Ming-Log's Blog

记录生活中的点点滴滴

0%

Python数据科学_27_案例：猫狗大战【计算机视觉】

发表于 更新于 分类于
本文字数： 31k 阅读时长 ≈ 28 分钟

项目环境介绍

本项目使用的深度学习框架为Tensorflow2，详细环境信息如下：

  • Python=3.8.16
  • Tensorflow=2.9.1
  • keras=2.10.0
  • cudatoolkit=11.3.1
  • cudnn=8.2.1

为了防止因数据集过大导致显存过载，首先配置一下Tensorflow调用GPU的规则，设置为动态显存申请。由于这个配置项必须在代码的最前方声明，故提前说明。

在TensorFlow中，GPU内存默认是一次性分配的，这意味着如果模型占用的内存超过可用内存的限制，将无法运行模型，而会出现OOM（Out Of Memory）错误。为了解决这个问题，TensorFlow提供了函数set_memory_growth，它可以让TensorFlow动态分配GPU内存，只使用所需的GPU内存。总之，使用set_memory_growth函数，可以在程序运行时分配所需的GPU内存，而不是在程序启动时将GPU内存分配给TensorFlow，这样可以避免在运行大型模型时出现内存不足的问题。

1
2
3
from tensorflow.config.experimental import list_physical_devices, set_memory_growth
physical_devices = list_physical_devices('GPU')
set_memory_growth(physical_devices[0], True)

项目背景与目标

项目背景

众所周知，猫和狗之间有很多的区别。

  • 从所属科目来看，猫属于猫科动物，狗属于犬科动物；
  • 从体型来看，一般的狗的体型要比猫大；
  • 从叫声来看，猫发出的叫声是喵喵~，而狗发出的叫声是汪汪~

等等，正是因为猫和狗之前的诸多不同，我们人类很容易就能将其区别开，那么我们能否让电脑去捕获这些特征，让电脑识别出猫和狗呢？

image-20230518190727287

项目目标

利用所提供的猫和狗的图片数据，设计并训练出一个能够根据图片区分出猫和狗的模型。

数据：数据集下载

  • 训练集：10000（猫）+10000（狗）
  • 验证集：1250（猫）+1250（狗）
  • 测试集：2500（猫+狗）

image-20230518190756069

数据预处理

根据对原始数据的观察，我们需要做以下预处理操作。

  1. 图像尺寸统一，将所有图片的尺寸统一为(150, 150)：这是由于神经网络在进行数据输入时，必须要求输入的数据是统一的。
  2. 像素归一化：神经网络训练时一般使用较小的权重值来进行拟合，当训练数据的值是较大整数值时，可能对模型的训练速度产生一定的影响，解决办法是将其进行像素值归一化处理。
  3. 图像增强：数据增强是深度学习中非常重要的一部分，它可以扩大数据集，提高模型的鲁棒性和泛化能力，降低模型过拟合的风险，进而提高模型的性能和效果。
    • 扩大数据集：数据增强技术通过改变原始图像，如旋转、裁剪、翻转等，可以产生一些新的图像，从而扩大数据集。这对于数据集较小的情况下特别有效，可以防止模型过拟合，提高模型的泛化能力。
    • 增强模型的鲁棒性：数据增强可以增加数据的多样性，使得模型更具有鲁棒性，即对于一定程度的图像扭曲、遮挡、噪声等干扰仍能产生较为准确的预测结果。
    • 降低模型过拟合：过度依赖训练数据样本可能导致过拟合，即在训练数据上表现很好，但是在测试数据上表现很差。通过数据增强，可以生成更多样本，在一定程度上减少过拟合现象，提高模型的泛化能力。
    • 提高模型的性能：数据增强可以让模型使用更多的信息来进行学习，使得模型训练得更充分，进而提高模型的性能。

数据增强

数据增强又分为在线数据增强和离线数据增强。

  1. 在线数据增强通常是在模型训练时动态地对数据进行增强，比如对图像进行旋转、缩放、平移、裁剪、翻转、加噪声等操作，从而增加数据的多样性。在线数据增强的优点是能够在模型训练过程中实时地增加训练数据，但也会造成额外的计算开销和内存占用。
  2. 离线数据增强则是在训练前对原始数据进行扩充，比如生成新的图像、音频、文本等，从而增加数据的数量和多样性。离线数据增强的优点是可以节省训练时间和计算资源，但需要额外的数据生成和存储工作。

在本项目中，由于训练使用的数据集数据量较大，如果采用离线数据增强的方式，会生成大量新图片数据存储到本地，造成时间和存储空间的浪费。

故在本项目中使用在线数据增强的方式进行处理，在Keras中存在对应的在线数据增强生成器from tensorflow.keras.preprocessing.image import ImageDataGenerator。该生成器对象可以在实例化的时候定义数据增强的方式。

