?? ：
- 语言环境：Python3
- 深度学习环境：TensorFlow2
?? ：
- 编译器教程：新手入门深度学习 | 1-2：编译器Jupyter Notebook
- 深度学习环境配置教程：新手入门深度学习 | 1-1:配置深度学习环境
- 一个深度学习小白需要的所有资料我都放这里了：新手入门深度学习 | 目录
建议您在学习本文之前先看下面的入门文章，以便您能更好地理解本文：
?? 新手入门深度学习 | 图像数据建模过程示例
强烈建议大家使用Jupyter Lab打开源译器的源代码，您的下一步操作将非常方便！
大家好，我是K同学！这次我为你准备了一个，我用了文章VGG16、ResNet50、IceptionV3、DenseNet121、LeNet-5、MobileNetV2、EfficientNetB0等待7个模型来识别天气，使用的数据集包括 5,531 对于不同类型天气的图像，最终模型的识别精度为 93.9% 。
?? 。
?? 您还可以在本代码中探索以下内容：
- 同一学习率对不同模型的影响
- 不同学习率下同一模型的性能
- Dropout层的作用(解决拟合问题)
??
我们的代码流程图如下：
文章目录
-
- ?? 前期准备工作
-
- ?? 导入数据
- ?? 设置评估指标metrics
- ?? 定义模型
-
- ?? VGG16模型
- ?? ResNet50模型
- ?? InceptionV3模型
- ?? DenseNet121算法模型
- ?? LeNet-5模型
- ?? MobileNetV2算法模型
- ?? EfficientNetB0算法模型
- ?? 训练模型
- ?? 结果分析
-
- ?? 精度对比分析
- ?? 损失对比分析
- ?? 召回率对比分析
- ?? 混淆矩阵
- ?? 生成评估指标
- ?? Loss/accuracy/Precision/Recall对比分析
- ?? AUC
- ?? 预测指定图片
- ?? 点击我获取 数据 代码 ??
?? 前期准备工作
import tensorflow as tf gpus = tf.config.list_physical_devices("GPU") if gpus: tf.config.experimental.set_memory_growth(gpus[0], True) #设置GPU按需使用显存用量 tf.config.set_visible_devices([gpus[0]],"GPU") plt.rcParams['font.sans-serif'] = ['SimHei'] # 用于正常显示中文标签 plt.rcParams['axes.unicode_minus'] = False # 用于正常显示负号 # 打印显卡信息，确认GPU可用 print(gpus)
[PhysicalDevic(name='/physical_device:GPU:0', device_type='GPU')]
🥗 导入数据
""" 关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/117018789 """
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
"./2-DataSet/",
validation_split=0.2,
subset="training",
label_mode = "categorical",
seed=12,
image_size=(img_height, img_width),
batch_size=batch_size)
Found 5531 files belonging to 9 classes.
Using 4425 files for training.
""" 关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/117018789 """
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
"./2-DataSet/",
validation_split=0.2,
subset="validation",
label_mode = "categorical",
seed=12,
image_size=(img_height, img_width),
batch_size=batch_size)
Found 5531 files belonging to 9 classes.
Using 1106 files for validation.
