前言

希望保研顺顺利利的，好慌=.=
csdn啥时候支持代码折叠啊啊啊啊啊
使用windows10+CPU+pycharm+tensorflow.keras+python搭建简单的卷积神经网络模型，并进行猫狗分类
数据来源于kaggle-cat vs. dog，包含train和test两个文件夹，train中图像有标签，test中无标签

正文

导入相应的文件处理、数据处理、深度学习、机器学习的库

主要是pandas, numpy, keras, tensorflow.keras, sklearn, random, os（keras和tensorflow.keras的区别？建议使用tensorflow.keras）
源码：

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

from keras.preprocessing.image import ImageDataGenerator, load_img
from keras.utils import to_categorical
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Conv2D, MaxPool2D,\
    Dropout, Flatten, Dense, Activation, BatchNormalization
from keras.callbacks import EarlyStopping, ReduceLROnPlateau

from sklearn.model_selection import train_test_split
import random
import os

标记数据标签

定义关于训练集和验证集图像的信息
此处图像命名规则为：cat/dog.*.jpg，所以将数据集与标签对应好，得到pd.DataFrame数据
源码：

# parameters
FAST_RUN = False
image_width = 128
image_height = 128
image_channels = 3
image_size = (image_width, image_height)
input_shape = (image_width, image_height, image_channels)

filenames = os.listdir('train')  # type:list
random.shuffle(filenames)  # no return
categories = []
for filename in filenames:
	categories.append(filename.split('.')[0])
#    if filename.split('.')[0] == 'dog':
#        categories.append(1)
#    else:
#        categories.append(0)

df = pd.DataFrame({
    'filename': filenames,
    'category': categories
})      # type:pd.DataFrame

搭建卷积神经网络模型

该神经网络模型总计有四层：（卷积层+最大池化层）x2、全连接层x1、输出层softmax，此处为二分类问题输出层可以使用sigmoid激活函数，模型如下图所示：
每层都使用dropout防止过拟合；并且使用BatchNormalizationj加速了网络训练速度、提高网络的鲁棒性；优化算法使用adam算法
源码：


# 2 build the model
model = Sequential()

model.add(Conv2D(32, (3, 3), activation='relu', input_shape=input_shape))
model.add(BatchNormalization())
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(BatchNormalization())
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(0.5))
model.add(Dense(2, activation='softmax'))

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()	# show the model

模型细节：

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 126, 126, 32)      896       
_________________________________________________________________
batch_normalization (BatchNo (None, 126, 126, 32)      128       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 63, 63, 32)        0         
_________________________________________________________________
dropout (Dropout)            (None, 63, 63, 32)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 61, 61, 64)        18496     
_________________________________________________________________
batch_normalization_1 (Batch (None, 61, 61, 64)        256       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 30, 30, 64)        0         
_________________________________________________________________
dropout_1 (Dropout)          (None, 30, 30, 64)        0         
_________________________________________________________________
flatten (Flatten)            (None, 57600)             0         
_________________________________________________________________
dense (Dense)                (None, 512)               29491712  
_________________________________________________________________
batch_normalization_2 (Batch (None, 512)               2048      
_________________________________________________________________
dropout_2 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 2)                 1026      
=================================================================
Total params: 29,514,562
Trainable params: 29,513,346
Non-trainable params: 1,216

数据增强

使用sklearn.model_selection.train_test.split将数据集划分为训练集和验证集
这里有个点不太明白，ImageDataGenerator和ImageDataGenerator.flow_from_dataframe如何区分，或者说，二者具体的职能是什么？按照名字来看，ImageDataGenerator是一个增强图像生成器，而调用ImageDataGenerator.flow_from_dataframe的过程就像是把图像喂给生成器然后产生增强后的图像，详情见keras中文文档
回调函数是
源码：

#df['category'] = df['category'].replace({
#    0: 'cat',
#    1: 'dog'
#})

train_df, validate_df = train_test_split(df, test_size=0.2, random_state=42)    # type:np.ndarray
train_df = train_df.reset_index(drop=True)
validate_df = validate_df.reset_index(drop=True)

total_train = train_df.shape[0]
total_validate = validate_df.shape[0]
batch_size = 15

train_datagen = ImageDataGenerator(
    rotation_range=15,
    rescale=1./255,
    shear_range=0.1,
    zoom_range=0.2,
    horizontal_flip=True,
    width_shift_range=0.1,
    height_shift_range=0.1
)

train_genertor = train_datagen.flow_from_dataframe(
    train_df,	# 训练集DataFrame格式
    'train',	# 数据集的位置
    x_col='filename',
    y_col='category',
    target_size=image_size,
    class_mode='categorical',	# 个人理解：多类中选择单类
    batch_size=batch_size
)

validation_datagen = ImageDataGenerator(rescale=1./255)
validation_generator = validation_datagen.flow_from_dataframe(
    validate_df,
    'train',
    x_col='filename',
    y_col='category',
    target_size=image_size,
    class_mode='categorical',
    batch_size=batch_size
)