例如：

  • rotation_range：整数，数据提升时图片随机转动的角度。随机选择图片的角度，是一个0~180的度数，取值为0~180。
  • width_shift_range：浮点数，图片宽度的某个比例，数据提升时图片随机水平偏移的幅度
  • height_shift_range：浮点数，图片高度的某个比例，数据提升时图片随机竖直偏移的幅度。
  • horizontal_flip：布尔值，进行随机水平翻转。随机的对图片进行水平翻转，这个参数适用于水平翻转不影响图片语义的时候。
  • vertical_flip：布尔值，进行随机竖直翻转。
  • rescale: 值将在执行其他处理前乘到整个图像上，我们的图像在RGB通道都是0~255的整数，这样的操作可能使图像的值过高或过低，所以我们将这个值定为0~1之间的数。

这里仅列出一小部分，详见Keras官方文档

定义相关路径和参数

1
2
3
4
# 定义所有图片所在路径
train_path = ''../data/dogVScat_data/train/'
val_path = ''../data/dogVScat_data/val/'
test_path = ''../data/dogVScat_data/test/'
1
2
3
4
5
6
# 定义图片的统一尺寸大小
size = (150, 150)
# 设置训练数据的batch_size
trian_batch_size = 32
# 设置验证集数据的batch_size
val_batch_size = 8

定义图像生成器

1
2
from tensorflow import keras
from tensorflow.keras.preprocessing.image import ImageDataGenerator
1
img_generator = ImageDataGenerator(rescale=1/255)  # 对原始像素做像素值归一化

使用图像生成器导入数据

在使用图像生成器时，有一个前提，必须要求所有的图片在一个大文件夹中，大文件夹中有多个子文件夹，用来存储不同类别的图片数据。

例如：

image-20230518192040802

定义训练集生成器

1
2
3
4
5
6
7
train_generator = img_generator.flow_from_directory(
    directory=train_path,  # 图片路径
    target_size=size,  # 图片生成的尺寸
    class_mode='binary',  # 二分类任务
    batch_size = trian_batch_size,  # 生成器一次生成多少个图片
    shuffle=True
)
Found 20000 images belonging to 2 classes.

定义验证集生成器

1
2
3
4
5
6
7
8
val_generator = img_generator.flow_from_directory(
    directory=val_path,  # 图片路径
    target_size=size,  # 图片生成的尺寸
    class_mode='binary',  # 二分类任务
    batch_size = val_batch_size,  # 生成器一次生成多少个图片
    shuffle=True
)

Found 2500 images belonging to 2 classes.

展示一个生成器的生成结果

1
res = next(train_generator)
1
2
3
# 取出图片数据
tmp_img = res[0]
tmp_img.shape
(32, 150, 150, 3)

1
2
3
# 取出标签数据
tmp_label = res[1]
tmp_label
array([0., 1., 0., 0., 1., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 1., 0.,
       0., 1., 1., 0., 0., 1., 0., 0., 0., 0., 0., 1., 0., 1., 0.],
      dtype=float32)

1
2
3
4
5
6
7
plt.figure(figsize=(10, 10), dpi=500)
for i in range(len(tmp_label[:16])):
    ax = plt.subplot(4, 4, i+1)
    ax.set_title(label_dic[tmp_label[i]])
    ax.axis('off')
    ax.imshow(tmp_img[i])
plt.show()

image-20230518192144636

模型搭建

在本项目中我们选用的模型为经典的LeNet-5神经网络模型，模型结构如下图所示：

image-20230518192200454

1
2
3
4
5
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras import losses
from tensorflow.keras import metrics
from tensorflow.keras import optimizers
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# 搭建LeNet5神经网络框架
model = keras.Sequential(name='LeNet5')
# 卷积层
# filters: 卷积核的个数
# kernel_size: 卷积核的大小
# striders: 卷积核移动的步长
# input_shape: 输入到模型中的数据的尺寸，这个只在网络的第一层中指定
# activation: 非线性激活函数
model.add(keras.layers.Conv2D(filters=6, kernel_size=(5, 5), strides=(1, 1), input_shape=(150, 150, 3), activation='relu'))
# 最大池化层
model.add(keras.layers.MaxPool2D())
# 卷积层
model.add(keras.layers.Conv2D(filters=16, kernel_size=(5, 5), strides=(1, 1), activation='relu'))
# 最大池化层
model.add(keras.layers.MaxPool2D())
# 展平层
model.add(keras.layers.Flatten())
# 全连接层
# units: 全连接层的神经元数量
model.add(keras.layers.Dense(units=120, activation='relu'))
# 全连接层
model.add(keras.layers.Dense(units=84, activation='relu'))
# 全连接层
model.add(keras.layers.Dense(units=1, activation='sigmoid'))

模型的编译

1
model.compile(optimizer='adam', loss=losses.binary_crossentropy, metrics=metrics.binary_accuracy)