class_names = train_ds.class_names
print(class_names)
['dew', 'fogsmog', 'frost', 'hail', 'lightning', 'rain', 'rainbow', 'rime', 'snow']
AUTOTUNE = tf.data.AUTOTUNE
# 归一化
def train_preprocessing(image,label):
return (image/255.0,label)
train_ds = (
train_ds.cache()
.map(train_preprocessing) # 这里可以设置预处理函数
.prefetch(buffer_size=AUTOTUNE)
)
val_ds = (
val_ds.cache()
.map(train_preprocessing) # 这里可以设置预处理函数
.prefetch(buffer_size=AUTOTUNE)
)
plt.figure(figsize=(14, 8)) # 图形的宽为10高为5
for images, labels in train_ds.take(1):
for i in range(28):
plt.subplot(4, 7, i + 1)
plt.xticks([])
plt.yticks([])
plt.grid(False)
# 显示图片
plt.imshow(images[i])
plt.title(class_names[np.argmax(labels[i])])
plt.show()
🥙 设置评估指标metrics
评估指标用于衡量深度学习算法模型的质量，评估深度学习算法模型对于任何项目都是必不可少的。在深度学习中，也有许多不同类型的评估指标可用于衡量算法模型，例如accuracy、precision、recall、auc等都是常用的评估指标。
关于评估指标metrics的详细介绍请参考：🍦新手入门深度学习 | 3-5：评估指标metrics
metrics = [
tf.keras.metrics.CategoricalAccuracy(name='accuracy'),
tf.keras.metrics.Precision(name='precision'),
tf.keras.metrics.Recall(name='recall'),
tf.keras.metrics.AUC(name='auc')
]
🍟 定义模型
🥪 VGG16模型
from tensorflow.keras import layers, models, Input
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Dense, Flatten, Dropout,BatchNormalization,Activation
# 加载预训练模型
vgg16_base_model = tf.keras.applications.vgg16.VGG16(weights='imagenet',
include_top=False,
# input_tensor=tf.keras.Input(shape=(img_width, img_height, 3)),
input_shape=(img_width, img_height, 3),
pooling='max')
for layer in vgg16_base_model.layers:
layer.trainable = True
X = vgg16_base_model.output
# X = Dropout(0.3)(X)
output = Dense(len(class_names), activation='softmax')(X)
vgg16_model = Model(inputs=vgg16_base_model.input, outputs=output)
# 设置初始学习率
initial_learning_rate = 1e-4
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate,
decay_steps=70, # 敲黑板！！！这里是指 steps，不是指epochs
decay_rate=0.92, # lr经过一次衰减就会变成 decay_rate*lr
staircase=True)
# 将指数衰减学习率送入优化器
optimizer = tf.keras.optimizers.Adam(lr_schedule) # 这里选择不使用动态学习率
vgg16_model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='categorical_crossentropy',
metrics = metrics)
# vgg16_model.summary()
🌮 ResNet50模型
# 加载预训练模型
resnet50_base_model = tf.keras.applications.resnet50.ResNet50(weights='imagenet',
include_top=False,
input_shape=(img_width, img_height, 3),
pooling='max')
for layer in resnet50_base_model.layers:
layer.trainable = True
X = resnet50_base_model.output
# X = Dropout(0.3)(X)
output = Dense(len(class_names), activation='softmax')(X)
resnet50_model = Model(inputs=resnet50_base_model.input, outputs=output)
# optimizer = tf.keras.optimizers.Adam(lr_schedule)
resnet50_model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='categorical_crossentropy',
metrics= metrics)
# resnet50_model.summary()
🌯 InceptionV3模型
# 加载预训练模型
InceptionV3_base_model = tf.keras.applications.inception_v3.InceptionV3(weights='imagenet',
include_top=False,
input_shape=(img_width, img_height, 3),
pooling='max')
for layer in InceptionV3_base_model.layers:
layer.trainable = True
X = InceptionV3_base_model.output
# X = Dropout(0.3)(X)
output = Dense(len(class_names), activation='softmax')(X)
InceptionV3_model = Model(inputs=InceptionV3_base_model.input, outputs=output)
InceptionV3_model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='categorical_crossentropy',
