Keras猫狗大战一：小样本4层卷积网络，74%精度

版权声明：本文为博主原创文章，欢迎转载，并请注明出处。联系方式：460356155@qq.com

 一、下载数据集

百度搜索“kaggle 猫狗数据集”，可找到网盘共享的猫狗数据集，有815M。

二、准备数据集

整个数据集有25000张图，猫狗各12500，从中选取1000、500、200分别作为训练、验证、测试集。

import os
import random
import shutil


# 随机得到样本子集
def get_sub_sample(sample_path, target_path, train, valid, test, file_name_format=None, class_name=None,
                   class_num=None):
    """
    sample_path: 样本全集目录
    target_path: 样本子集目录
    train, valid, test: 随机选取训练、验证、测试样本数
    file_name_format：文件名格式过滤
    class_name：样本类型名
    class_num: 样本数
    """
    # 得到样本全集目录下的所有文件，不遍历子目录
    all_files = [f for f in os.listdir(sample_path) if os.path.isfile(os.path.join(sample_path, f))]
    total = len(all_files)

    if file_name_format:
        # 针对一个目录放多种类型情况
        num_per_class = int(total / class_num)
        fnames = [file_name_format.format(i) for i in range(num_per_class)]
    else:
        fnames = all_files

    # 打乱顺序
    random.shuffle(fnames)

    os.makedirs(os.path.join(target_path, 'train', class_name))
    os.makedirs(os.path.join(target_path, 'valid', class_name))
    os.makedirs(os.path.join(target_path, 'test', class_name))

    for i in range(train):
        src = os.path.join(sample_path, fnames[i])
        dst = os.path.join(target_path, 'train', class_name, fnames[i])
        shutil.copyfile(src, dst)

    for i in range(train, train + valid):
        src = os.path.join(sample_path, fnames[i])
        dst = os.path.join(target_path, 'valid', class_name, fnames[i])
        shutil.copyfile(src, dst)

    for i in range(train + valid, train + valid + test):
        src = os.path.join(sample_path, fnames[i])
        dst = os.path.join(target_path, 'test', class_name, fnames[i])
        shutil.copyfile(src, dst)


src_path = r'D:BaiduNetdiskDownload	rain'
dst_path = r'D:BaiduNetdiskDownloadsmall'
train_dir = os.path.join(dst_path, 'train')
validation_dir = os.path.join(dst_path, 'valid')
class_name = ['cat', 'dog']

if os.path.exists(dst_path):
    shutil.rmtree(dst_path)

os.makedirs(dst_path)

for cls in class_name:
    get_sub_sample(src_path, dst_path, 1000, 500, 200, file_name_format='%s.{}.jpg' % (cls), class_name=cls,
                   class_num=2)

三、模型建立

from keras import layers
from keras import models

model = models.Sequential()

# 输出图片尺寸：150-3+1=148*148，参数数量：32*3*3*3+32=896
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(150, 150, 3)))
model.add(layers.MaxPooling2D((2, 2)))  # 输出图片尺寸：148/2=74*74

# 输出图片尺寸：74-3+1=72*72，参数数量：64*3*3*32+64=18496
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))  # 输出图片尺寸：72/2=36*36

# 输出图片尺寸：36-3+1=34*34，参数数量：128*3*3*64+128=73856
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))  # 输出图片尺寸：34/2=17*17

# 输出图片尺寸：17-3+1=15*15，参数数量：128*3*3*128+128=147584
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))  # 输出图片尺寸：15/2=7*7

#  多维转为一维：7*7*128=6272
model.add(layers.Flatten())

#  参数数量：6272*512+512=3211776
model.add(layers.Dense(512, activation='relu'))

#  参数数量：512*1+1=513
model.add(layers.Dense(1, activation='sigmoid'))

model.summary()

四、模型compile

from keras import optimizers

#  二分类用binary_crossentropy
model.compile(loss='binary_crossentropy',
              optimizer=optimizers.RMSprop(lr=1e-4),
              metrics=['acc'])

五、建立训练和验证数据

from keras.preprocessing.image import ImageDataGenerator

#  归一化
train_datagen = ImageDataGenerator(rescale=1. / 255)
test_datagen = ImageDataGenerator(rescale=1. / 255)

train_generator = train_datagen.flow_from_directory(
    train_dir,
    #  输入训练图像尺寸
    target_size=(150, 150),
    batch_size=20,
    #  二分类
    class_mode='binary')

validation_generator = test_datagen.flow_from_directory(
    validation_dir,
    target_size=(150, 150),
    batch_size=20,
    class_mode='binary')

六、训练

history = model.fit_generator(
    train_generator,
    # 2000张图 / 20 batch size
    steps_per_epoch=100,
    epochs=30,
    validation_data=validation_generator,
    # 1000张图 / 20 batch size
    validation_steps=50)

