会有新的结果,也就是生成的src和tgt两个序列,其子句的长度每次都是随机的,这里改成生成固定长度的序列:
src_len = torch.Tensor([11, 9]).to(torch.int32)
tgt_len = torch.Tensor([10, 11]).to(torch.int32)
将生成的src_len、tgt_len输出,此时就固定好了序列长度了:
tensor([11, 9], dtype=torch.int32) tensor([10, 11], dtype=torch.int32)
接着是要实现单词索引构成的句子,首先定义单词表的大小和序列的最大长度。
# 单词表大小
max_num_src_words = 10
max_num_tgt_words = 10
# 序列的最大长度
max_src_seg_len = 12
max_tgt_seg_len = 12
以生成原序列为例,使用torch.randint()生成第一个句子和第二个句子,然后放到列表中:
src_seq = [torch.randint(1, max_num_src_words, (L,)) for L in src_len]
[tensor([5, 3, 7, 5, 6, 3, 4, 3]), tensor([1, 6, 3, 1, 1, 7, 4])]
可以发现生成的两个序列长度不一样(因为我们自己定义的时候就是不一样的),在这里需要使用F.pad()函数进行padding保证序列长度一致:
src_seq = [F.pad(torch.randint(1, max_num_src_words, (L,)), (0, max_src_seg_len-L)) for L in src_len]
[tensor([8, 5, 2, 4, 6, 8, 1, 4, 0, 0, 0, 0]), tensor([5, 5, 5, 3, 7, 9, 3, 0, 0, 0, 0, 0])]
此时已经填充为同样的长度了,但是不同的句子各为一个张量,需要使用torch.cat()函数把不同句子的tensor转化为二维的tensor,在此之前需要先把每个张量变成二维的,使用torch.unsqueeze()函数:
src_seq = torch.cat([torch.unsqueeze(F.pad(torch.randint(1, max_num_src_words, (L,)),(0, max_src_seg_len-L)), 0) for L in src_len])
tensor([[9, 7, 7, 4, 7, 3, 9, 4, 7, 8, 8, 0], [1, 1, 5, 9, 5, 6, 2, 7, 4, 0, 0, 0]]) tensor([[3, 3, 2, 8, 3, 4, 1, 2, 9, 4, 0, 0], [1, 6, 3, 8, 5, 1, 5, 5, 1, 5, 3, 0]])
这里把tgt的也补充了,得到的就是src和tgt的内容各自在一个二维张量里(batch_size,max_seg_len),batch_size也就是句子数,max_seg_len也就是句子的单词数(分为src的长度跟tgt两种)。
补充:可以看到上面三次运行出来的结果都不一样,因为三次运行的时候,每次都是随机生成,所以结果肯定不一样,第三次为什么有两个二维的tensor是因为第三次把tgt的部分也补上去了,所以就有两个二维的tensor。
接下来就是构造embedding了,这里nn.Embedding()传入了两个参数,第一个是embedding的长度,也就是单词个数+1,+1的原因是因为有个0是作为填充的,第二个参数就是embedding的维度,也就是一个单词会被映射为多少维度的向量。
然后调用forward,得到我们的src和tgt的embedding
src_embedding_table = nn.Embedding(max_num_src_words+1, model_dim)
tgt_embedding_table = nn.Embedding(max_num_tgt_words+1, model_dim)
src_embedding = src_embedding_table(src_seq)
tgt_embedding = tgt_embedding_table(tgt_seq)
print(src_embedding_table.weight) # 每一行代表一个embedding向量,第0行让给pad,从第1行到第行分配给各个单词,单词的索引是多少就取对应的行位置的向量
print(src_embedding) # 根据src_seq,从src_embedding_table获取得到的embedding vector,三维张量:batch_size、max_seq_len、model_dim
print(tgt_embedding)
此时src_embedding_table.weight的输出内容如下,第一行为填充(0)的向量:
tensor([[-0.3412, 1.5198, -1.7252, 0.6905, -0.3832, -0.8586, -2.0788, 0.3269],
[-0.5613, 0.3953, 1.6818, -2.0385, 1.1072, 0.2145, -0.9349, -0.7091],
[ 1.5881, -0.2389, -0.0347, 0.3808, 0.5261, 0.7253, 0.8557, -1.0020],
[-0.2725, 1.3238, -0.4087, 1.0758, 0.5321, -0.3466, -0.9051, -0.8938],
[-1.5393, 0.4966, -1.4887, 0.2795, -1.6751, -0.8635, -0.4689, -0.0827],
[ 0.6798, 0.1168, -0.5410, 0.5363, -0.0503, 0.4518, -0.3134, -0.6160],
