旨在使用keras構(gòu)建出二分類和多分類模型，給出相關(guān)代碼。
機(jī)器學(xué)習(xí)問題中，二分類和多分類問題是最為常見，下面使用keras在imdb和newswires數(shù)據(jù)上進(jìn)行相應(yīng)的實(shí)驗(yàn)。
[code: https://github.com/zylhub/More_Python/blob/master/keras_TOT/simple-network-imdb.py ]
imdb 二分類
- 文本獲取
- 文本預(yù)處理
- 定義模型/神經(jīng)網(wǎng)絡(luò)
- 定義評價(jià)指標(biāo)和優(yōu)化方法
- 劃分?jǐn)?shù)據(jù)集，進(jìn)行訓(xùn)練
- 測試集對模型進(jìn)行測試
# coding=utf-8
# @Time : 17-12-22 下午11:01
# @Author : knight
# @File : simple-network-imdb.py
from keras import layers
from keras import models
from keras.datasets import imdb
import numpy as np
# 1. 加載數(shù)據(jù)
(train_data, train_labels), (test_data, test_labels) = imdb.load_data(num_words=10000)
# 2. 定義模型
model = models.Sequential()
model.add(layers.Dense(32, activation='relu', input_shape=(10000,)))
model.add(layers.Dense(16, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
######################
# input_tensor = layers.Input(shape=(784,))
# x = layers.Dense(32, activation='relu')(input_tensor)
# output_tensor = layers.Dense(10, activation='softmax')(x)
# model = models.Model(input=input_tensor, output=output_tensor)
# 3.1 文本數(shù)據(jù)預(yù)處理
word_index = imdb.get_word_index() # 獲取詞索引 id to word
reverse_word_index = dict((value, key) for key, value in word_index.items()) # word to id
decoded_review = ' '.join([reverse_word_index.get(i-3, '?') for i in train_data[0]])
# 3.2 特征向量化
def vectorize_sequences(sequences, dim=10000):
"""
詞袋模型，獲取詞向量
:param sequences: [[sentence1], [sentence2], [...]]
:param dim: 詞袋大小，詞特征維度
:return:
"""
results = np.zeros((len(sequences), dim))
for i, sequence in enumerate(sequences):
results[i, sequence] = 1.0
return results
x_train = vectorize_sequences(train_data)
x_test = vectorize_sequences(test_data)
y_train = np.asarray(train_labels).astype('float32')
y_test = np.asarray(test_labels).astype('float32')
from keras import metrics, losses
model.compile(optimizer='rmsprop',
loss=losses.binary_crossentropy,
metrics=[metrics.binary_accuracy]) # metrics 傳入list，可以使用多種評價(jià)方式
# 劃分驗(yàn)證集
x_val = x_train[:10000]
partial_x_train = x_train[10000:]
y_val = y_train[:10000]
partial_y_train = y_train[10000:]
# 4. 訓(xùn)練模型
history = model.fit(partial_x_train, partial_y_train,
epochs=20,
batch_size=512,
validation_data=(x_val, y_val)) # 驗(yàn)證集
# history 獲取訓(xùn)練過程的acc loss val_acc val_loss
history_dict = history.history
print(history_dict.keys())
# 5. Plotting the training and validation loss
import matplotlib.pyplot as plt
# 畫出訓(xùn)練集和驗(yàn)證集的損失和精度變化，分析模型狀態(tài)
acc = history.history['binary_accuracy'] # 訓(xùn)練集acc
val_acc = history.history['val_binary_accuracy'] # 驗(yàn)證集 acc
loss = history.history['loss'] # 訓(xùn)練損失
val_loss = history.history['val_loss'] # 驗(yàn)證損失
epochs = range(1, len(acc)+1) # 迭代次數(shù)
plt.plot(epochs, loss, 'bo', label='Training loss') # bo for blue dot 藍(lán)色點(diǎn)
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.clf() # clar figure
plt.plot(epochs, acc, 'bo', label='Training acc') # bo for blue dot 藍(lán)色點(diǎn)
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()
plt.show()
# ----------------------------------------使用model 定義網(wǎng)絡(luò)
input_tensor = layers.Input(shape=(10000,))
x = layers.Dense(32, activation='relu')(input_tensor)
x = layers.Dense(16, activation='relu')(x)
x = layers.Dense(16, activation='relu')(x)
output_tensor = layers.Dense(1, activation='sigmoid')(x)
network = models.Model(inputs=input_tensor, outputs=output_tensor)
network.compile(optimizer='rmsprop',
loss='binary_crossentropy',
metrics=['accuracy'])
