使用RNN来根据语言类型和首字母生成名字

xiaoxiao2025-09-14 39

文章目录

搭建网络结构训练网络准备训练所需的数据(与上篇一致)开始训练绘制损失函数使用神经网络进行预测尝试Importance sampling

搭建网络结构

import torch import torch.nn as nn class GenerateRNN(nn.Module): def __init__(self, input_size, hidden_size, output_size): super(GenerateRNN, self).__init__() self.hidden_size = hidden_size self.c2h = nn.Linear(n_categories, hidden_size) self.i2h = nn.Linear(input_size, hidden_size) self.h2h = nn.Linear(hidden_size, hidden_size) self.activation = nn.Tanh() self.h2o = nn.Linear(hidden_size, output_size) self.dropout = nn.Dropout(0.2) self.softmax = nn.LogSoftmax(dim=1) def forward(self, category, input, hidden): c2h = self.c2h(category) i2h = self.i2h(input) h2h = self.h2h(hidden) hidden = self.activation( c2h+i2h+h2h ) dropout = self.dropout(self.h2o(hidden)) output = self.softmax(dropout) return output, hidden def initHidden(self, is_cuda=True): if is_cuda: return torch.zeros(1, self.hidden_size).cuda() else: return torch.zeros(1, self.hidden_size)

训练网络

准备训练所需的数据(与上篇一致)

all_categories ['Greek', 'Korean', 'Arabic', 'English', 'Chinese', 'Czech', 'Vietnamese', 'German', 'Polish', 'Japanese', 'Dutch', 'Italian', 'Irish', 'Portuguese', 'Russian', 'French', 'Spanish', 'Scottish'] category_lines['Chinese'] ['Ang', 'AuYong', 'Bai', 'Ban', 'Bao', 'Bei', 'Bian', 'Bui', 'Cai', 'Cao', 'Cen', 'Chai', 'Chaim', 'Chan', 'Chang', 'Chao', 'Che', 'Chen', 'Cheng', 'Cheung', 'Chew', 'Chieu', 'Chin', 'Chong', 'Chou', 'Chu', 'Cui', 'Dai', 'Deng', 'Ding', 'Dong', 'Dou', 'Duan', 'Eng', 'Fan', 'Fei', 'Feng', 'Foong', 'Fung', 'Gan', 'Gauk', 'Geng', 'Gim', 'Gok', 'Gong', 'Guan', 'Guang', 'Guo', 'Gwock', 'Han', 'Hang', 'Hao', 'Hew', 'Hiu', 'Hong', 'Hor', 'Hsiao', 'Hua', 'Huan', 'Huang', 'Hui', 'Huie', 'Huo', 'Jia', 'Jiang', 'Jin', 'Jing', 'Joe', 'Kang', 'Kau', 'Khoo', 'Khu', 'Kong', 'Koo', 'Kwan', 'Kwei', 'Kwong', 'Lai', 'Lam', 'Lang', 'Lau', 'Law', 'Lew', 'Lian', 'Liao', 'Lim', 'Lin', 'Ling', 'Liu', 'Loh', 'Long', 'Loong', 'Luo', 'Mah', 'Mai', 'Mak', 'Mao', 'Mar', 'Mei', 'Meng', 'Miao', 'Min', 'Ming', 'Moy', 'Mui', 'Nie', 'Niu', 'OuYang', 'OwYang', 'Pan', 'Pang', 'Pei', 'Peng', 'Ping', 'Qian', 'Qin', 'Qiu', 'Quan', 'Que', 'Ran', 'Rao', 'Rong', 'Ruan', 'Sam', 'Seah', 'See ', 'Seow', 'Seto', 'Sha', 'Shan', 'Shang', 'Shao', 'Shaw', 'She', 'Shen', 'Sheng', 'Shi', 'Shu', 'Shuai', 'Shui', 'Shum', 'Siew', 'Siu', 'Song', 'Sum', 'Sun', 'Sze ', 'Tan', 'Tang', 'Tao', 'Teng', 'Teoh', 'Thean', 'Thian', 'Thien', 'Tian', 'Tong', 'Tow', 'Tsang', 'Tse', 'Tsen', 'Tso', 'Tze', 'Wan', 'Wang', 'Wei', 'Wen', 'Weng', 'Won', 'Wong', 'Woo', 'Xiang', 'Xiao', 'Xie', 'Xing', 'Xue', 'Xun', 'Yan', 'Yang', 'Yao', 'Yap', 'Yau', 'Yee', 'Yep', 'Yim', 'Yin', 'Ying', 'Yong', 'You', 'Yuan', 'Zang', 'Zeng', 'Zha', 'Zhan', 'Zhang', 'Zhao', 'Zhen', 'Zheng', 'Zhong', 'Zhou', 'Zhu', 'Zhuo', 'Zong', 'Zou', 'Bing', 'Chi', 'Chu', 'Cong', 'Cuan', 'Dan', 'Fei', 'Feng', 'Gai', 'Gao', 'Gou', 'Guan', 'Gui', 'Guo', 'Hong', 'Hou', 'Huan', 'Jian', 'Jiao', 'Jin', 'Jiu', 'Juan', 'Jue', 'Kan', 'Kuai', 'Kuang', 'Kui', 'Lao', 'Liang', 'Lu', 'Luo', 'Man', 'Nao', 'Pian', 'Qiao', 'Qing', 'Qiu', 'Rang', 'Rui', 'She', 'Shi', 'Shuo', 'Sui', 'Tai', 'Wan', 'Wei', 'Xian', 'Xie', 'Xin', 'Xing', 'Xiong', 'Xuan', 'Yan', 'Yin', 'Ying', 'Yuan', 'Yue', 'Yun', 'Zha', 'Zhai', 'Zhang', 'Zhi', 'Zhuan', 'Zhui']

