NLP FROM SCRATCH: CLASSIFYING NAMES WITH A CHARACTER-LEVEL RNN
原文来自于pytorch官网教程。
文章实现了一个字母级别的基础RNN模型来分类单词。其中并没有用已经提炼过的pytorch中的RNN方法，以展示RNN模型是怎样工作的。
这个模型将单词读成一个字母序列，每一步都会输出当前预测和隐藏层，隐藏层会传递给下一个字母，用最后一个结果可以对这个单词进行分类。
数据为几千个来自18个国家的人名的姓氏。通过输入这些姓氏，我们的模型应当能够判断这是哪个国家的。
$ python predict.py Hinton
(-0.47) Scottish
(-1.52) English
(-3.57) Irish
$ python predict.py Schmidhuber
(-0.19) German
(-2.48) Czech
(-2.68) Dutch
推荐阅读
pytorch相关：
- https://pytorch.org/ For installation instructions
- Deep Learning with PyTorch: A 60 Minute Blitz to get started with PyTorch in general
- Learning PyTorch with Examples for a wide and deep overview
- PyTorch for Former Torch Users if you are former Lua Torch user
RNN相关：
- The Unreasonable Effectiveness of Recurrent Neural Networks shows a bunch of real life examples
- Understanding LSTM Networks is about LSTMs specifically but also informative about RNNs in general
准备数据
在data/names目录下，有18个命名为[Language].txt的文件。每个文件里面都有很多名字，一个一行，大多数都已经转化成了我们看得懂的字母，但是仍然需要进一步规则化。
最后我们想要的是一个dictionary，每个语言作为索引可以找到它所有的名字组成的list。（{Language: [names...]}）
from __future__ import unicode_literals, print_function, division
from io import open
import glob
import os
def findFiles(path): return glob.glob(path)
print(findFiles('data/names/*.txt'))
import unicodedata
import string
all_letters = string.ascii_letters + " .,;'"
n_letters = len(all_letters)
# Turn a Unicode string to plain ASCII, thanks to https://stackoverflow.com/a/518232/2809427
def unicodeToAscii(s):
return ''.join(
c for c in unicodedata.normalize('NFD', s)
if unicodedata.category(c) != 'Mn'
and c in all_letters
)
print(unicodeToAscii('Ślusàrski'))
# Build the category_lines dictionary, a list of names per language
category_lines = {}
all_categories = []
# Read a file and split into lines
def readLines(filename):
lines = open(filename, encoding='utf-8').read().strip().split('
')
return [unicodeToAscii(line) for line in lines]
for filename in findFiles('data/names/*.txt'):
category = os.path.splitext(os.path.basename(filename))[0]
all_categories.append(category)
lines = readLines(filename)
category_lines[category] = lines
n_categories = len(all_categories)
输出：
['data/names/French.txt', 'data/names/Czech.txt', 'data/names/Dutch.txt', 'data/names/Polish.txt', 'data/names/Scottish.txt', 'data/names/Chinese.txt', 'data/names/English.txt', 'data/names/Italian.txt', 'data/names/Portuguese.txt', 'data/names/Japanese.txt', 'data/names/German.txt', 'data/names/Russian.txt', 'data/names/Korean.txt', 'data/names/Arabic.txt', 'data/names/Greek.txt', 'data/names/Vietnamese.txt', 'data/names/Spanish.txt', 'data/names/Irish.txt']
Slusarski
现在我们有category_lines作为字典，其中用语言名字可以找到对应语言文件中所有的单词。我们还需要all_categories和n_categories后面用。
将名字转化为张量
现在我们已经组织了所有名称，我们需要将它们转换为张量以使用它们。
在这里直接使用独热编码，一个字母张量为1 X n_letters，一个单词的维度大小就是line_length X 1 X n_letters。
额外的1维是因为PyTorch假定所有东西都是成批的-我们在这里只使用1的批处理大小。
import torch
# Find letter index from all_letters, e.g. "a" = 0
def letterToIndex(letter):
return all_letters.find(letter)
# Just for demonstration, turn a letter into a <1 x n_letters> Tensor
def letterToTensor(letter):
tensor = torch.zeros(1, n_letters)
tensor[0][letterToIndex(letter)] = 1
return tensor
# Turn a line into a <line_length x 1 x n_letters>,
# or an array of one-hot letter vectors
def lineToTensor(line):
tensor = torch.zeros(len(line), 1, n_letters)
for li, letter in enumerate(line):
tensor[li][0][letterToIndex(letter)] = 1
return tensor
print(letterToTensor('J'))
print(lineToTensor('Jones').size())
输出：
tensor([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0.]])