network.fit(x_train, y_train, epochs=4, batch_size=512)
print("test data evaluate, epochs=20", model.evaluate(x_test, y_test))
print("test data evaluate, epochs=4 ", network.evaluate(x_test, y_test))
epochs=20_acc.png
epochs=20_loss.png
根據(jù)訓(xùn)練集和驗(yàn)證集的loss和acc，可以判斷出模型過擬合了，此時(shí)應(yīng)該嘗試早停止，比如在Epochs=4的時(shí)候（迭代次數(shù)到4之后，訓(xùn)練集的loss繼續(xù)降低，驗(yàn)證集loss增加了，同時(shí)驗(yàn)證集的acc開始下降，模型學(xué)習(xí)的目的是對未來的數(shù)據(jù)有較好的泛化能力，因此選擇驗(yàn)證集較好的時(shí)候作為可用的模型，會(huì)得到較好的結(jié)果 ）。
newswires 多標(biāo)簽多分類
構(gòu)建一個(gè)基本的神經(jīng)網(wǎng)絡(luò)來解決文本分類問題，先給出網(wǎng)絡(luò)結(jié)構(gòu)
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 64) 640064
_________________________________________________________________
dense_2 (Dense) (None, 64) 4160
_________________________________________________________________
dense_3 (Dense) (None, 46) 2990
=================================================================
Total params: 647,214
Trainable params: 647,214
Non-trainable params: 0
_________________________________________________________________
# coding=utf-8
# ---------------------------
# @Time : 17-12-23 下午10:05
# @Author : knight
# @File : simple-network-newswires.py
# ---------------------------
from __future__ import absolute_import
from keras.datasets import reuters
from keras_TOT.utils_text import vectorize_sequences, to_one_hot
(train_data, train_labels), (test_data, test_labels) = reuters.load_data(num_words=10000)
# 8,982 training examples and 2,246 test examples:
print('train len :', len(train_data))
print('test len: ', len(test_data))
word_index = reuters.get_word_index() # 獲取詞袋
revserse_word_index = dict((value, key) for key, value in word_index.items())
decoded_newswire = ' '.join(revserse_word_index.get(i-3, '?') for i in train_data[0])
print('decoded_newswire', decoded_newswire)
x_train = vectorize_sequences(train_data)
x_test = vectorize_sequences(test_data)
###------------------------------------###
one_hot_train_labels = to_one_hot(train_labels)
one_hot_test_labels = to_one_hot(test_labels)
# 也可以直接使用keras提供的one-hot 方法
# from keras.utils.np_utils import to_categorical
# one_hot_train_labels = to_categorical(train_labels)
# one_hot_test_labels = to_categorical(test_labels)
###------------------------------------###
from keras import models
from keras import layers
network = models.Sequential() # 序列方式構(gòu)建模型
network.add(layers.Dense(64, activation='relu', input_shape=(10000,)))
network.add(layers.Dense(64, activation='relu'))
network.add(layers.Dense(46, activation='softmax')) # 多分類問題常用的激活函數(shù)softmax
network.compile(optimizer='rmsprop',
loss='categorical_crossentropy', # 交叉熵
metrics=['accuracy'])
x_val = x_train[:1000]
partial_x_train = x_train[1000:]
y_val = one_hot_train_labels[:1000]
partial_y_train = one_hot_train_labels[1000:]
history = network.fit(partial_x_train, partial_y_train, epochs=20, validation_data=(x_val, y_val))
import matplotlib.pyplot as plt
loss = history.history['loss']
val_loss = history.history['val_loss']
epoch = range(1, len(loss)+1)
plt.plot(epoch, loss, 'bo', label='Training loss')
plt.plot(epoch, val_loss, 'b', label='Training val_loss')
plt.title("Training and validation loss")
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()
plt.show()
plt.clf()
acc = history.history['acc']
val_acc = history.history['val_acc']
plt.plot(epoch, acc, 'bo', label='Training acc')
plt.plot(epoch, val_loss, 'b', label='Training val_loss')
plt.title('Training and validation acc')
plt.xlabel('Epochs')
plt.ylabel('acc')
plt.legend()
plt.show()