模型的训练

1
history = model.fit(train_generator, steps_per_epoch=100, epochs=10, verbose=1, validation_data=val_generator, validation_steps=8)
Epoch 1/10
100/100 [==============================] - 5s 42ms/step - loss: 7.4331 - binary_accuracy: 0.5678 - val_loss: 0.6062 - val_binary_accuracy: 0.7656
Epoch 2/10
100/100 [==============================] - 4s 40ms/step - loss: 0.6712 - binary_accuracy: 0.6006 - val_loss: 0.6717 - val_binary_accuracy: 0.6562
Epoch 3/10
100/100 [==============================] - 4s 40ms/step - loss: 0.6344 - binary_accuracy: 0.6606 - val_loss: 0.6986 - val_binary_accuracy: 0.5156
Epoch 4/10
100/100 [==============================] - 4s 40ms/step - loss: 0.6029 - binary_accuracy: 0.6800 - val_loss: 0.7661 - val_binary_accuracy: 0.5781
Epoch 5/10
100/100 [==============================] - 4s 39ms/step - loss: 0.6031 - binary_accuracy: 0.6856 - val_loss: 0.5327 - val_binary_accuracy: 0.7188
Epoch 6/10
100/100 [==============================] - 4s 40ms/step - loss: 0.5514 - binary_accuracy: 0.7309 - val_loss: 0.5909 - val_binary_accuracy: 0.6719
Epoch 7/10
100/100 [==============================] - 4s 39ms/step - loss: 0.5325 - binary_accuracy: 0.7428 - val_loss: 0.5306 - val_binary_accuracy: 0.6406
Epoch 8/10
100/100 [==============================] - 4s 42ms/step - loss: 0.4777 - binary_accuracy: 0.7763 - val_loss: 0.7278 - val_binary_accuracy: 0.6562
Epoch 9/10
100/100 [==============================] - 4s 41ms/step - loss: 0.4717 - binary_accuracy: 0.7809 - val_loss: 0.7141 - val_binary_accuracy: 0.6250
Epoch 10/10
100/100 [==============================] - 4s 41ms/step - loss: 0.4344 - binary_accuracy: 0.8081 - val_loss: 0.6835 - val_binary_accuracy: 0.7188

设置回调函数

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
# 只保存最佳模型
model_check_point = keras.callbacks.ModelCheckpoint(
    filepath='best_model.h5',   # 模型的保存地址
    monitor='val_binary_accuracy',   # 检测的指标
    verbose=1,   # 在控制台中打印出检测过程
    save_best_only=True,  # 仅保存最优模型
    save_weights_only=False,   # 是否仅保存权重
    mode='auto')   # 设置检测参数最优的方向
    
# 提前结束训练函数
early_stopping = keras.callbacks.EarlyStopping(
    monitor='val_binary_accuracy',  # 检测的指标
    min_delta=0,  # 设置最小增量
    patience=10,  # 为改善的时间数
    verbose=1, 
    mode='auto')

# 设置权重衰减
from keras import backend as K
# 每训练5轮学习速率下降10%
def scheduler(epoch):
    if epoch % 5 == 0 and epoch != 0:
        lr = K.get_value(model.optimizer.lr)
        K.set_value(model.optimizer.lr, lr * 0.9)
        print("lr changed to {}".format(lr * 0.9))
    return K.get_value(model.optimizer.lr)
learning_rate_scheduler = keras.callbacks.LearningRateScheduler(scheduler, verbose=1)

# 保存训练日志
csv_logger = keras.callbacks.CSVLogger('training.log', separator=',', append=False)

callbacks = [model_check_point, early_stopping, learning_rate_scheduler, csv_logger]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# 搭建LeNet5神经网络框架
model = keras.Sequential(name='LeNet5')
# 卷积层
# filters: 卷积核的个数
# kernel_size: 卷积核的大小
# striders: 卷积核移动的步长
# input_shape: 输入到模型中的数据的尺寸，这个只在网络的第一层中指定
# activation: 非线性激活函数
model.add(keras.layers.Conv2D(filters=6, kernel_size=(5, 5), strides=(1, 1), input_shape=(150, 150, 3), activation='relu'))
# 最大池化层
model.add(keras.layers.MaxPool2D())
# 卷积层
model.add(keras.layers.Conv2D(filters=16, kernel_size=(5, 5), strides=(1, 1), activation='relu'))
# 最大池化层
model.add(keras.layers.MaxPool2D())
# 展平层
model.add(keras.layers.Flatten())
# 全连接层
# units: 全连接层的神经元数量
model.add(keras.layers.Dense(units=120, activation='relu'))
# 全连接层
model.add(keras.layers.Dense(units=84, activation='relu'))
# 全连接层
model.add(keras.layers.Dense(units=1, activation='sigmoid'))
1
model.compile(optimizer='adam', loss=losses.binary_crossentropy, metrics=metrics.binary_accuracy)
1
history = model.fit(train_generator, steps_per_epoch=100, epochs=100, verbose=1, validation_data=val_generator, validation_steps=8, callbacks=callbacks)
Epoch 1: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 1/100
 99/100 [============================>.] - ETA: 0s - loss: 0.6766 - binary_accuracy: 0.5732
Epoch 1: val_binary_accuracy improved from -inf to 0.68750, saving model to best_model.h5
100/100 [==============================] - 5s 46ms/step - loss: 0.6754 - binary_accuracy: 0.5756 - val_loss: 0.5892 - val_binary_accuracy: 0.6875 - lr: 0.0010

