RNN姓氏分类：官方教程翻译

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

数据下载 https://download.pytorch.org/tutorial/data.zip

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。

Github项目

官方

我的