WARNING:tensorflow:From d:program filespython37libsite-packages	ensorflowpythonopsmath_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.cast instead.
Epoch 1/30
100/100 [==============================] - 41s 409ms/step - loss: 0.6903 - acc: 0.5255 - val_loss: 0.6730 - val_acc: 0.6070
Epoch 2/30
100/100 [==============================] - 41s 406ms/step - loss: 0.6599 - acc: 0.6070 - val_loss: 0.6350 - val_acc: 0.6510
Epoch 3/30
100/100 [==============================] - 41s 408ms/step - loss: 0.6135 - acc: 0.6710 - val_loss: 0.6223 - val_acc: 0.6400
Epoch 4/30
100/100 [==============================] - 41s 410ms/step - loss: 0.5816 - acc: 0.6960 - val_loss: 0.5798 - val_acc: 0.6950
Epoch 5/30
100/100 [==============================] - 41s 411ms/step - loss: 0.5582 - acc: 0.7160 - val_loss: 0.5757 - val_acc: 0.6970
Epoch 6/30
100/100 [==============================] - 42s 420ms/step - loss: 0.5278 - acc: 0.7360 - val_loss: 0.5788 - val_acc: 0.6790
Epoch 7/30
100/100 [==============================] - 41s 412ms/step - loss: 0.5096 - acc: 0.7485 - val_loss: 0.5551 - val_acc: 0.7140
Epoch 8/30
100/100 [==============================] - 42s 418ms/step - loss: 0.4809 - acc: 0.7715 - val_loss: 0.5871 - val_acc: 0.6870
Epoch 9/30
100/100 [==============================] - 42s 416ms/step - loss: 0.4645 - acc: 0.7850 - val_loss: 0.5309 - val_acc: 0.7370
Epoch 10/30
100/100 [==============================] - 42s 415ms/step - loss: 0.4348 - acc: 0.7960 - val_loss: 0.5618 - val_acc: 0.7160
Epoch 11/30
100/100 [==============================] - 42s 420ms/step - loss: 0.4133 - acc: 0.8050 - val_loss: 0.5714 - val_acc: 0.7210
Epoch 12/30
100/100 [==============================] - 41s 409ms/step - loss: 0.3847 - acc: 0.8215 - val_loss: 0.5937 - val_acc: 0.7030
Epoch 13/30
100/100 [==============================] - 41s 413ms/step - loss: 0.3523 - acc: 0.8465 - val_loss: 0.6225 - val_acc: 0.7030
Epoch 14/30
100/100 [==============================] - 42s 416ms/step - loss: 0.3339 - acc: 0.8535 - val_loss: 0.5339 - val_acc: 0.7500
Epoch 15/30
100/100 [==============================] - 43s 428ms/step - loss: 0.3013 - acc: 0.8650 - val_loss: 0.5404 - val_acc: 0.7520
Epoch 16/30
100/100 [==============================] - 42s 417ms/step - loss: 0.2736 - acc: 0.8885 - val_loss: 0.5885 - val_acc: 0.7380
Epoch 17/30
100/100 [==============================] - 41s 415ms/step - loss: 0.2562 - acc: 0.8995 - val_loss: 0.5636 - val_acc: 0.7420
Epoch 18/30
100/100 [==============================] - 41s 415ms/step - loss: 0.2294 - acc: 0.9115 - val_loss: 0.5722 - val_acc: 0.7490
Epoch 19/30
100/100 [==============================] - 42s 415ms/step - loss: 0.2004 - acc: 0.9210 - val_loss: 0.6201 - val_acc: 0.7390
Epoch 20/30
100/100 [==============================] - 41s 413ms/step - loss: 0.1812 - acc: 0.9315 - val_loss: 0.6323 - val_acc: 0.7390
Epoch 21/30
100/100 [==============================] - 42s 423ms/step - loss: 0.1551 - acc: 0.9495 - val_loss: 0.5949 - val_acc: 0.7530
Epoch 22/30
100/100 [==============================] - 50s 500ms/step - loss: 0.1438 - acc: 0.9505 - val_loss: 0.6145 - val_acc: 0.7500
Epoch 23/30
100/100 [==============================] - 45s 447ms/step - loss: 0.1131 - acc: 0.9660 - val_loss: 0.7587 - val_acc: 0.7340
Epoch 24/30
100/100 [==============================] - 42s 415ms/step - loss: 0.1012 - acc: 0.9650 - val_loss: 0.7000 - val_acc: 0.7500
Epoch 25/30
100/100 [==============================] - 42s 425ms/step - loss: 0.0852 - acc: 0.9765 - val_loss: 0.7501 - val_acc: 0.7400
Epoch 26/30
100/100 [==============================] - 43s 427ms/step - loss: 0.0730 - acc: 0.9785 - val_loss: 0.7945 - val_acc: 0.7500
Epoch 27/30
100/100 [==============================] - 41s 410ms/step - loss: 0.0643 - acc: 0.9825 - val_loss: 0.7769 - val_acc: 0.7480
Epoch 28/30
100/100 [==============================] - 41s 415ms/step - loss: 0.0544 - acc: 0.9860 - val_loss: 0.8410 - val_acc: 0.7530
Epoch 29/30
100/100 [==============================] - 41s 410ms/step - loss: 0.0435 - acc: 0.9910 - val_loss: 0.8678 - val_acc: 0.7670
Epoch 30/30
100/100 [==============================] - 41s 411ms/step - loss: 0.0370 - acc: 0.9920 - val_loss: 0.8941 - val_acc: 0.7640

在第9次迭代时，验证损失达到最小，验证精度在74%左右，随着迭代次数增加，出现了过拟合。显示训练曲线：

% matplotlib inline
import matplotlib.pyplot as plt

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc))

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()

plt.figure()

plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()

plt.show()

七、保存模型

model.save('cats_and_dogs_small_1.h5')