Epoch 2: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 2/100
100/100 [==============================] - ETA: 0s - loss: 0.6036 - binary_accuracy: 0.6597
Epoch 2: val_binary_accuracy improved from 0.68750 to 0.75000, saving model to best_model.h5
100/100 [==============================] - 5s 45ms/step - loss: 0.6036 - binary_accuracy: 0.6597 - val_loss: 0.5050 - val_binary_accuracy: 0.7500 - lr: 0.0010

Epoch 3: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 3/100
 99/100 [============================>.] - ETA: 0s - loss: 0.5689 - binary_accuracy: 0.6995
Epoch 3: val_binary_accuracy did not improve from 0.75000
100/100 [==============================] - 4s 44ms/step - loss: 0.5700 - binary_accuracy: 0.6988 - val_loss: 0.6376 - val_binary_accuracy: 0.6875 - lr: 0.0010

Epoch 4: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 4/100
100/100 [==============================] - ETA: 0s - loss: 0.5661 - binary_accuracy: 0.7138
Epoch 4: val_binary_accuracy improved from 0.75000 to 0.76562, saving model to best_model.h5
100/100 [==============================] - 4s 43ms/step - loss: 0.5661 - binary_accuracy: 0.7138 - val_loss: 0.4945 - val_binary_accuracy: 0.7656 - lr: 0.0010

Epoch 5: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 5/100
 99/100 [============================>.] - ETA: 0s - loss: 0.5255 - binary_accuracy: 0.7421
Epoch 5: val_binary_accuracy improved from 0.76562 to 0.81250, saving model to best_model.h5
100/100 [==============================] - 4s 43ms/step - loss: 0.5253 - binary_accuracy: 0.7422 - val_loss: 0.4567 - val_binary_accuracy: 0.8125 - lr: 0.0010
lr changed to 0.0009000000427477062

Epoch 6: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 6/100
 99/100 [============================>.] - ETA: 0s - loss: 0.4889 - binary_accuracy: 0.7620
Epoch 6: val_binary_accuracy improved from 0.81250 to 0.82812, saving model to best_model.h5
100/100 [==============================] - 4s 42ms/step - loss: 0.4880 - binary_accuracy: 0.7622 - val_loss: 0.4246 - val_binary_accuracy: 0.8281 - lr: 9.0000e-04

Epoch 7: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 7/100
 99/100 [============================>.] - ETA: 0s - loss: 0.4685 - binary_accuracy: 0.7743
Epoch 7: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 41ms/step - loss: 0.4685 - binary_accuracy: 0.7744 - val_loss: 0.4741 - val_binary_accuracy: 0.7500 - lr: 9.0000e-04

Epoch 8: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 8/100
 99/100 [============================>.] - ETA: 0s - loss: 0.4374 - binary_accuracy: 0.7989
Epoch 8: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 42ms/step - loss: 0.4376 - binary_accuracy: 0.7987 - val_loss: 0.5664 - val_binary_accuracy: 0.6406 - lr: 9.0000e-04

Epoch 9: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 9/100
 99/100 [============================>.] - ETA: 0s - loss: 0.4539 - binary_accuracy: 0.7926
Epoch 9: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 41ms/step - loss: 0.4530 - binary_accuracy: 0.7934 - val_loss: 0.5788 - val_binary_accuracy: 0.7188 - lr: 9.0000e-04

Epoch 10: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 10/100
 99/100 [============================>.] - ETA: 0s - loss: 0.4373 - binary_accuracy: 0.8040
Epoch 10: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 41ms/step - loss: 0.4379 - binary_accuracy: 0.8034 - val_loss: 0.5112 - val_binary_accuracy: 0.7969 - lr: 9.0000e-04
lr changed to 0.0008100000384729356

Epoch 11: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 11/100
 99/100 [============================>.] - ETA: 0s - loss: 0.4136 - binary_accuracy: 0.8122
Epoch 11: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 41ms/step - loss: 0.4139 - binary_accuracy: 0.8128 - val_loss: 0.4942 - val_binary_accuracy: 0.7969 - lr: 8.1000e-04

Epoch 12: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 12/100
 99/100 [============================>.] - ETA: 0s - loss: 0.3779 - binary_accuracy: 0.8330
Epoch 12: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 42ms/step - loss: 0.3782 - binary_accuracy: 0.8331 - val_loss: 0.4771 - val_binary_accuracy: 0.8125 - lr: 8.1000e-04

Epoch 13: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 13/100
 99/100 [============================>.] - ETA: 0s - loss: 0.3853 - binary_accuracy: 0.8314
Epoch 13: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 41ms/step - loss: 0.3836 - binary_accuracy: 0.8325 - val_loss: 0.6342 - val_binary_accuracy: 0.6719 - lr: 8.1000e-04

Epoch 14: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 14/100
 99/100 [============================>.] - ETA: 0s - loss: 0.3323 - binary_accuracy: 0.8573
Epoch 14: val_binary_accuracy did not improve from 0.82812
100/100 [==============================] - 4s 42ms/step - loss: 0.3314 - binary_accuracy: 0.8575 - val_loss: 0.4888 - val_binary_accuracy: 0.7656 - lr: 8.1000e-04

