本教程的目的是带领大家学会如何给 stack overflow 上的问题进行打标签
首先我们需要导入要用到的函数库
代码语言:javascript复制import matplotlib.pyplot as plt
import os
import re
import shutil
import string
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras import losses
from tensorflow.keras import preprocessing
from tensorflow.keras.layers.experimental.preprocessing import TextVectorization
接下来我们看下 stack overflow 数据集,该数据集有 4 个类别标签,分别是 csharp、java、javascript、python ,每个类别有 2000 个样本,数据集下载地址: http://storage.googleapis.com/download.tensorflow.org/data/stack_overflow_16k.tar.gz
下一步是加载数据集,我们用的是 tf.keras.preprocessing.text_dataset_from_directory()
,要求的数据存放结构如下图所示
main_directory/
...class_a/
......a_text_1.txt
......a_text_2.txt
...class_b/
......b_text_1.txt
......b_text_2.txt
在开始训练前,我们需要把数据集划分成训练集、验证集、测试集,不过我们看下目录可以发现,已经存在训练集和测试集,那么还缺验证集,这个可以用validation_split
从训练集里划分出来,代码如下所示
batch_size = 32
seed = 42
raw_train_ds = tf.keras.preprocessing.text_dataset_from_directory(
'stack_overflow/train',
batch_size=batch_size,
validation_split=0.2,
subset='training',
seed=seed
)
raw_val_ds = tf.keras.preprocessing.text_dataset_from_directory(
'stack_overflow/train',
batch_size=batch_size,
validation_split=0.2,
subset='validation',
seed=seed
)
raw_test_ds = tf.keras.preprocessing.text_dataset_from_directory(
'stack_overflow/test',
batch_size=batch_size
)
在开始训练之前我们还需要对数据进行一些处理,可以通过调用 tf.keras.layers.experimental.preprocessing.TextVectorization
来进行数据的 standardize , tokenize , and vectorize
standardize: 用于移除 remove punctuation or HTML elements
tokenize: 把 strings 切分成 tokens
vectorize: 把 tokens 转化成 numbers ,然后可以送入神经网络
代码语言:javascript复制def custom_standardization(input_data):
lowercase = tf.strings.lower(input_data)
stripped_html = tf.strings.regex_replace(lowercase, '<br />', ' ')
return tf.strings.regex_replace(stripped_html,
'[%s]' % re.escape(string.punctuation),
'')
max_features = 10000
sequence_length = 125
vectorize_layer = TextVectorization(
standardize=custom_standardization,
max_tokens=max_features,
output_mode='int',
output_sequence_length=sequence_length
)
train_text = raw_train_ds.map(lambda x, y: x)
vectorize_layer.adapt(train_text)
def vectorize_text(text, label):
text = tf.expand_dims(text, -1)
return vectorize_layer(text), label
我们可以一起看下处理过后的数据长什么样子,如下图所示
到这一步,我们还需要对数据进行最后一步处理,然后就可以开始训练了
代码语言:javascript复制train_ds = raw_train_ds.map(vectorize_text)
val_ds = raw_val_ds.map(vectorize_text)
test_ds = raw_test_ds.map(vectorize_text)
AUTOTUNE = tf.data.experimental.AUTOTUNE
train_ds = train_ds.cache().prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)
test_ds = test_ds.cache().prefetch(buffer_size=AUTOTUNE)
接下来我们开始搭建模型
代码语言:javascript复制embedding_dim = 16
model = tf.keras.Sequential([
layers.Embedding(max_features 1, embedding_dim),
layers.Dropout(0.2),
layers.GlobalAveragePooling1D(),
layers.Dropout(0.2),
layers.Dense(4)
])
model.summary()
model.compile(loss=losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer='adam',
metrics=['accuracy'])
开始模型训练
代码语言:javascript复制epochs = 20
history = model.fit(
train_ds,
validation_data=val_ds,
epochs=epochs
)
绘制训练结果图
代码语言:javascript复制history_dict = history.history
history_dict.keys()
acc = history_dict['accuracy']
val_acc = history_dict['val_accuracy']
loss = history_dict['loss']
val_loss = history_dict['val_loss']
epochs = range(1, len(acc) 1)
# "bo" is for "blue dot"
plt.plot(epochs, loss, 'bo', label='Training loss')
# b is for "solid blue line"
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()
plt.show()
plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(loc='lower right')
plt.show()
我们来分析下,上面的两个图,第一个图反应的是训练损失值和验证损失值的曲线,我们发现模型过拟合了,针对这种情况我们可以用tf.keras.callbacks.EarlyStopping
来处理,只要在模型的验证损失值不再下降的地方,停止训练就好
训练完模型之后,我们可以对样本进行预测,比如 examples 里面有 3 个样本,分别截取自 stack overflow 数据集,关于预测效果,大家可以自行测试
代码地址: https://codechina.csdn.net/csdn_codechina/enterprise_technology/-/blob/master/text_classification.ipynb