模型训练

模型训练需要：训练集、验证集、总迭代次数、批量数、回调函数列表（训练时使用）
在训练时出现警告WARNING:tensorflow:Reduce LR on plateau conditioned on metric `val_acc` which is not available. Available metrics are: loss,accuracy,val_loss,val_accuracy,lr，有人说是验证集太小导致的
一个epoch大概训练10多分钟，用CPU训练了一晚上=.=，如果只是想练习，建议把FAST_RUN = True
源码：

epochs = 3 if FAST_RUN else 50		# total iterations

early_stop = EarlyStopping(patience=10)		# early_stop
# learn rate decay
learn_rate_reduction = ReduceLROnPlateau(monitor='val_acc', patience=2,
                                         verbose=1, factor=0.5, min_lr=0.00001)
callbacks = [early_stop, learn_rate_reduction]

history = model.fit_generator(
    train_genertor,
    epochs=epochs,
    validation_data=validation_generator,
    validation_steps=total_validate//batch_size,
    steps_per_epoch=total_train//batch_size,
    callbacks=callbacks			# list of callback functions
)

model.save_weights('model.h5')	# save the parameters of model to h5 file

训练结果：

···
1333/1333 [==============================] - 685s 514ms/step - loss: 0.4195 - accuracy: 0.8084 - val_loss: 0.4752 - val_accuracy: 0.7862
Epoch 46/50
1333/1333 [==============================] - 681s 511ms/step - loss: 0.4291 - accuracy: 0.8017 - val_loss: 0.3621 - val_accuracy: 0.8432
WARNING:tensorflow:Reduce LR on plateau conditioned on metric `val_acc` which is not available. Available metrics are: loss,accuracy,val_loss,val_accuracy,lr
Epoch 47/50
1333/1333 [==============================] - 679s 509ms/step - loss: 0.4165 - accuracy: 0.8087 - val_loss: 0.3811 - val_accuracy: 0.8308
WARNING:tensorflow:Reduce LR on plateau conditioned on metric `val_acc` which is not available. Available metrics are: loss,accuracy,val_loss,val_accuracy,lr
Epoch 48/50
1333/1333 [==============================] - 687s 515ms/step - loss: 0.4407 - accuracy: 0.7967 - val_loss: 0.3781 - val_accuracy: 0.8374
WARNING:tensorflow:Reduce LR on plateau conditioned on metric `val_acc` which is not available. Available metrics are: loss,accuracy,val_loss,val_accuracy,lr
Epoch 49/50
1333/1333 [==============================] - 689s 517ms/step - loss: 0.4129 - accuracy: 0.8113 - val_loss: 0.4771 - val_accuracy: 0.7722
WARNING:tensorflow:Reduce LR on plateau conditioned on metric `val_acc` which is not available. Available metrics are: loss,accuracy,val_loss,val_accuracy,lr
Epoch 50/50
1333/1333 [==============================] - 680s 510ms/step - loss: 0.4152 - accuracy: 0.8113 - val_loss: 1.6649 - val_accuracy: 0.5908
WARNING:tensorflow:Reduce LR on plateau conditioned on metric `val_acc` which is not available. Available metrics are: loss,accuracy,val_loss,val_accuracy,lr

Process finished with exit code 0

模型测试

用和得到训练集和验证集一样的方法得到测试集
使用predict_genrator预测测试集
将结果存起来，最后转成csv格式
源码：

# get the test set genrator
test_filenames = os.listdir('test')
test_df = pd.DataFrame({
    'filename': test_filenames
})
number_samples = test_df.shape[0]
test_gen = ImageDataGenerator(rescale=1./255)
test_generator = test_gen.flow_from_dataframe(
    test_df,
    'test',
    x_col='filename',
    y_col=None,
    class_mode=None,
    target_size=image_size,
    batch_size=batch_size,
    shuffle=False
)

# predict the result
predict = model.predict_generator(test_generator,
                                  steps=np.ceil(number_samples/batch_size))
test_df['category'] = np.argmax(predict, axis=-1)	# return index value 0，1

# tranform label
# 不懂=.=
label_map = dict((v, k) for k, v in train_generator.class_indices.items())
test_df['category'] = test_df['category'].replace(label_map)
test_df['category'] = test_df['category'].replace({
    'dog': 1,
    'cat': 0
})

# convert to csv format
submission_df = test_df.copy()
submission_df['id'] = submission_df['filename'].str.split('.').str[0]
submission_df.drop(['filename', 'category'], axis=1, inplace=True)
submission_df.to_csv('submission.csv', index=False)