Epoch 15: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 15/100
100/100 [==============================] - ETA: 0s - loss: 0.3395 - binary_accuracy: 0.8534
Epoch 15: val_binary_accuracy improved from 0.82812 to 0.84375, saving model to best_model.h5
100/100 [==============================] - 4s 42ms/step - loss: 0.3395 - binary_accuracy: 0.8534 - val_loss: 0.4309 - val_binary_accuracy: 0.8438 - lr: 8.1000e-04
lr changed to 0.0007290000503417104

Epoch 16: LearningRateScheduler setting learning rate to 0.0007290000794455409.
Epoch 16/100
 99/100 [============================>.] - ETA: 0s - loss: 0.3116 - binary_accuracy: 0.8753
Epoch 16: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 41ms/step - loss: 0.3119 - binary_accuracy: 0.8756 - val_loss: 0.3444 - val_binary_accuracy: 0.8281 - lr: 7.2900e-04

Epoch 17: LearningRateScheduler setting learning rate to 0.0007290000794455409.
Epoch 17/100
 99/100 [============================>.] - ETA: 0s - loss: 0.2847 - binary_accuracy: 0.8842
Epoch 17: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 41ms/step - loss: 0.2850 - binary_accuracy: 0.8834 - val_loss: 0.4776 - val_binary_accuracy: 0.8125 - lr: 7.2900e-04

Epoch 18: LearningRateScheduler setting learning rate to 0.0007290000794455409.
Epoch 18/100
 99/100 [============================>.] - ETA: 0s - loss: 0.2925 - binary_accuracy: 0.8810
Epoch 18: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 41ms/step - loss: 0.2925 - binary_accuracy: 0.8813 - val_loss: 0.3710 - val_binary_accuracy: 0.8125 - lr: 7.2900e-04

Epoch 19: LearningRateScheduler setting learning rate to 0.0007290000794455409.
Epoch 19/100
 99/100 [============================>.] - ETA: 0s - loss: 0.2544 - binary_accuracy: 0.9018
Epoch 19: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 41ms/step - loss: 0.2542 - binary_accuracy: 0.9016 - val_loss: 0.4350 - val_binary_accuracy: 0.8125 - lr: 7.2900e-04

Epoch 20: LearningRateScheduler setting learning rate to 0.0007290000794455409.
Epoch 20/100
100/100 [==============================] - ETA: 0s - loss: 0.2393 - binary_accuracy: 0.9038
Epoch 20: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 41ms/step - loss: 0.2393 - binary_accuracy: 0.9038 - val_loss: 0.8589 - val_binary_accuracy: 0.6875 - lr: 7.2900e-04
lr changed to 0.0006561000715009868

Epoch 21: LearningRateScheduler setting learning rate to 0.0006561000482179224.
Epoch 21/100
 99/100 [============================>.] - ETA: 0s - loss: 0.2071 - binary_accuracy: 0.9192
Epoch 21: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 41ms/step - loss: 0.2074 - binary_accuracy: 0.9187 - val_loss: 0.4638 - val_binary_accuracy: 0.7969 - lr: 6.5610e-04

Epoch 22: LearningRateScheduler setting learning rate to 0.0006561000482179224.
Epoch 22/100
 99/100 [============================>.] - ETA: 0s - loss: 0.1757 - binary_accuracy: 0.9343
Epoch 22: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 40ms/step - loss: 0.1748 - binary_accuracy: 0.9347 - val_loss: 0.5891 - val_binary_accuracy: 0.7500 - lr: 6.5610e-04

Epoch 23: LearningRateScheduler setting learning rate to 0.0006561000482179224.
Epoch 23/100
 99/100 [============================>.] - ETA: 0s - loss: 0.1737 - binary_accuracy: 0.9378
Epoch 23: val_binary_accuracy did not improve from 0.84375
100/100 [==============================] - 4s 41ms/step - loss: 0.1728 - binary_accuracy: 0.9381 - val_loss: 0.6565 - val_binary_accuracy: 0.7500 - lr: 6.5610e-04

Epoch 24: LearningRateScheduler setting learning rate to 0.0006561000482179224.
Epoch 24/100
 99/100 [============================>.] - ETA: 0s - loss: 0.1539 - binary_accuracy: 0.9489
Epoch 24: val_binary_accuracy improved from 0.84375 to 0.85938, saving model to best_model.h5
100/100 [==============================] - 4s 43ms/step - loss: 0.1549 - binary_accuracy: 0.9484 - val_loss: 0.3733 - val_binary_accuracy: 0.8594 - lr: 6.5610e-04

Epoch 25: LearningRateScheduler setting learning rate to 0.0006561000482179224.
Epoch 25/100
 99/100 [============================>.] - ETA: 0s - loss: 0.1448 - binary_accuracy: 0.9508
Epoch 25: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.1455 - binary_accuracy: 0.9506 - val_loss: 0.6035 - val_binary_accuracy: 0.7656 - lr: 6.5610e-04
lr changed to 0.0005904900433961303

Epoch 26: LearningRateScheduler setting learning rate to 0.0005904900608584285.
Epoch 26/100
 99/100 [============================>.] - ETA: 0s - loss: 0.1366 - binary_accuracy: 0.9542
Epoch 26: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.1393 - binary_accuracy: 0.9541 - val_loss: 0.4766 - val_binary_accuracy: 0.8281 - lr: 5.9049e-04