代表终止符

all_letters[56] "'" import random def randomChoice(l): return l[random.randint(0, len(l) - 1)] # 随机生成训练对 def randomTrainingPair(): category = randomChoice(all_categories) line = randomChoice(category_lines[category]) return category, line # 将语言类型转换为onehot向量 def categoryTensor(category): li = all_categories.index(category) tensor = torch.zeros(1, n_categories) tensor[0][li] = 1 if is_cuda: tensor = tensor.cuda() return tensor # 将名字转换为onehot矩阵 def inputTensor(line): tensor = torch.zeros(len(line), 1, n_letters) for li in range(len(line)): letter = line[li] tensor[li][0][all_letters.find(letter)] = 1 if is_cuda: tensor = tensor.cuda() return tensor # 返回的是每个正确字母对应的索引，包含EOF def targetTensor(line): letter_indexes = [all_letters.find(line[li]) for li in range(1, len(line))] letter_indexes.append(n_letters - 1) # EOS # 把最后一个字母作为EOF标记 tensor = torch.LongTensor(letter_indexes) if is_cuda: tensor = tensor.cuda() return tensor

随机生成训练集

''' 生成训练数据，其中 category_tensor是描述语言的一维onehot向量 input_line_tensor是描述输入的onehot矩阵 target_line_tensor是描述答案的索引向量，其实每一个索引可以看做是一种分类 ''' def randomTrainingExample(): category, line = randomTrainingPair() category_tensor = categoryTensor(category) input_line_tensor = inputTensor(line) target_line_tensor = targetTensor(line) return category_tensor, input_line_tensor, target_line_tensor

开始训练

criterion = nn.NLLLoss() learning_rate = 0.0005 def train(category_tensor, input_line_tensor, target_line_tensor): target_line_tensor.unsqueeze_(-1) hidden = rnn.initHidden() rnn.zero_grad() loss = 0 # Take care of the loss function, # it could be visualized as the following figure for i in range(input_line_tensor.size(0)): output, hidden = rnn(category_tensor, input_line_tensor[i], hidden) l = criterion(output, target_line_tensor[i]) loss += l loss.backward() # 通过这种方法来进行损失函数的回传 for p in rnn.parameters(): if hasattr(p.grad, "data"): p.data.add_(-learning_rate, p.grad.data) return output, loss.item() / input_line_tensor.size(0)