[-1.1223, 0.3817, -0.6903, 0.0479, -0.6894, 0.7666, 0.9695, -1.0962],
[ 0.9608, 0.0764, 0.0914, 1.1949, -1.3853, 1.1089, -0.9282, -0.9793],
[-0.9118, -1.4221, -2.4675, -0.1321, 0.7458, -0.8015, 0.5114, -0.5023],
[-1.7504, 0.0824, 2.2088, -0.4486, 0.7324, 1.8790, 1.7644, 1.2731],
[-0.3791, 1.9915, -1.0117, 0.8238, -2.1784, -1.2824, -0.4275, 0.3202]],
requires_grad=True)
src_embedding的输出结果如下所示,往前看src_seq的第一个句子前三个为9 7 7,往前看第9+1行与第7+1行的向量,就是现在输出的前3个向量:
tensor([[[-1.7504, 0.0824, 2.2088, -0.4486, 0.7324, 1.8790, 1.7644, 1.2731], [ 0.9608, 0.0764, 0.0914, 1.1949, -1.3853, 1.1089, -0.9282, -0.9793], [ 0.9608, 0.0764, 0.0914, 1.1949, -1.3853, 1.1089, -0.9282, -0.9793], [-1.5393, 0.4966, -1.4887, 0.2795, -1.6751, -0.8635, -0.4689, -0.0827], [ 0.9608, 0.0764, 0.0914, 1.1949, -1.3853, 1.1089, -0.9282, -0.9793], [-0.2725, 1.3238, -0.4087, 1.0758, 0.5321, -0.3466, -0.9051, -0.8938], [-1.7504, 0.0824, 2.2088, -0.4486, 0.7324, 1.8790, 1.7644, 1.2731], [-1.5393, 0.4966, -1.4887, 0.2795, -1.6751, -0.8635, -0.4689, -0.0827], [ 0.9608, 0.0764, 0.0914, 1.1949, -1.3853, 1.1089, -0.9282, -0.9793], [-0.9118, -1.4221, -2.4675, -0.1321, 0.7458, -0.8015, 0.5114, -0.5023], [-0.9118, -1.4221, -2.4675, -0.1321, 0.7458, -0.8015, 0.5114, -0.5023], [-0.3412, 1.5198, -1.7252, 0.6905, -0.3832, -0.8586, -2.0788, 0.3269]], [[-0.5613, 0.3953, 1.6818, -2.0385, 1.1072, 0.2145, -0.9349, -0.7091], [-0.5613, 0.3953, 1.6818, -2.0385, 1.1072, 0.2145, -0.9349, -0.7091], [ 0.6798, 0.1168, -0.5410, 0.5363, -0.0503, 0.4518, -0.3134, -0.6160], [-1.7504, 0.0824, 2.2088, -0.4486, 0.7324, 1.8790, 1.7644, 1.2731], [ 0.6798, 0.1168, -0.5410, 0.5363, -0.0503, 0.4518, -0.3134, -0.6160], [-1.1223, 0.3817, -0.6903, 0.0479, -0.6894, 0.7666, 0.9695, -1.0962], [ 1.5881, -0.2389, -0.0347, 0.3808, 0.5261, 0.7253, 0.8557, -1.0020], [ 0.9608, 0.0764, 0.0914, 1.1949, -1.3853, 1.1089, -0.9282, -0.9793], [-1.5393, 0.4966, -1.4887, 0.2795, -1.6751, -0.8635, -0.4689, -0.0827], [-0.3412, 1.5198, -1.7252, 0.6905, -0.3832, -0.8586, -2.0788, 0.3269], [-0.3412, 1.5198, -1.7252, 0.6905, -0.3832, -0.8586, -2.0788, 0.3269], [-0.3412, 1.5198, -1.7252, 0.6905, -0.3832, -0.8586, -2.0788, 0.3269]]], grad_fn=)
同理tgt_embedding的输出结果如下所示:
tensor([[[-1.3681, -0.1619, -0.3676, 0.4312, -1.3842, -0.6180, 0.3685, 1.6281], [-1.3681, -0.1619, -0.3676, 0.4312, -1.3842, -0.6180, 0.3685, 1.6281], [-2.6519, -0.8566, 1.2268, 2.6479, -0.2011, -0.1394, -0.2449, 1.0309], [-0.8919, 0.5235, -3.1833, 0.9388, -0.6213, -0.5146, 0.7913, 0.5126], [-1.3681, -0.1619, -0.3676, 0.4312, -1.3842, -0.6180, 0.3685, 1.6281], [-0.4984, 0.2948, -0.2804, -1.1943, -0.4495, 0.3793, -0.1562, -1.0122], [ 0.8976, 0.5226, 0.0286, 0.1434, -0.2600, -0.7661, 0.1225, -0.7869], [-2.6519, -0.8566, 1.2268, 2.6479, -0.2011, -0.1394, -0.2449, 1.0309], [ 2.2026, 1.8504, -0.6285, -0.0996, -0.0994, -0.0828, 0.6004, -0.3173], [-0.4984, 0.2948, -0.2804, -1.1943, -0.4495, 0.3793, -0.1562, -1.0122], [ 0.3637, 0.4256, 0.7674, 1.4321, -0.1164, -0.6032, -0.8182, -0.6119], [ 0.3637, 0.4256, 0.7674, 1.4321, -0.1164, -0.6032, -0.8182, -0.6119]], [[ 0.8976, 0.5226, 0.0286, 0.1434, -0.2600, -0.7661, 0.1225, -0.7869], [-1.0356, 0.8212, 1.0538, 0.4510, 0.2734, 0.3254, 0.4503, 0.1694], [-1.3681, -0.1619, -0.3676, 0.4312, -1.3842, -0.6180, 0.3685, 1.6281], [-0.8919, 0.5235, -3.1833, 0.9388, -0.6213, -0.5146, 0.7913, 0.5126], [-0.4783, -1.5936, 0.5033, 0.3483, -1.3354, 1.4553, -1.1344, -1.9280], [ 0.8976, 0.5226, 0.0286, 0.1434, -0.2600, -0.7661, 0.1225, -0.7869], [-0.4783, -1.5936, 0.5033, 0.3483, -1.3354, 1.4553, -1.1344, -1.9280], [-0.4783, -1.5936, 0.5033, 0.3483, -1.3354, 1.4553, -1.1344, -1.9280], [ 0.8976, 0.5226, 0.0286, 0.1434, -0.2600, -0.7661, 0.1225, -0.7869], [-0.4783, -1.5936, 0.5033, 0.3483, -1.3354, 1.4553, -1.1344, -1.9280], [-1.3681, -0.1619, -0.3676, 0.4312, -1.3842, -0.6180, 0.3685, 1.6281], [ 0.3637, 0.4256, 0.7674, 1.4321, -0.1164, -0.6032, -0.8182, -0.6119]]], grad_fn=)
实际想要把文本句子嵌入到Embedding中,需要先根据自己的词典,将文本信息转化为每个词在词典中的位置,然后第0个位置依旧要让给Padding,得到索引然后构建Batch再去构造Embedding,以索引为输入得到每个样本的Embedding。
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
# 句子数
batch_size = 2
# 单词表大小
max_num_src_words = 10
max_num_tgt_words = 10
# 序列的最大长度
max_src_seg_len = 12
max_tgt_seg_len = 12
# 模型的维度
model_dim = 8
# 生成固定长度的序列
src_len = torch.Tensor([11, 9]).to(torch.int32)
tgt_len = torch.Tensor([10, 11]).to(torch.int32)
print(src_len)
print(tgt_len)
#单词索引构成的句子
src_seq = torch.cat([torch.unsqueeze(F.pad(torch.randint(1, max_num_src_words, (L,)),(0, max_src_seg_len-L)), 0) for L in src_len])
tgt_seq = torch.cat([torch.unsqueeze(F.pad(torch.randint(1, max_num_tgt_words, (L,)),(0, max_tgt_seg_len-L)), 0) for L in tgt_len])
print(src_seq)
print(tgt_seq)
# 构造embedding
src_embedding_table = nn.Embedding(max_num_src_words+1, model_dim)
tgt_embedding_table = nn.Embedding(max_num_tgt_words+1, model_dim)
src_embedding = src_embedding_table(src_seq)
tgt_embedding = tgt_embedding_table(tgt_seq)
print(src_embedding_table.weight)
print(src_embedding)
print(tgt_embedding)
torch.randint — PyTorch 2.3 documentation
torch.nn.functional.pad — PyTorch 2.3 文档
F.pad 的理解_domain:luyixian.cn-CSDN博客
嵌入 — PyTorch 2.3 文档