Epoch 27: LearningRateScheduler setting learning rate to 0.0005904900608584285.
Epoch 27/100
 99/100 [============================>.] - ETA: 0s - loss: 0.1207 - binary_accuracy: 0.9596
Epoch 27: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.1211 - binary_accuracy: 0.9591 - val_loss: 0.9246 - val_binary_accuracy: 0.7500 - lr: 5.9049e-04

Epoch 28: LearningRateScheduler setting learning rate to 0.0005904900608584285.
Epoch 28/100
100/100 [==============================] - ETA: 0s - loss: 0.1131 - binary_accuracy: 0.9625
Epoch 28: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 42ms/step - loss: 0.1131 - binary_accuracy: 0.9625 - val_loss: 1.0707 - val_binary_accuracy: 0.7188 - lr: 5.9049e-04

Epoch 29: LearningRateScheduler setting learning rate to 0.0005904900608584285.
Epoch 29/100
 99/100 [============================>.] - ETA: 0s - loss: 0.0898 - binary_accuracy: 0.9719
Epoch 29: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.0893 - binary_accuracy: 0.9722 - val_loss: 0.8163 - val_binary_accuracy: 0.7656 - lr: 5.9049e-04

Epoch 30: LearningRateScheduler setting learning rate to 0.0005904900608584285.
Epoch 30/100
 99/100 [============================>.] - ETA: 0s - loss: 0.1082 - binary_accuracy: 0.9624
Epoch 30: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.1075 - binary_accuracy: 0.9628 - val_loss: 0.7376 - val_binary_accuracy: 0.7969 - lr: 5.9049e-04
lr changed to 0.0005314410547725857

Epoch 31: LearningRateScheduler setting learning rate to 0.0005314410664141178.
Epoch 31/100
 99/100 [============================>.] - ETA: 0s - loss: 0.0848 - binary_accuracy: 0.9722
Epoch 31: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 41ms/step - loss: 0.0847 - binary_accuracy: 0.9722 - val_loss: 0.6468 - val_binary_accuracy: 0.7656 - lr: 5.3144e-04

Epoch 32: LearningRateScheduler setting learning rate to 0.0005314410664141178.
Epoch 32/100
 99/100 [============================>.] - ETA: 0s - loss: 0.0616 - binary_accuracy: 0.9804
Epoch 32: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.0640 - binary_accuracy: 0.9797 - val_loss: 0.8088 - val_binary_accuracy: 0.7344 - lr: 5.3144e-04

Epoch 33: LearningRateScheduler setting learning rate to 0.0005314410664141178.
Epoch 33/100
 99/100 [============================>.] - ETA: 0s - loss: 0.0685 - binary_accuracy: 0.9811
Epoch 33: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.0706 - binary_accuracy: 0.9806 - val_loss: 0.9013 - val_binary_accuracy: 0.7969 - lr: 5.3144e-04

Epoch 34: LearningRateScheduler setting learning rate to 0.0005314410664141178.
Epoch 34/100
 99/100 [============================>.] - ETA: 0s - loss: 0.0432 - binary_accuracy: 0.9858
Epoch 34: val_binary_accuracy did not improve from 0.85938
100/100 [==============================] - 4s 40ms/step - loss: 0.0432 - binary_accuracy: 0.9859 - val_loss: 0.8229 - val_binary_accuracy: 0.7188 - lr: 5.3144e-04
Epoch 34: early stopping

引入回调函数，模型一共训练了34轮，最优模型的精度为85.938%

接下来使用该模型对测试集进行预测并将预测结果可视化。

1
2
3
from tensorflow import keras
import cv2
import os
1
best_model = keras.models.load_model('best_model.h5')
1
2
3
4
5
6
7
8
9
# 在测试集种随机读取16张图片进行预测
all_test_imgs = os.listdir(test_path)
index_img = np.random.choice(range(len(all_test_imgs)), 16, replace=False)
imgs = []
for i in index_img:
    img = cv2.imread(test_path + all_test_imgs[i])
    img = cv2.resize(img, (150, 150))
    imgs.append(img)
imgs = np.array(imgs)
1
2
3
y_pred = best_model.predict(imgs)
y_pred = [1 if i > 0.5 else 0 for i in y_pred]
y_pred
1/1 [==============================] - 1s 583ms/step

1
import matplotlib.pyplot as plt
1
label_dic = {0:'cat', 1:'dog'}
1
2
3
4
5
6
7
plt.figure(figsize=(10, 10))
for i in range(len(y_pred)):
    ax = plt.subplot(4, 4, i+1)
    ax.set_title(label_dic[y_pred[i]])
    ax.axis('off')
    ax.imshow(imgs[i][:, :, ::-1])
plt.show()

image-20230518185039790

从可视化结果可以看出，模型的效果不好，16张图片预测错误了4张。

使用迁移学习去改善模型训练效果

1
2
3
from tensorflow.keras.applications import resnet50
# include_top：是否包含全连接层
conv_base = resnet50.ResNet50(weights='imagenet', include_top=False, input_shape=(150, 150, 3))
1
2
3
4
5
6
7
8
for layers in conv_base.layers[:]:
    layers.trainable = False  # 冻结resnet50所有的参数，不参与模型参数更新

x = conv_base.output
x = keras.layers.GlobalAveragePooling2D()(x)
x = keras.layers.Dense(256, activation='relu')(x)
predictions = keras.layers.Dense(1, activation='sigmoid')(x)
model = keras.Model(inputs=conv_base.input, outputs=predictions)
1
2
3
# optimizer = keras.optimizers.RMSprop(lr=1e-3)

model.compile(optimizer='adam', loss=losses.binary_crossentropy, metrics=metrics.binary_accuracy)