metrics= metrics)
# InceptionV3_model.summary()
🥫 DenseNet121算法模型
# 加载预训练模型
DenseNet121_base_model = tf.keras.applications.densenet.DenseNet121(weights='imagenet',
include_top=False,
input_shape=(img_width, img_height, 3),
pooling='max')
for layer in DenseNet121_base_model.layers:
layer.trainable = True
X = DenseNet121_base_model.output
# X = Dropout(0.3)(X)
output = Dense(len(class_names), activation='softmax')(X)
DenseNet121_model = Model(inputs=DenseNet121_base_model.input, outputs=output)
DenseNet121_model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='categorical_crossentropy',
metrics= metrics)
# model.summary()
🍛 LeNet-5模型
LeNet5_model = keras.Sequential([
# 卷积层1
keras.layers.Conv2D(6, 5), # 使用6个5*5的卷积核对单通道32*32的图片进行卷积，结果得到6个28*28的特征图
keras.layers.MaxPooling2D(pool_size=2, strides=2), # 对28*28的特征图进行2*2最大池化，得到14*14的特征图
keras.layers.ReLU(), # ReLU激活函数
# 卷积层2
keras.layers.Conv2D(16, 5), # 使用16个5*5的卷积核对6通道14*14的图片进行卷积，结果得到16个10*10的特征图
keras.layers.MaxPooling2D(pool_size=2, strides=2), # 对10*10的特征图进行2*2最大池化，得到5*5的特征图
keras.layers.ReLU(), # ReLU激活函数
keras.layers.Flatten(),
# 全连接层1
keras.layers.Dense(500, activation='relu'), # 32*84
# 全连接层2
keras.layers.Dense(100, activation='relu'), # 32*84
# 全连接层2
keras.layers.Dense(len(class_names), activation='softmax') # 32*10
])
LeNet5_model.build(input_shape=(batch_size, img_width,img_height,3))
# LeNet5_model.summary()
LeNet5_model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics= metrics)
🍜 MobileNetV2算法模型
# 加载预训练模型
MobileNetV2_base_model = tf.keras.applications.mobilenet_v2.MobileNetV2(weights='imagenet',
include_top=False,
input_shape=(img_width, img_height, 3),
pooling='max')
for layer in MobileNetV2_base_model.layers:
layer.trainable = True
X = MobileNetV2_base_model.output
# X = Dropout(0.3)(X)
output = Dense(len(class_names), activation='softmax')(X)
MobileNetV2_model = Model(inputs=MobileNetV2_base_model.input, outputs=output)
MobileNetV2_model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='categorical_crossentropy',
metrics= metrics)
# model.summary()
🥡 EfficientNetB0算法模型
EfficientNetB0_base_model = tf.keras.applications.efficientnet.EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=(img_width, img_height, 3),
pooling='max')
for layer in EfficientNetB0_base_model.layers:
layer.trainable = True
X = EfficientNetB0_base_model.output
# X = Dropout(0.3)(X)
output = Dense(len(class_names), activation='softmax')(X)
EfficientNetB0_model = Model(inputs=EfficientNetB0_base_model.input, outputs=output)
EfficientNetB0_model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='categorical_crossentropy',
metrics= metrics)
# EfficientNetB0_model.summary()
🌭 训练模型
vgg16_history = vgg16_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)
Epoch 1/30
139/139 [==============================] - 35s 184ms/step - loss: 0.8009 - accuracy: 0.7313 - precision: 0.8380 - recall: 0.6466 - auc: 0.9579 - val_loss: 0.4339 - val_accuracy: 0.8472 - val_precision: 0.8938 - val_recall: 0.8065 - val_auc: 0.9867
Epoch 2/30
139/139 [==============================] - 18s 133ms/step - loss: 0.3605 - accuracy: 0.8744 - precision: 0.9090 - recall: 0.8447 - auc: 0.9904 - val_loss: 0.3253 - val_accuracy: 0.8933 - val_precision: 0.9251 - val_recall: 0.8599 - val_auc: 0.9924
......