import time import math def timeSince(since): now = time.time() s = now - since m = math.floor(s / 60) s -= m * 60 return '%dm %ds' % (m, s) rnn = GenerateRNN(n_letters, 128, n_letters) rnn = rnn.cuda() if is_cuda else rnn n_iters = 100000 print_every = 5000 plot_every = 500 all_losses = [] total_loss = 0 # Reset every plot_every iters start = time.time() for iter in range(1, n_iters + 1): output, loss = train(*randomTrainingExample()) total_loss += loss if iter % print_every == 0: print('%s (%d %d%%) %.4f' % (timeSince(start), iter, iter / n_iters * 100, total_loss/iter)) if iter % plot_every == 0: all_losses.append(total_loss / iter) 0m 17s (5000 5%) 3.1435 0m 35s (10000 10%) 3.0208 0m 52s (15000 15%) 2.9548 1m 10s (20000 20%) 2.9098 1m 27s (25000 25%) 2.8753 1m 45s (30000 30%) 2.8430 2m 2s (35000 35%) 2.8165 2m 21s (40000 40%) 2.7949 2m 38s (45000 45%) 2.7756 2m 55s (50000 50%) 2.7574 3m 12s (55000 55%) 2.7420 3m 30s (60000 60%) 2.7273 3m 48s (65000 65%) 2.7144 4m 6s (70000 70%) 2.7016 4m 23s (75000 75%) 2.6902 4m 40s (80000 80%) 2.6800 4m 58s (85000 85%) 2.6693 5m 16s (90000 90%) 2.6585 5m 35s (95000 95%) 2.6489 5m 54s (100000 100%) 2.6394

绘制损失函数

import matplotlib.pyplot as plt import matplotlib.ticker as ticker plt.figure() plt.plot(all_losses)

使用神经网络进行预测

max_length = 20 # 输入语言的种类和开始的第一个字母，预测最终的结果 def sample(category, start_letter='A'): with torch.no_grad(): category_tensor = categoryTensor(category) input = inputTensor(start_letter) hidden = rnn.initHidden() output_name = start_letter for i in range(max_length): output, hidden = rnn(category_tensor, input[0], hidden) topv, topi = output.topk(1) topi = topi[0][0] # 判断是否遇到EOS if topi == n_letters - 1: break else: letter = all_letters[topi] output_name += letter input = inputTensor(letter) return output_name # 获取多个结果 def samples(category, start_letters='ABC'): for start_letter in start_letters: print(sample(category, start_letter)) print("\n") samples('Russian', 'CCZZYY') samples('German', 'CCZZYY') samples('Spanish', 'CCZZYY') samples('Chinese', 'CCZZYY') Charinov Cakin Zakin Zherinov Youlovev Yulovevk

Chien Cherer Zamerr Zaner Yaner Yanter

Care Carera Zoura Zales Youle Yaner

Can Chan Zhang Zan Yai Yai

samples('Chinese', 'WWWWWWWW') Wai Wan Wan Wan Wang Win Wan Wan

可以看出来对于同一个输入会产生不同的输出，这是由于dropout的存在产生的结果。

rnn.dropout = nn.Dropout(0) samples('Chinese', 'WWWWWWWW') Wan Wan Wan Wan Wan Wan Wan Wan

尝试Importance sampling

Importance sampling指的是，在生成单词每一个字母的时候，一般的生成模型都有是选择输出结果中概率值最大的字母作为预测，而Importance sampling的方式则是按照各个字母的输出概率来选择输出哪个模型。举个栗子，当程序输入一个字母C的时候，程序预测下一个字母为[a, o, e]的概率分别为[0.7, 0.2, 0.1], 普通的做法是直接选择概率值最大的a作为预测结果，而Importance Sampling的方式则是以70%的概率选择a，以20%的概率选择o，以10%的该路选择e，最终选到哪个有一定的随机性。

Tips: 直接使用模型输出的概率值会比较麻烦，因为模型的概率输出值会很小，因为这里有50多种字符，所以，一个更加明智的做法是只在概率值最高的前3个或者5个中选择，记得要将概率使用softmax等函数将其转换到和为1.

Tips: 更加简单的做法是，仅仅使用模型输出的概率值进行排序，然后使用指定的概率值进行选择，如，指定选择概率最大的那个字母的概率值为0.5，第二大的为0.3，第三大的为0.2，然后可以直接输入[0.5, 0.3, 0.2].

import numpy as np max_length = 20 def sample(category, start_letter='A'): with torch.no_grad(): category_tensor = categoryTensor(category) input = inputTensor(start_letter) hidden = rnn.initHidden() output_name = start_letter for i in range(max_length): output, hidden = rnn(category_tensor, input[0], hidden) output = torch.exp(output) topv, topi = output.topk(3) topi = topi[0] r = random.random() topv = torch.nn.functional.softmax(topv[0]) if r < topv[0]: topi = topi[0] elif r < topv[0]+topv[1]: topi = topi[1] else: topi = topi[2] if topi == n_letters - 1: break else: letter = all_letters[topi] output_name += letter input = inputTensor(letter) return output_name def samples(category, start_letters='ABC'): for start_letter in start_letters: print(sample(category, start_letter)) print("\n")

最新回复(0)