torch.Size([5, 1, 57])
建立模型
在进行自动微分之前，在Torch中创建一个递归神经网络需要在多个epoch中克隆网络层(layers)的参数，layers保留了隐藏状态和梯度，这些layers现在完全由张量图本身处理。这意味着您可以以非常“纯粹”的方式实现RNN，作为常规的前馈层。
这个RNN模块（大部分是从PyTorch for Torch用户教程中复制的）只有2个线性层，它们在输入和隐藏状态下运行，输出之后是LogSoftmax层。
注：我在实际操作中将模型有所更改，达到了相对更好一点的训练效果，下面的模型是官方的模型。我的模型在文末。
import torch.nn as nn
class RNN(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(RNN, self).__init__()
self.hidden_size = hidden_size
self.i2h = nn.Linear(input_size + hidden_size, hidden_size)
self.i2o = nn.Linear(input_size + hidden_size, output_size)
self.softmax = nn.LogSoftmax(dim=1)
def forward(self, input, hidden):
combined = torch.cat((input, hidden), 1)
hidden = self.i2h(combined)
output = self.i2o(combined)
output = self.softmax(output)
return output, hidden
def initHidden(self):
return torch.zeros(1, self.hidden_size)
n_hidden = 128
rnn = RNN(n_letters, n_hidden, n_categories)
要运行此网络的步骤，我们需要传递输入（在本例中为当前字母的张量）和先前的隐藏状态（首先将其初始化为零）。我们将返回输出（每种语言的概率）和下一个隐藏状态（我们将其保留用于下一步）。
input = letterToTensor('A')
hidden =torch.zeros(1, n_hidden)
output, next_hidden = rnn(input, hidden)
为了提高效率，我们不想为每个步骤都创建一个新的Tensor，因此我们将使用lineToTensor代替 letterToTensor和使用slice。这可以通过预先计算一批张量来进一步优化。
input = lineToTensor('Albert')
hidden = torch.zeros(1, n_hidden)
output, next_hidden = rnn(input[0], hidden)
print(output)
输出
tensor([[-2.9504, -2.8402, -2.9195, -2.9136, -2.9799, -2.8207, -2.8258, -2.8399,
-2.9098, -2.8815, -2.8313, -2.8628, -3.0440, -2.8689, -2.9391, -2.8381,
-2.9202, -2.8717]], grad_fn=<LogSoftmaxBackward>)
训练
准备训练
在开始训练之前，我们应该做一些辅助功能。首先是解释网络的输出，我们知道这是每个类别的可能性。我们可以Tensor.topk用来获取最高可能性的值：
def categoryFromOutput(output):
top_n, top_i = output.topk(1)
category_i = top_i[0].item()
return all_categories[category_i], category_i
print(categoryFromOutput(output))
输出：
('Chinese', 5)
我们还将需要快速地获取一组随机训练数据：
import random
def randomChoice(l):
return l[random.randint(0, len(l) - 1)]
def randomTrainingExample():
category = randomChoice(all_categories)
line = randomChoice(category_lines[category])
category_tensor = torch.tensor([all_categories.index(category)], dtype=torch.long)
line_tensor = lineToTensor(line)
return category, line, category_tensor, line_tensor
for i in range(10):
category, line, category_tensor, line_tensor = randomTrainingExample()
print('category =', category, '/ line =', line)
输出：
category = Italian / line = Pastore
category = Arabic / line = Toma
category = Irish / line = Tracey
category = Portuguese / line = Lobo
category = Arabic / line = Sleiman
category = Polish / line = Sokolsky
category = English / line = Farr
category = Polish / line = Winogrodzki
category = Russian / line = Adoratsky
category = Dutch / line = Robert
训练模型
现在，训练该网络所需要做的就是向它展示大量示例，进行预测，并告诉它是否错误。
损失函数用nn.NLLLoss，因为RNN的最后一层是nn.LogSoftmax。
criterion = nn.NLLLoss()
每个epoch中：
- 创建输入和目标张量
- 创建为零的初始隐藏状态
- 输入每个字母
- 记录下一个字母的隐藏状态
- 比较最终输出与目标
- 反向传播
- 返回输出和损失
learning_rate = 0.005 # If you set this too high, it might explode. If too low, it might not learn
def train(category_tensor, line_tensor):
hidden = rnn.initHidden()
rnn.zero_grad()
for i in range(line_tensor.size()[0]):
output, hidden = rnn(line_tensor[i], hidden)
loss = criterion(output, category_tensor)
loss.backward()
# Add parameters' gradients to their values, multiplied by learning rate
for p in rnn.parameters():
p.data.add_(-learning_rate, p.grad.data)
return output, loss.item()
