Tensorflow keras动物分类[keras学习笔记]

学习笔记

猫，狗， 象，牛，猪五分类问题

样本

库

网络

• 模型
• 编译
• 训练
• 结果

样本

猫(cat, 200)

大象(elefante, 215)

狗(dog, 205)

牛(mucca, 210)

猪(pig, 147)

总共977张，5类图片

数据集下载

链接：https://pan.baidu.com/s/1N2fHG0r3YkiQvQ7VOUF7Hw 
提取码：8ulu 
复制这段内容后打开百度网盘手机App，操作更方便哦--来自百度网盘超级会员V5的分享

库(推荐使用anaconda进行包管理)

python 3.6
tensorflow 1.14.0
numpy 1.19.5
glob2 0.7
matplotlib 3.3.3

代码

import tensorflow as tf
import numpy as np
import glob
import matplotlib.pyplot as plt

数据预处理

imags_path = glob.glob('./data/animals/*/*.jpg') # 读取所有图片路径

all_labels_name = [imag_p.split("\\")[1] for imag_p in imags_path] # 获取标签名

label_names = np.unique(all_labels_name) # 去除重复的

label_to_index = dict((name, i) for i, name in enumerate(label_names))
print(label_to_index) # 名字与标号一一对应， 例如1 dog，2 cat

index_to_label = dict((v, k) for k, v in label_to_index.items()) # 标号-名字

all_labels = [label_to_index.get(name) for name in all_labels_name] # 将标签名字全部转化为数字标号

np.random.seed()
random_index = np.random.permutation(len(imags_path)) # 随机标号

imags_path = np.array(imags_path)[random_index] # 图片路径和标签用同一随机数打乱
all_labels = np.array(all_labels)[random_index]

i = int(len(imags_path)*0.8) # 取80%为训练集，20%为测试集
train_path = imags_path[ :i]
train_labels = all_labels[ :i]
test_path = imags_path[i: ]
test_labels = all_labels[i: ]

train_ds = tf.data.Dataset.from_tensor_slices((train_path, train_labels)) # 数据保存为dataset
test_ds = tf.data.Dataset.from_tensor_slices((test_path, test_labels))

# 加载图片，进行预处理
def load_img(path, label):
    image = tf.io.read_file(path)
    image = tf.image.decode_jpeg(image, channels=3)
    image = tf.image.resize(image, [128, 128])
    image = tf.cast(image, tf.float32)
    image = image/255
    return image, label

AUTOYUNE = tf.data.experimental.AUTOTUNE
train_ds = train_ds.map(load_img, num_parallel_calls=AUTOYUNE)
test_ds = test_ds.map(load_img, num_parallel_calls=AUTOYUNE)

BATCH_SIZE = 8
train_ds = train_ds.repeat().shuffle(200).batch(BATCH_SIZE) # 打乱数据
test_ds = test_ds.batch(BATCH_SIZE)

网络

使用Sequential模型, 使用语法：model = tf.keras.Sequential()

添加网络层：

输入层：model.add(tf.keras.layers.Conv2D(64, (3, 3), input_shape=(128, 128, 3), activation='relu'))
Conv2D函数参数

添加层：model.add(tf.keras.layers.BatchNormalization())
BatchNormalization作用

继续添加层，增加网络深度，并添加池化层
池化层：model.add(tf.keras.layers.MaxPooling2D())

最后添加输出层:model.add(tf.keras.layers.Dense(5, activation='softmax'))
因为该网络为5分类，所以Dense层输出5维度
Dense参数

添加全局池化层：model.add(tf.keras.layers.GlobalAveragePooling2D())
GlobalAveragePooling2D

池化层

代码

model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv2D(64, (3, 3), input_shape=(128, 128, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(64, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(128, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(128, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())

model.add(tf.keras.layers.Conv2D(256, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(256, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())

model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())

model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.GlobalAveragePooling2D())

model.add(tf.keras.layers.Dense(1024, activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Dense(5, activation='softmax'))

网络编译

compile参数

model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
              loss=tf.keras.losses.SparseCategoricalCrossentropy(),
              metrics=['acc']
              )

