本例程是一个处理图像数据的标准的分布过程，利用CNN 实现对狗的品种的鉴定。

实现过程在代码中会有相关注释。

本例程数据集资源在这 https://download.csdn.net/download/rh8866/12532963

import os
#这里用到opencv，需要安装此包（用了清华镜像）
#pip install -i https://mirrors.aliyun.com/pypi/simple/  opencv-python
import cv2
import numpy as np
import pandas as pd
#Tqdm 是一个快速，可扩展的Python进度条，可以在 Python 长循环中添加一个进度提示信息
from tqdm import tqdm

import seaborn as sns
import matplotlib.pyplot as plt

import sklearn
from sklearn.model_selection import train_test_split

#读取数据的标签
labels = pd.read_csv('./dataset/dog-breed/labels.csv')
labels.head()
#id是train数据集样本图片的名称 ，breed是标签

#看看数据的分类统计
breed_count = labels['breed'].value_counts()
breed_count.head(10)

breed_count.shape
#可以看到有120个种类

# (120,)

#将标签转成one_hot形式
targets = pd.Series(labels['breed'])
one_hot = pd.get_dummies(targets,sparse=True)
one_hot_labels = np.asarray(one_hot)

#设置样本输入模型的参数
img_rows = 128
img_cols = 128
num_channel=1

#输出一张图片看看样式
img_1 = cv2.imread('./dataset/dog-breed/train/001513dfcb2ffafc82cccf4d8bbaba97.jpg',0)
plt.title('Original Image')
plt.imshow(img_1)

#输出图片shape
print(img_1.shape)

#(375, 500)

#由上面的输出可以看到图片是(375, 500) ，为了方便计算我们将其调整大小为128*128
img_1_resize = cv2.resize(img_1,(img_rows,img_cols))
print(img_1_resize.shape)
#(128, 128)
plt.title('Resized Image')
plt.imshow(img_1_resize)

x_feature = []
y_feature = []

i = 0
#加载训练数据和标签
for f,img in tqdm(labels.values):
    train_img = cv2.imread('./dataset/dog-breed/train/{}.jpg'.format(f),0)
    
    label = one_hot_labels[i]
    train_img_resize = cv2.resize(train_img,(img_rows,img_cols))
    x_feature.append(train_img_resize)
    y_feature.append(label)
    i += 1

#将数据归一化
x_train_data = np.array(x_feature,np.float32) / 255.
print(x_train_data.shape)
#为了符合keras训练数据的格式，我们给数据增加一个纬度
x_train_data = np.expand_dims(x_train_data,axis = 3)
print(x_train_data.shape)


#(10222, 128, 128)
#(10222, 128, 128, 1)

y_train_data = np.array(y_feature,np.uint8)
print(y_train_data.shape)

#(10222, 120)

#分割训练集和验证集
x_train,x_val,y_train,y_val = train_test_split(x_train_data,y_train_data,test_size=0.2,random_state=2)
print(x_train.shape)
print(x_val.shape)

#(8177, 128, 128, 1)
#(2045, 128, 128, 1)

#接下来我们对测试数据做同样的操作
submission = pd.read_csv('./dataset/dog-breed/sample_submission.csv')
test_img = submission['id']
test_img.head()

x_test_feature = []
i = 0
for f in tqdm(test_img.values):
    img = cv2.imread('./dataset/dog-breed/test/{}.jpg'.format(f),0)
    img_resize = cv2.resize(img,(img_rows,img_cols))
    x_test_feature.append(img_resize)

#数据归一化
x_test_data = np.array(x_test_feature, np.float32) / 255. 
print (x_test_data.shape)
#增加一个纬度
x_test_data = np.expand_dims(x_test_data, axis = 3)
print (x_test_data.shape)


# (10357, 128, 128)
# (10357, 128, 128, 1)

#现在我们的训练数据，验证数据以及测试聚居都准备妥当，接下来我们开始构建模型

from keras.models import Sequential
from keras.layers import Dense,Dropout
from keras.layers import Convolution2D
from keras.layers import MaxPooling2D
from keras.layers import Flatten

#我们构建一个简单的CNN 模型

model = Sequential()
#卷积层
model.add(Convolution2D(filters = 64,kernel_size = (4,4),padding = 'Same',
                       activation='relu',input_shape = (img_rows,img_cols,num_channel)))
#池化 最大池化
model.add(MaxPooling2D(pool_size=(2,2)))
#卷积层
model.add(Convolution2D(filters = 64,kernel_size = (4,4),padding='Same',
                       activation='relu'))
#最大池化
model.add(MaxPooling2D(pool_size=(2,2)))
#展平数据
model.add(Flatten())
#全连接
model.add(Dense(units=120,activation='relu'))
#输出层（由上面可知有120个输出）
model.add(Dense(units=120,activation='softmax'))

#编译模型
model.compile(optimizer='adam',loss = "categorical_crossentropy",metrics=['accuracy'])
model.summary()
#模型结构如下

#开始训练，我们设batch_size = 128，为了能节约时间 epochs为 2(我们只是关注该案例的个工作方法和流程，至于计算速度和准确度暂时我们不考虑，将在后续的博客中探讨这个问题)
batch_size = 128
nb_epochs = 2
history = model.fit(x_train,y_train,
                   batch_size=batch_size,
                   epochs=nb_epochs,
                   verbose=2,
                   validation_data=(x_val,y_val),
                   initial_epoch=0)

results = model.predict(x_test_data)
prediction = pd.DataFrame(results)
col_names = one_hot.columns.values
prediction.columns = col_names
prediction.insert(0, 'id', submission['id'])

submission = prediction

submission.head()

 

#保存预测结果
submission.to_csv('new_submission.csv', index=False)

至此本例程就结束了，精确度暂不考虑，勿喷，有好的建议欢迎评论讨论。