重新定义回调函数

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
# 只保存最佳模型
model_check_point = keras.callbacks.ModelCheckpoint(
    filepath='best_model.h5',   # 模型的保存地址
    monitor='val_binary_accuracy',   # 检测的指标
    verbose=1,   # 在控制台中打印出检测过程
    save_best_only=True,  # 仅保存最优模型
    save_weights_only=False,   # 是否仅保存权重
    mode='auto')   # 设置检测参数最优的方向
    
# 提前结束训练函数
early_stopping = keras.callbacks.EarlyStopping(
    monitor='val_binary_accuracy',  # 检测的指标
    min_delta=0,  # 设置最小增量
    patience=10,  # 为改善的时间数
    verbose=1, 
    mode='auto')

# 设置权重衰减
from keras import backend as K
# 每训练5轮学习速率下降10%
def scheduler(epoch):
    if epoch % 5 == 0 and epoch != 0:
        lr = K.get_value(model.optimizer.lr)
        K.set_value(model.optimizer.lr, lr * 0.9)
        print("lr changed to {}".format(lr * 0.9))
    return K.get_value(model.optimizer.lr)
learning_rate_scheduler = keras.callbacks.LearningRateScheduler(scheduler, verbose=1)

# 保存训练日志
csv_logger = keras.callbacks.CSVLogger('training.log', separator=',', append=False)

callbacks = [model_check_point, early_stopping, learning_rate_scheduler, csv_logger]

定义生成器

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
img_generator = ImageDataGenerator(preprocessing_function=keras.applications.resnet.preprocess_input)  # 定义图片生成器
# train_img_path: 训练数据地址，会自动加载其文件夹下面所有的图片
# 包括子文件夹下的图片，并且每个子文件夹下面存放一类数据
train_generator = img_generator.flow_from_directory(
    train_path, 
    target_size=size, 
    batch_size=16, 
    class_mode='binary', 
    shuffle=True)  # binary 2分类，一维标签
val_generator = img_generator.flow_from_directory(
    val_path, 
    target_size=size, 
    batch_size=8, 
    class_mode='binary', 
    shuffle=True)  # binary 2分类，一维标签
Found 20000 images belonging to 2 classes.
Found 2500 images belonging to 2 classes.

1
2
3
4
5
6
7
8
history = model.fit(
      train_generator,
      steps_per_epoch=100,
      epochs=100,
      verbose=1,
      validation_data=val_generator,
      validation_steps=8,
      callbacks=callbacks)
Epoch 1: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 1/100
100/100 [==============================] - ETA: 0s - loss: 0.2755 - binary_accuracy: 0.9044
Epoch 1: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 13s 112ms/step - loss: 0.2755 - binary_accuracy: 0.9044 - val_loss: 0.1234 - val_binary_accuracy: 0.9531 - lr: 0.0010

Epoch 2: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 2/100
100/100 [==============================] - ETA: 0s - loss: 0.1476 - binary_accuracy: 0.9406
Epoch 2: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 105ms/step - loss: 0.1476 - binary_accuracy: 0.9406 - val_loss: 0.1852 - val_binary_accuracy: 0.9062 - lr: 0.0010

Epoch 3: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 3/100
100/100 [==============================] - ETA: 0s - loss: 0.1350 - binary_accuracy: 0.9463
Epoch 3: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 105ms/step - loss: 0.1350 - binary_accuracy: 0.9463 - val_loss: 0.0791 - val_binary_accuracy: 0.9375 - lr: 0.0010

Epoch 4: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 4/100
100/100 [==============================] - ETA: 0s - loss: 0.1435 - binary_accuracy: 0.9450
Epoch 4: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 105ms/step - loss: 0.1435 - binary_accuracy: 0.9450 - val_loss: 0.1381 - val_binary_accuracy: 0.9219 - lr: 0.0010

Epoch 5: LearningRateScheduler setting learning rate to 0.0010000000474974513.
Epoch 5/100
100/100 [==============================] - ETA: 0s - loss: 0.1267 - binary_accuracy: 0.9463
Epoch 5: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.1267 - binary_accuracy: 0.9463 - val_loss: 0.0838 - val_binary_accuracy: 0.9531 - lr: 0.0010
lr changed to 0.0009000000427477062

Epoch 6: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 6/100
100/100 [==============================] - ETA: 0s - loss: 0.1179 - binary_accuracy: 0.9556
Epoch 6: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.1179 - binary_accuracy: 0.9556 - val_loss: 0.0130 - val_binary_accuracy: 1.0000 - lr: 9.0000e-04