训练

fit参数

train_count = len(train_path)
test_count = len(test_path)
step_per_epoch = train_count//BATCH_SIZE
validation_step = test_count//BATCH_SIZE

history = model.fit(train_ds, epochs=500, steps_per_epoch=step_per_epoch,
                        validation_data=test_ds,
                        validation_steps=validation_step
                        )

模型保存

模型保存
model.save("保存路径")

训练效果

训练集损失函数loss（下），测试集损失函数val_loss

训练集正确率acc，测试集正确率val_acc

测试结果最好有70%的准确率

完整代码

import tensorflow as tf
import numpy as np
import glob
import matplotlib.pyplot as plt
import os

tf.compat.v1.disable_eager_execution() //版本兼容问题，不报错可去掉

gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
    tf.config.experimental.set_virtual_device_configuration(gpus[0], [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=4500)])
    # 设置GPU显存

os.environ["CUDA_VISIBLE_DEVICES"] = "0" # 使用0号GPU进行计算

imags_path = glob.glob('./data/animals/*/*.jpg') # 读取所有图片路径


all_labels_name = [imag_p.split("\\")[1] for imag_p in imags_path] # 获取标签名

label_names = np.unique(all_labels_name) # 去除重复的

label_to_index = dict((name, i) for i, name in enumerate(label_names))
print(label_to_index) # 名字与标号一一对应， 例如1 dog，2 cat

index_to_label = dict((v, k) for k, v in label_to_index.items()) # 标号-名字

all_labels = [label_to_index.get(name) for name in all_labels_name] # 将标签名字全部转化为数字标号

np.random.seed()
random_index = np.random.permutation(len(imags_path)) # 随机标号

imags_path = np.array(imags_path)[random_index] # 图片路径和标签用同一随机数打乱
all_labels = np.array(all_labels)[random_index]


i = int(len(imags_path)*0.8) # 取80%为训练集，20%为测试集
train_path = imags_path[ :i]
train_labels = all_labels[ :i]
test_path = imags_path[i: ]
test_labels = all_labels[i: ]

train_ds = tf.data.Dataset.from_tensor_slices((train_path, train_labels)) # 数据保存为dataset
test_ds = tf.data.Dataset.from_tensor_slices((test_path, test_labels))

# 加载图片，进行预处理
def load_img(path, label):
    image = tf.io.read_file(path)
    image = tf.image.decode_jpeg(image, channels=3)
    image = tf.image.resize(image, [128, 128])
    image = tf.cast(image, tf.float32)
    image = image/255
    return image, label


AUTOYUNE = tf.data.experimental.AUTOTUNE
train_ds = train_ds.map(load_img, num_parallel_calls=AUTOYUNE)
test_ds = test_ds.map(load_img, num_parallel_calls=AUTOYUNE)

BATCH_SIZE = 8
train_ds = train_ds.repeat().shuffle(200).batch(BATCH_SIZE) # 打乱数据
test_ds = test_ds.batch(BATCH_SIZE)

# 模型
model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv2D(64, (3, 3), input_shape=(128, 128, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(64, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(128, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(128, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())
model.add(tf.keras.layers.Conv2D(256, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(256, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.GlobalAveragePooling2D())
model.add(tf.keras.layers.Dense(1024, activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Dense(5, activation='softmax'))

model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
              loss=tf.keras.losses.SparseCategoricalCrossentropy(),
              metrics=['acc']
              )

train_count = len(train_path)
test_count = len(test_path)
step_per_epoch = train_count//BATCH_SIZE
validation_step = test_count//BATCH_SIZE

history = model.fit(train_ds, epochs=500, steps_per_epoch=step_per_epoch,
                        validation_data=test_ds,
                        validation_steps=validation_step
                        )


plt.figure()
plt.plot(history.epoch, history.history.get('loss'), label='loss')
plt.plot(history.epoch, history.history.get('val_loss'), label='val_loss')
plt.show()
plt.figure()
plt.plot(history.epoch, history.history.get('acc'), label='acc')
plt.plot(history.epoch, history.history.get('val_acc'), label='val_acc')
plt.show()

参考：
B站：日月光华

参考：
B站：[日月光华](https://space.bilibili.com/387143299?spm_id_from=333.788.b_765f7570696e666f.1)