完整源码

import pandas as pd
import numpy as np

from keras.preprocessing.image import ImageDataGenerator
from keras.utils import to_categorical
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Conv2D, MaxPool2D,\
    Dropout, Flatten, Dense, Activation, BatchNormalization
from keras.callbacks import EarlyStopping, ReduceLROnPlateau

from sklearn.model_selection import train_test_split
import random
import os

# 1 preprocessing
# input parameters
FAST_RUN = True
image_width = 128
image_height = 128
image_channels = 3
image_size = (image_width, image_height)
input_shape = (image_width, image_height, image_channels)

filenames = os.listdir('train')  # type:list
random.shuffle(filenames)  # no return
categories = []
for filename in filenames:
    categories.append(filename.split('.')[0])
    # if filename.split('.')[0] == 'dog':
    #     categories.append(1)
    # else:
    #     categories.append(0)

df = pd.DataFrame({
    'filename': filenames,
    'category': categories
})      # type:pd.DataFrame

# 2 build the model
model = Sequential()

model.add(Conv2D(32, (3, 3), activation='relu', input_shape=input_shape))
model.add(BatchNormalization())
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(BatchNormalization())
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(0.5))
model.add(Dense(2, activation='softmax'))

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()

early_stop = EarlyStopping(patience=10)
learn_rate_reduction = ReduceLROnPlateau(monitor='val_acc', patience=2,
                                         verbose=1, factor=0.5, min_lr=0.00001)
callbacks = [early_stop, learn_rate_reduction]

# df['category'] = df['category'].replace({
#     0: 'cat',
#     1: 'dog'
# })

train_df, validate_df = train_test_split(df, test_size=0.2, random_state=42)    # type:np.ndarray
train_df = train_df.reset_index(drop=True)
validate_df = validate_df.reset_index(drop=True)

total_train = train_df.shape[0]
total_validate = validate_df.shape[0]
batch_size = 15

train_datagen = ImageDataGenerator(
    rotation_range=15,
    rescale=1./255,
    shear_range=0.1,
    zoom_range=0.2,
    horizontal_flip=True,
    width_shift_range=0.1,
    height_shift_range=0.1
)

train_generator = train_datagen.flow_from_dataframe(
    train_df,
    'train',
    x_col='filename',
    y_col='category',
    target_size=image_size,
    class_mode='categorical',
    batch_size=batch_size
)

validation_datagen = ImageDataGenerator(rescale=1./255)
validation_generator = validation_datagen.flow_from_dataframe(
    validate_df,
    'train',
    x_col='filename',
    y_col='category',
    target_size=image_size,
    class_mode='categorical',
    batch_size=batch_size
)

epochs = 3 if FAST_RUN else 50
history = model.fit_generator(
    train_generator,
    epochs=epochs,
    validation_data=validation_generator,
    validation_steps=total_validate//batch_size,
    steps_per_epoch=total_train//batch_size,
    callbacks=callbacks
)

model.save_weights('model.h5')

test_filenames = os.listdir('test')
test_df = pd.DataFrame({
    'filename': test_filenames
})
number_samples = test_df.shape[0]
test_gen = ImageDataGenerator(rescale=1./255)
test_generator = test_gen.flow_from_dataframe(
    test_df,
    'test',
    x_col='filename',
    y_col=None,
    class_mode=None,
    target_size=image_size,
    batch_size=batch_size,
    shuffle=False
)

predict = model.predict_generator(test_generator,
                                  steps=np.ceil(number_samples/batch_size))
test_df['category'] = np.argmax(predict, axis=-1)
label_map = dict((v, k) for k, v in train_generator.class_indices.items())
print(label_map)
test_df['category'] = test_df['category'].replace(label_map)

test_df['category'] = test_df['category'].replace({
    'dog': 1,
    'cat': 0
})

submission_df = test_df.copy()
submission_df['id'] = submission_df['filename'].str.split('.').str[0]
submission_df.drop(['filename', 'category'], axis=1, inplace=True)
submission_df.to_csv('submission.csv', index=False)

总结

基本卷积神经网络的训练流程
以下问题对应python不熟悉的同学（比如我>.<）来说，肯定会经常搞不明白，下面是我的总结

os.listdir('.')
# 本级目录

test_df['category'] = np.argmax(predict, axis=-1)
# axis=-1表示倒数第一维

train_df.shape[0]
# 行数，第一维数

axis=0 vs. axis=1
# 个人理解：0、1、2表示第0、1、2维度，在二维空间中，0表示行、1表示列，当在函数中出
# 现时，如axis=0，表示跨行进行运算，即在某一列上进行运算，但本质上应该按照跨行理解，
# 因为在高维空间中时，这样理解方面运算

index=False vs. index=True
# 一般存在一些数据‘清洗’函数中
# index = False：输出不显示 index (索引)值
# index = True：输出显示 index (索引)值

inplace=False vs. inplace=True
# inplace表示是否原地操作，即是否直接对对象进行修改，比如：x=x+5，你是否要存储x+5的值？
# inplace=True means that it will modify the input directly, without allocating any
# additional output. It can sometimes slightly decrease the memory usage, but may 
# not always be a valid operation (because the original input is destroyed).
# However, if you don’t see an error, it means that your use case is valid.