【原创首发】训练集停止训练后，验证集加入训练（在深度学习算法中）

原理

• 在深度学习中，为了防止过拟合，被切分出来的验证集用来停止训练，但木有用于训练而导致浪费
• 本文尝试训练集停止训练后，验证集加入训练，并用多组参数进行实验

Python极简算法

图像识别示例

from keras.datasets import cifar10  # 小图像10分类
from keras.utils import to_categorical  # 独热码
from keras.models import Sequential  # 顺序模型
from keras.layers import Dense, Flatten, Conv2D, MaxPooling2D
from keras.callbacks import EarlyStopping
from sklearn.model_selection import train_test_split

# 超参
verbose = 2  # 训练过程打印
batch_size = 256  # 批大小
patience = 1  # 没有进步的轮数
callbacks = [EarlyStopping('val_acc', patience=patience)]
# 读数据
(x1, y1), (x2, y2) = cifar10.load_data()
x1, x2 = x1 / 255, x2 / 255  # 像素值→[0,1]
y1 = to_categorical(y1, 10)
y2 = to_categorical(y2, 10)


def experiment(validation_size=.1):
    # 验证集切分
    x11, x12, y11, y12 = train_test_split(x1, y1, test_size=validation_size)
    # 建模
    model = Sequential()
    for i in (32, 64):
        model.add(Conv2D(i, (3, 3), padding='same', activation='relu'))
        model.add(Conv2D(i, (3, 3), padding='same', activation='relu'))
        model.add(MaxPooling2D(pool_size=(2, 2)))
    model.add(Flatten())
    model.add(Dense(512, activation='relu'))
    model.add(Dense(10, activation='softmax'))
    model.compile('adam', 'categorical_crossentropy', ['acc'])  # 编译
    # 训练、评估
    h = model.fit(x11, y11, batch_size, 999, verbose, callbacks, validation_data=(x12, y12))
    print(validation_size, model.evaluate(x2, y2, batch_size, verbose))
    # 验证集加入训练、评估
    model.fit(x12, y12, batch_size, len(h.history['acc'])-patience, verbose)
    print(model.evaluate(x2, y2, batch_size, verbose))


experiment(.1)
experiment(.2)
experiment(.3)
experiment(.4)
experiment(.5)
experiment(.6)
experiment(.7)
experiment(.8)
experiment(.9)

文本分类示例

from gensim.models import Word2Vec
from keras.utils import to_categorical
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Dense, GRU
from keras.callbacks import EarlyStopping
from sklearn.model_selection import train_test_split
from jieba import cut
from data10 import load_xy  # https://download.csdn.net/download/Yellow_python/12862983

maxlen = 200  # 序列长度
batch_size = 256  # 每批数据量大小
verbose = 0  # 训练过程展示
callbacks = [EarlyStopping('val_acc', patience=1)]  # 没有进步的训练轮数
C = ['science', 'car', 'finance', 'sports', 'military', 'medicine', 'entertainment', 'politics', 'education', 'fashion']
num_classes = len(C)

# 读数据
x, (x1, x2, y1, y2) = load_xy()
y1 = to_categorical([C.index(i) for i in y1], num_classes)
y2 = to_categorical([C.index(i) for i in y2], num_classes)
# 词向量
wv = Word2Vec([list(cut(t)) for t in x], window=7, sg=1).wv
vectors = wv.vectors
w2i = {w: i for i, w in enumerate(wv.index2word)}
# pad
x1 = pad_sequences([[vectors[w2i[w]] for w in cut(t) if w in w2i] for t in x1], maxlen, dtype='float')
x2 = pad_sequences([[vectors[w2i[w]] for w in cut(t) if w in w2i] for t in x2], maxlen, dtype='float')


def experiment(validation_size=.1):
    # 训练数据、分词
    x11, x12, y11, y12 = train_test_split(x1, y1, test_size=validation_size, random_state=7)
    # 建模
    model = Sequential()
    model.add(GRU(64))
    model.add(Dense(units=num_classes, activation='softmax'))
    # 训练
    model.compile('adam', 'categorical_crossentropy', ['acc'])
    history = model.fit(x11, y11, batch_size, 99, verbose, callbacks, validation_data=(x12, y12))
    # 预测
    print(model.evaluate(x2, y2, batch_size, verbose), validation_size, len(history.history['acc']))
    # 验证集加入训练
    model.fit(x12, y12, batch_size, len(history.history['acc'])-1, verbose)
    print(model.evaluate(x2, y2, batch_size, verbose))


experiment(.1)
experiment(.2)
experiment(.3)
experiment(.4)
experiment(.5)
experiment(.6)
experiment(.7)
experiment(.8)
experiment(.9)

结果及结论

附1组结果

结论

图像识别结果有提升，非常不稳定但可控，建议在选择10%作为验证集并在训练结束后加入至少1次
文本分类结果有明显提升，不稳定但可控
无论哪种分类，训练集结束后，进行至少一轮验证集，能有效提升模型准确率