pytorch ---神经网络开篇之作 NNLM <A Neural Probabilistic Language Model>

论文地址：http://www.iro.umontreal.ca/~vincentp/Publications/lm_jmlr.pdf

论文给出了NNLM的框架图：

针对论文，实现代码如下：

 1 # -*- coding: utf-8 -*-
 2 # @time : 2019/10/26  12:20
 3
 4 import numpy as np
 5 import torch
 6 import torch.nn as nn
 7 import torch.optim as optim
 8 from torch.autograd import Variable
 9
10 dtype = torch.FloatTensor
11
12 sentences = [ "i like dog", "i love coffee", "i hate milk"]
13
14 word_list = " ".join(sentences).split()
15 word_list = list(set(word_list))
16 word_dict = {w: i for i, w in enumerate(word_list)} # {‘i‘: 0, ‘like‘: 1, ‘love‘: 2, ‘hate‘: 3, ‘milk‘: 4, ‘dog‘: 5, ‘coffee‘: 6}}
17 number_dict = {i: w for i, w in enumerate(word_list)}
18 n_class = len(word_dict) # number of Vocabulary
19
20 # NNLM Parameter
21 n_step = 2 # n-1 in paper    ->3gram
22 n_hidden = 2 # h in paper   ->number hidden unit
23 m = 2 # m in paper   ->embedding size
24
25 # make data batch (input,target)
26 # input: [[0,1],[0,2],[0,3]]
27 # target: [5,6,4]
28 def make_batch(sentences):
29     input_batch = []
30     target_batch = []
31
32     for sen in sentences:
33         word = sen.split()
34         input = [word_dict[n] for n in word[:-1]]
35         target = word_dict[word[-1]]
36
37         input_batch.append(input)
38         target_batch.append(target)
39
40     return input_batch, target_batch
41
42 # Model
43 class NNLM(nn.Module):
44     def __init__(self):
45         super(NNLM, self).__init__()
46         self.C = nn.Embedding(n_class, m)
47         self.H = nn.Parameter(torch.randn(n_step * m, n_hidden).type(dtype))
48         self.W = nn.Parameter(torch.randn(n_step * m, n_class).type(dtype))
49         self.d = nn.Parameter(torch.randn(n_hidden).type(dtype))
50         self.U = nn.Parameter(torch.randn(n_hidden, n_class).type(dtype))
51         self.b = nn.Parameter(torch.randn(n_class).type(dtype))
52
53     def forward(self, X):
54         X = self.C(X)
55         X = X.view(-1, n_step * m) # [batch_size, n_step * m]
56         tanh = torch.tanh(self.d + torch.mm(X, self.H)) # [batch_size, n_hidden]
57         output = self.b + torch.mm(X, self.W) + torch.mm(tanh, self.U) # [batch_size, n_class]
58         return output
59
60 model = NNLM()
61
62 criterion = nn.CrossEntropyLoss()
63 optimizer = optim.Adam(model.parameters(), lr=0.001)
64
65 input_batch, target_batch = make_batch(sentences)
66 input_batch = Variable(torch.LongTensor(input_batch))
67 target_batch = Variable(torch.LongTensor(target_batch))
68
69 # Training
70 for epoch in range(5000):
71
72     optimizer.zero_grad()
73     output = model(input_batch)
74
75     # output : [batch_size, n_class], target_batch : [batch_size] (LongTensor, not one-hot)
76     loss = criterion(output, target_batch)
77     if (epoch + 1)%1000 == 0:
78         print(‘Epoch:‘, ‘%04d‘ % (epoch + 1), ‘cost =‘, ‘{:.6f}‘.format(loss))
79
80     loss.backward()
81     optimizer.step()
82
83 # Predict [5,6,4]   (equal with target)
84 predict = model(input_batch).data.max(1, keepdim=True)[1]
85
86 # print to visual
87 print([sen.split()[:2] for sen in sentences], ‘->‘, [number_dict[n.item()] for n in predict.squeeze()])

原文地址：https://www.cnblogs.com/dhName/p/11825300.html