Epoch 30/30
139/139 [==============================] - 19s 134ms/step - loss: 7.2972e-05 - accuracy: 1.0000 - precision: 1.0000 - recall: 1.0000 - auc: 1.0000 - val_loss: 0.3784 - val_accuracy: 0.9394 - val_precision: 0.9410 - val_recall: 0.9367 - val_auc: 0.9852
resnet50_history = resnet50_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)
Epoch 1/30
139/139 [==============================] - 21s 124ms/step - loss: 1.5570 - accuracy: 0.8165 - precision: 0.8231 - recall: 0.8134 - auc: 0.9424 - val_loss: 8.3198 - val_accuracy: 0.0642 - val_precision: 0.0643 - val_recall: 0.0642 - val_auc: 0.4878
Epoch 2/30
139/139 [==============================] - 16s 115ms/step - loss: 0.0972 - accuracy: 0.9756 - precision: 0.9762 - recall: 0.9749 - auc: 0.9969 - val_loss: 12.0035 - val_accuracy: 0.1646 - val_precision: 0.1646 - val_recall: 0.1646 - val_auc: 0.5218
。。。。。。
Epoch 30/30
139/139 [==============================] - 16s 115ms/step - loss: 3.2757e-06 - accuracy: 1.0000 - precision: 1.0000 - recall: 1.0000 - auc: 1.0000 - val_loss: 0.4380 - val_accuracy: 0.9358 - val_precision: 0.9365 - val_recall: 0.9340 - val_auc: 0.9823
InceptionV3_history = InceptionV3_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)
Epoch 1/30
139/139 [==============================] - 21s 109ms/step - loss: 0.6351 - accuracy: 0.8291 - precision: 0.8708 - recall: 0.7968 - auc: 0.9752 - val_loss: 0.4674 - val_accuracy: 0.8553 - val_precision: 0.8660 - val_recall: 0.8418 - val_auc: 0.9839
Epoch 2/30
139/139 [==============================] - 13s 92ms/step - loss: 0.0451 - accuracy: 0.9876 - precision: 0.9907 - recall: 0.9835 - auc: 0.9999 - val_loss: 0.3189 - val_accuracy: 0.9105 - val_precision: 0.9201 - val_recall: 0.9060 - val_auc: 0.9904
......
Epoch 30/30
139/139 [==============================] - 13s 92ms/step - loss: 1.6310e-05 - accuracy: 1.0000 - precision: 1.0000 - recall: 1.0000 - auc: 1.0000 - val_loss: 0.3224 - val_accuracy: 0.9231 - val_precision: 0.9264 - val_recall: 0.9222 - val_auc: 0.9884
DenseNet121_history = DenseNet121_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)
Epoch 1/30
139/139 [==============================] - 30s 158ms/step - loss: 0.9439 - accuracy: 0.8006 - precision: 0.8226 - recall: 0.7872 - auc: 0.9602 - val_loss: 0.4637 - val_accuracy: 0.8725 - val_precision: 0.8860 - val_recall: 0.8644 - val_auc: 0.9817
Epoch 2/30
139/139 [==============================] - 19s 135ms/step - loss: 0.0615 - accuracy: 0.9792 - precision: 0.9812 - recall: 0.9772 - auc: 0.9994 - val_loss: 0.3314 - val_accuracy: 0.9105 - val_precision: 0.9150 - val_recall: 0.9051 - val_auc: 0.9897
。。。。。。
Epoch 30/30
139/139 [==============================] - 19s 135ms/step - loss: 3.9534e-05 - accuracy: 1.0000 - precision: 1.0000 - recall: 1.0000 - auc: 1.0000 - val_loss: 0.2790 - val_accuracy: 0.9304 - val_precision: 0.9318 - val_recall: 0.9259 - val_auc: 0.9911
LeNet5_history = LeNet5_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)
Epoch 1/30
139/139 [==============================] - 3s 17ms/step - loss: 1.5704 - accuracy: 0.5485 - precision: 0.8639 - recall: 0.3410 - auc: 0.8849 - val_loss: 1.2588 - val_accuracy: 0.5696 - val_precision: 0.7702 - val_recall: 0.3879 - val_auc: 0.9004
Epoch 2/30
139/139 [==============================] - 2s 15ms/step - loss: 1.1065 - accuracy: 0.6350 - precision: 0.8069 - recall: 0.4495 - auc: 0.9224 - val_loss: 1.0188 - val_accuracy: 0.6754 - val_precision: 0.7986 - val_recall: 0.5054 - val_auc: 0.9342
......