Epoch 7: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 7/100
100/100 [==============================] - ETA: 0s - loss: 0.1143 - binary_accuracy: 0.9506
Epoch 7: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.1143 - binary_accuracy: 0.9506 - val_loss: 0.0915 - val_binary_accuracy: 0.9688 - lr: 9.0000e-04

Epoch 8: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 8/100
100/100 [==============================] - ETA: 0s - loss: 0.0939 - binary_accuracy: 0.9594
Epoch 8: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.0939 - binary_accuracy: 0.9594 - val_loss: 0.0665 - val_binary_accuracy: 0.9688 - lr: 9.0000e-04

Epoch 9: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 9/100
100/100 [==============================] - ETA: 0s - loss: 0.1167 - binary_accuracy: 0.9469
Epoch 9: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.1167 - binary_accuracy: 0.9469 - val_loss: 0.0630 - val_binary_accuracy: 0.9688 - lr: 9.0000e-04

Epoch 10: LearningRateScheduler setting learning rate to 0.0009000000427477062.
Epoch 10/100
100/100 [==============================] - ETA: 0s - loss: 0.1151 - binary_accuracy: 0.9500
Epoch 10: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 105ms/step - loss: 0.1151 - binary_accuracy: 0.9500 - val_loss: 0.0387 - val_binary_accuracy: 1.0000 - lr: 9.0000e-04
lr changed to 0.0008100000384729356

Epoch 11: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 11/100
100/100 [==============================] - ETA: 0s - loss: 0.0924 - binary_accuracy: 0.9681
Epoch 11: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.0924 - binary_accuracy: 0.9681 - val_loss: 0.0819 - val_binary_accuracy: 0.9688 - lr: 8.1000e-04

Epoch 12: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 12/100
100/100 [==============================] - ETA: 0s - loss: 0.1084 - binary_accuracy: 0.9556
Epoch 12: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.1084 - binary_accuracy: 0.9556 - val_loss: 0.1482 - val_binary_accuracy: 0.9375 - lr: 8.1000e-04

Epoch 13: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 13/100
100/100 [==============================] - ETA: 0s - loss: 0.0998 - binary_accuracy: 0.9569
Epoch 13: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 107ms/step - loss: 0.0998 - binary_accuracy: 0.9569 - val_loss: 0.0780 - val_binary_accuracy: 0.9688 - lr: 8.1000e-04

Epoch 14: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 14/100
100/100 [==============================] - ETA: 0s - loss: 0.0989 - binary_accuracy: 0.9631
Epoch 14: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.0989 - binary_accuracy: 0.9631 - val_loss: 0.0957 - val_binary_accuracy: 0.9688 - lr: 8.1000e-04

Epoch 15: LearningRateScheduler setting learning rate to 0.0008100000559352338.
Epoch 15/100
100/100 [==============================] - ETA: 0s - loss: 0.1124 - binary_accuracy: 0.9600
Epoch 15: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.1124 - binary_accuracy: 0.9600 - val_loss: 0.0735 - val_binary_accuracy: 0.9688 - lr: 8.1000e-04
lr changed to 0.0007290000503417104

Epoch 16: LearningRateScheduler setting learning rate to 0.0007290000794455409.
Epoch 16/100
100/100 [==============================] - ETA: 0s - loss: 0.0928 - binary_accuracy: 0.9644
Epoch 16: val_binary_accuracy did not improve from 1.00000
100/100 [==============================] - 11s 106ms/step - loss: 0.0928 - binary_accuracy: 0.9644 - val_loss: 0.0382 - val_binary_accuracy: 1.0000 - lr: 7.2900e-04
Epoch 16: early stopping

可以发现使用迁移学习后模型在验证集上的精确率到了100%，效果显著提升。

模型的预测

刚才我们已经使用3种方式生成了模型，我们使用最后一次迁移学习生成的模型进行预测

1
2
3
from tensorflow import keras
import cv2
import os
1
best_model = keras.models.load_model('best_model.h5')
1
2
3
4
5
6
7
8
9
# 在测试集种随机读取16张图片进行预测
all_test_imgs = os.listdir(test_path)
index_img = np.random.choice(range(len(all_test_imgs)), 16, replace=False)
imgs = []
for i in index_img:
    img = cv2.imread(test_path + all_test_imgs[i])
    img = cv2.resize(img, (150, 150))
    imgs.append(img)
imgs = np.array(imgs)
1
2
3
y_pred = best_model.predict(imgs)
y_pred = [1 if i > 0.5 else 0 for i in y_pred]
y_pred
1/1 [==============================] - 1s 583ms/step

[0, 0, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1, 0]

1
import matplotlib.pyplot as plt
1
label_dic = {0:'cat', 1:'dog'}
1
2
3
4
5
6
7
plt.figure(figsize=(10, 10))
for i in range(len(y_pred)):
    ax = plt.subplot(4, 4, i+1)
    ax.set_title(label_dic[y_pred[i]])
    ax.axis('off')
    ax.imshow(imgs[i][:, :, ::-1])
plt.show()

output_45_0_202303092127

-------------本文结束感谢您的阅读-------------