现在，我们只需要运行大量samples。由于 train函数同时返回输出和损失，因此我们可以打印其猜测并跟踪绘制损失。由于有1000个示例，因此我们仅打印每个print_every示例，并对损失进行平均。
import time
import math
n_iters = 100000
print_every = 5000
plot_every = 1000
# Keep track of losses for plotting
current_loss = 0
all_losses = []
def timeSince(since):
now = time.time()
s = now - since
m = math.floor(s / 60)
s -= m * 60
return '%dm %ds' % (m, s)
start = time.time()
for iter in range(1, n_iters + 1):
category, line, category_tensor, line_tensor = randomTrainingExample()
output, loss = train(category_tensor, line_tensor)
current_loss += loss
# Print iter number, loss, name and guess
if iter % print_every == 0:
guess, guess_i = categoryFromOutput(output)
correct = '✓' if guess == category else '✗ (%s)' % category
print('%d %d%% (%s) %.4f %s / %s %s' % (iter, iter / n_iters * 100, timeSince(start), loss, line, guess, correct))
# Add current loss avg to list of losses
if iter % plot_every == 0:
all_losses.append(current_loss / plot_every)
current_loss = 0
输出：
5000 5% (0m 12s) 3.1806 Olguin / Irish ✗ (Spanish)
10000 10% (0m 21s) 2.1254 Dubnov / Russian ✓
15000 15% (0m 29s) 3.1001 Quirke / Polish ✗ (Irish)
20000 20% (0m 38s) 0.9191 Jiang / Chinese ✓
25000 25% (0m 46s) 2.3233 Marti / Italian ✗ (Spanish)
30000 30% (0m 54s) nan Amari / Russian ✗ (Arabic)
35000 35% (1m 3s) nan Gudojnik / Russian ✓
40000 40% (1m 11s) nan Finn / Russian ✗ (Irish)
45000 45% (1m 20s) nan Napoliello / Russian ✗ (Italian)
50000 50% (1m 28s) nan Clark / Russian ✗ (Irish)
55000 55% (1m 37s) nan Roijakker / Russian ✗ (Dutch)
60000 60% (1m 46s) nan Kalb / Russian ✗ (Arabic)
65000 65% (1m 54s) nan Hanania / Russian ✗ (Arabic)
70000 70% (2m 3s) nan Theofilopoulos / Russian ✗ (Greek)
75000 75% (2m 11s) nan Pakulski / Russian ✗ (Polish)
80000 80% (2m 20s) nan Thistlethwaite / Russian ✗ (English)
85000 85% (2m 29s) nan Shadid / Russian ✗ (Arabic)
90000 90% (2m 37s) nan Finnegan / Russian ✗ (Irish)
95000 95% (2m 46s) nan Brannon / Russian ✗ (Irish)
100000 100% (2m 54s) nan Gomulka / Russian ✗ (Polish)
绘制结果
从中绘出历史损失all_losses显示网络学习情况：
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker
plt.figure()
plt.plot(all_losses)
评估结果
# Keep track of correct guesses in a confusion matrix
confusion = torch.zeros(n_categories, n_categories)
n_confusion = 10000
# Just return an output given a line
def evaluate(line_tensor):
hidden = rnn.initHidden()
for i in range(line_tensor.size()[0]):
output, hidden = rnn(line_tensor[i], hidden)
return output
# Go through a bunch of examples and record which are correctly guessed
for i in range(n_confusion):
category, line, category_tensor, line_tensor = randomTrainingExample()
output = evaluate(line_tensor)
guess, guess_i = categoryFromOutput(output)
category_i = all_categories.index(category)
confusion[category_i][guess_i] += 1
# Normalize by dividing every row by its sum
for i in range(n_categories):
confusion[i] = confusion[i] / confusion[i].sum()
# Set up plot
fig = plt.figure()
ax = fig.add_subplot(111)
cax = ax.matshow(confusion.numpy())
fig.colorbar(cax)
# Set up axes
ax.set_xticklabels([''] + all_categories, rotation=90)
ax.set_yticklabels([''] + all_categories)
# Force label at every tick
ax.xaxis.set_major_locator(ticker.MultipleLocator(1))
ax.yaxis.set_major_locator(ticker.MultipleLocator(1))
# sphinx_gallery_thumbnail_number = 2
plt.show()
这个准确率表示方法有必要学一下，非常直观。我对我的模型也做了这样的评估。
尝试nn.LSTM或者nn.GRU，再加上一些更复杂的网络层，可以达到更好的效果。