Pytorch在NLP中的簡單應(yīng)用詳解

import torch
import torch.autograd as autograd #Autograd為Tensor所有操作提供自動求導(dǎo)方法
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

#tensor
x = torch.Tensor([[1,2,3],[4,5,6]])
#size為2x3x4的隨機數(shù)隨機數(shù)
x = torch.randn((2,3,4))

x = torch.Tensor([[1,2],[3,4]])
y = torch.Tensor([[5,6],[7,8]])
z = x+y

x = torch.randn(2,3,4)
#拉直
x = x.view(-1)
#4*6維度
x = x.view(4,6)

#將Tensor變?yōu)閂ariable
x = autograd.Variable(torch.Tensor([1,2,3]),requires_grad = True)
#將Variable變?yōu)門ensor
y = x.data

x = autograd.Variable(torch.Tensor([1,2]),requires_grad=True)
y = autograd.Variable(torch.Tensor([3,4]),requires_grad=True)
z = x+y
#求和
s = z.sum()
#反向梯度傳播
s.backward()
print(x.grad)

linear = nn.Linear(3,5) #三維線性映射到五維
x = autograd.Variable(torch.randn(4,3))
#輸出為（4,5）維
y = linear(x)

x = autograd.Variable(torch.randn(5))
#relu激活函數(shù)
x_relu = F.relu(x)
print(x_relu)
x_soft = F.softmax(x)
#softmax激活函數(shù)
print(x_soft)
print(x_soft.sum())

Variable containing:
-0.9347
-0.9882
 1.3801
-0.1173
 0.9317
[torch.FloatTensor of size 5]
 
Variable containing:
 0.0481
 0.0456
 0.4867
 0.1089
 0.3108
[torch.FloatTensor of size 5]
 
Variable containing:
 1
[torch.FloatTensor of size 1]
 
Variable containing:
-3.0350
-3.0885
-0.7201
-2.2176
-1.1686
[torch.FloatTensor of size 5]

word_to_idx = {'hello':0,'world':1}
embeds = nn.Embedding(2,5) #即兩個單詞，單詞的詞嵌入維度為5
hello_idx = torch.LongTensor([word_to_idx['hello']])
hello_idx = autograd.Variable(hello_idx)
hello_embed = embeds(hello_idx)
print(hello_embed)

Variable containing:
-0.6982 0.3909 -1.0760 -1.6215 0.4429
[torch.FloatTensor of size 1x5]

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.autograd as autograd
import torch.optim as optim
 
from six.moves import xrange

context_size = 2
embed_dim = 10
text_sequence = """When forty winters shall besiege thy brow,
And dig deep trenches in thy beauty's field,
Thy youth's proud livery so gazed on now,
Will be a totter'd weed of small worth held:
Then being asked, where all thy beauty lies,
Where all the treasure of thy lusty days;
To say, within thine own deep sunken eyes,
Were an all-eating shame, and thriftless praise.
How much more praise deserv'd thy beauty's use,
If thou couldst answer 'This fair child of mine
Shall sum my count, and make my old excuse,'
Proving his beauty by succession thine!
This were to be new made when thou art old,
And see thy blood warm when thou feel'st it cold.""".split()
#分詞
trigrams = [ ([text_sequence[i], text_sequence[i+1]], text_sequence[i+2]) for i in xrange(len(text_sequence) - 2) ]
trigrams[:10]

#建立vocab索引
vocab = set(text_sequence)
word_to_ix = {word: i for i,word in enumerate(vocab)}

#N-Gram Language model
class NGramLanguageModeler(nn.Module): 
 def __init__(self, vocab_size, embed_dim, context_size):
  super(NGramLanguageModeler, self).__init__()
  #詞嵌入
  self.embedding = nn.Embedding(vocab_size, embed_dim)
  #兩層線性分類器
  self.linear1 = nn.Linear(embed_dim*context_size, 128)
  self.linear2 = nn.Linear(128, vocab_size)
  
 def forward(self, input):
  embeds = self.embedding(input).view((1, -1)) #2,10拉直為20
  out = F.relu(self.linear1(embeds))
  out = F.relu(self.linear2(out))
  log_probs = F.log_softmax(out)
  return log_probs  

#輸出模型看一下網(wǎng)絡(luò)結(jié)構(gòu)
model = NGramLanguageModeler(96,10,2)
print(model)

#定義損失函數(shù)以及優(yōu)化器
loss_function = nn.NLLLoss()
optimizer = optim.SGD(model.parameters(),lr = 0.01)
model = NGramLanguageModeler(len(vocab), embed_dim, context_size)
losses = []

#模型訓(xùn)練
for epoch in xrange(10):
 total_loss = torch.Tensor([0])
 for context, target in trigrams:
  #1.處理數(shù)據(jù)輸入為索引向量
  #print(context)
  #注：python3中map函數(shù)前要加上list()轉(zhuǎn)換為列表形式
  context_idxs = list(map(lambda w: word_to_ix[w], context))
  #print(context_idxs)
  context_var = autograd.Variable( torch.LongTensor(context_idxs) )
 
  
  #2.梯度清零
  model.zero_grad()
  
  #3.前向傳播，計算下一個單詞的概率
  log_probs = model(context_var)
  
  #4.損失函數(shù)
  loss = loss_function(log_probs, autograd.Variable(torch.LongTensor([word_to_ix[target]])))
  
  #反向傳播及梯度更新
  loss.backward()
  optimizer.step()
  
  total_loss += loss.data 
 losses.append(total_loss)
print(losses)