当前位置:网站首页>Transformer中position encoding实践
Transformer中position encoding实践
2022-07-04 14:54:00 【初学者chris】
近年来,transformer由于其可以实现并行计算且可以解决长序列的依赖问题在nlp领域和cv领域大放异彩。
原理图如下所示:
这里我们主要关注一个小部分,即position encoding部分,因为transformer取消了循环依赖,为了体现位置属性,所以给每个元素进行位置编码。
代码如下所示,至于为什么会这么写,可以参考作者原文,或者参考一下文章。https://zhuanlan.zhihu.com/p/338592312
代码如下:
class PositionalEncoding(torch.nn.Module):
def __init__(self, d_model, max_len=5000):
super(PositionalEncoding, self).__init__()
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).transpose(0, 1)#(max-len,1,d_model)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:x.size(1), :].squeeze(1)
#x = x + self.pe[:x.size(1), :]
return x
为了测试,我们定义两个输入矩阵,分别为全0、全1tensor。
d_model = 4
a=torch.zeros(2,3,4)
pos=PositionalEncoding(d_model)
b=pos(a)
c=torch.ones(2,3,4)
b1=pos(c)
很明显,输入矩阵为
输出为b,b1,如下所示:;
可以看出,都是在输入的基础之上,加上了固定值,而那些固定值就是编码得到的,与输入无关,与d_model有关,d_model可以理解为单词的embedding大小。
边栏推荐
- Research Report on market supply and demand and strategy of China's four sided flat bag industry
- Research Report on plastic recycling machine industry - market status analysis and development prospect forecast
- What should ABAP do when it calls a third-party API and encounters garbled code?
- Stress, anxiety or depression? Correct diagnosis and retreatment
- Ten clothing stores have nine losses. A little change will make you buy every day
- Vscode prompt Please install clang or check configuration 'clang executable‘
- 线程池的使用和原理
- Find numbers
- [Chongqing Guangdong education] National Open University spring 2019 1248 public sector human resource management reference questions
- Lv166 turned over
猜你喜欢
Talking about Net core how to use efcore to inject multiple instances of a context annotation type for connecting to the master-slave database
Blood cases caused by Lombok use
[North Asia data recovery] data recovery case of database data loss caused by HP DL380 server RAID disk failure
MySQL learning notes - data type (2)
Application of clock wheel in RPC
Principle and general steps of SQL injection
Preliminary practice of niuke.com (10)
DC-2靶场搭建及渗透实战详细过程(DC靶场系列)
@EnableAspectAutoJAutoProxy_ Exposeproxy property
The new generation of domestic ORM framework sagacity sqltoy-5.1.25 release
随机推荐
多年锤炼,迈向Kata 3.0 !走进开箱即用的安全容器体验之旅| 龙蜥技术
Intranet penetrating FRP: hidden communication tunnel technology
Big God explains open source buff gain strategy live broadcast
[Chongqing Guangdong education] National Open University spring 2019 1396 pharmaceutical administration and regulations (version) reference questions
How can floating point numbers be compared with 0?
MFC implementation of ACM basic questions encoded by the number of characters
Preliminary practice of niuke.com (10)
Application of clock wheel in RPC
Common knowledge of unity Editor Extension
Variable cannot have type 'void'
Statistical learning: logistic regression and cross entropy loss (pytoch Implementation)
Accounting regulations and professional ethics [6]
Research Report on market supply and demand and strategy of China's well completion equipment industry
[book club issue 13] packaging format and coding format of audio files
Selenium browser (2)
AI system content recommendation issue 24
Anta is actually a technology company? These operations fool netizens
Will the memory of ParticleSystem be affected by maxparticles
How to save the contents of div as an image- How to save the contents of a div as a image?
2021 Google vulnerability reward program review