Epoch 30/30
139/139 [==============================] - 2s 15ms/step - loss: 0.3052 - accuracy: 0.9306 - precision: 0.9611 - recall: 0.8884 - auc: 0.9943 - val_loss: 0.7361 - val_accuracy: 0.7649 - val_precision: 0.8367 - val_recall: 0.7134 - val_auc: 0.9646
MobileNetV2_history = MobileNetV2_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)
Epoch 1/30
139/139 [==============================] - 19s 117ms/step - loss: 1.3687 - accuracy: 0.7252 - precision: 0.7581 - recall: 0.7064 - auc: 0.9281 - val_loss: 1.4920 - val_accuracy: 0.7043 - val_precision: 0.7203 - val_recall: 0.6917 - val_auc: 0.9184
Epoch 2/30
139/139 [==============================] - 15s 109ms/step - loss: 0.0676 - accuracy: 0.9792 - precision: 0.9798 - recall: 0.9765 - auc: 0.9992 - val_loss: 1.4331 - val_accuracy: 0.7260 - val_precision: 0.7380 - val_recall: 0.7206 - val_auc: 0.9241
。。。。。。
Epoch 30/30
139/139 [==============================] - 15s 108ms/step - loss: 5.0026e-05 - accuracy: 1.0000 - precision: 1.0000 - recall: 1.0000 - auc: 1.0000 - val_loss: 0.4606 - val_accuracy: 0.8933 - val_precision: 0.8955 - val_recall: 0.8915 - val_auc: 0.9807 1.0000 - precision: 1.0000 - recall: 1.0000 -
EfficientNetB0_history = EfficientNetB0_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)
Epoch 1/30
139/139 [==============================] - 28s 162ms/step - loss: 1.6410 - accuracy: 0.6706 - precision: 0.7080 - recall: 0.6478 - auc: 0.9135 - val_loss: 9.5856 - val_accuracy: 0.0696 - val_precision: 0.0696 - val_recall: 0.0696 - val_auc: 0.4577
Epoch 2/30
139/139 [==============================] - 21s 151ms/step - loss: 0.3428 - accuracy: 0.8856 - precision: 0.9044 - recall: 0.8723 - auc: 0.9895 - val_loss: 3.5993 - val_accuracy: 0.0832 - val_precision: 0.0911 - val_recall: 0.0642 - val_auc: 0.5274
......
Epoch 30/30
139/139 [==============================] - 21s 151ms/step - loss: 0.0067 - accuracy: 0.9989 - precision: 0.9991 - recall: 0.9984 - auc: 1.0000 - val_loss: 0.9306 - val_accuracy: 0.7920 - val_precision: 0.8105 - val_recall: 0.7848 - val_auc: 0.9535
🍿 结果分析
🦪 准确率对比分析
plot_model(["VGG16","InceptionV3","DenseNet121","LeNet5","MobileNetV2"],
[vgg16_history,InceptionV3_history,DenseNet121_history,LeNet5_history,MobileNetV2_history],
["accuracy","val_accuracy"],
marker = ".")
🍣 损失对比分析
plot_model(["VGG16","resnet50","InceptionV3","DenseNet121","LeNet5","MobileNetV2"],
[vgg16_history,resnet50_history,InceptionV3_history,DenseNet121_history,LeNet5_history,MobileNetV2_history],
["loss","val_loss"],
marker = "1")
🍤 召回率对比分析
召回率(Recall)，也称查全率，是。
plot_model(["VGG16","resnet50","InceptionV3","DenseNet121","LeNet5","MobileNetV2"],
[vgg16_history,resnet50_history,InceptionV3_history,DenseNet121_history,LeNet5_history,MobileNetV2_history],
["recall","val_recall"],
marker = "")
🍘 混淆矩阵
val_pre = []
val_label = []
for images, labels in val_ds:#这里可以取部分验证数据（.take(1)）生成混淆矩阵
for image, label in zip(images, labels):
# 需要给图片增加一个维度
img_array = tf.expand_dims(image, 0)
prediction = DenseNet121_model.predict(img_array)
val_pre.append(np.argmax(prediction))
val_label.append([np.argmax(one_hot)for one_hot in [label]])
plot_cm(val_label, val_pre)
🍙 评估指标生成
support：当前行的类别在测试数据中的样本总量；precision：被判定为正例（反例）的样本中，真正的正例样本（反例样本）的比例，精度=正确预测的个数(TP)/被预测正确的个数(TP+FP)。recall：被正确分类的正例（反例）样本，占所有正例（反