本文为为🔗365天深度学习训练营内部文章
原作者：K同学啊
import matplotlib.pyplot as plt
import tensorflow as tf
import warnings as w
w.filterwarnings('ignore')
# 支持中文
plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号import os,PIL,pathlib#隐藏警告
import warnings
warnings.filterwarnings('ignore')data_dir = "./J3-data"
data_dir = pathlib.Path(data_dir)image_count = len(list(data_dir.glob('*/*')))print("图片数为：",image_count)图片数为： 13403
batch_size = 64
img_height = 224
img_width = 224train_ds = tf.keras.preprocessing.image_dataset_from_directory(data_dir,validation_split=0.3,subset="training",seed=12,image_size=(img_height, img_width),batch_size=batch_size)Found 13403 files belonging to 2 classes. Using 9383 files for training.
val_ds = tf.keras.preprocessing.image_dataset_from_directory(data_dir,validation_split=0.3,subset="validation",seed=12,image_size=(img_height, img_width),batch_size=batch_size)Found 13403 files belonging to 2 classes. Using 4020 files for validation.
class_names = train_ds.class_names
print(class_names)for image_batch, labels_batch in train_ds:print(image_batch.shape)print(labels_batch.shape)breakAUTOTUNE = tf.data.AUTOTUNEdef preprocess_image(image,label):return (image/255.0,label)# 归一化处理
train_ds = train_ds.map(preprocess_image, num_parallel_calls=AUTOTUNE)
val_ds = val_ds.map(preprocess_image, num_parallel_calls=AUTOTUNE)train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)plt.figure(figsize=(15, 10)) # 图形的宽为15高为10for images, labels in train_ds.take(1):for i in range(15):ax = plt.subplot(3, 5, i + 1) plt.imshow(images[i])plt.title(class_names[labels[i]])plt.axis("off")
import tensorflow as tf
from tensorflow.keras import layers, models
from tensorflow.keras.applications import ResNet50, DenseNet121
from tensorflow.keras.layers import GlobalAveragePooling2D, Concatenate, Dense, Dropout, BatchNormalization, Multiply, Reshape# SE-Net 模块
def se_block(input_tensor, ratio=16):channels = input_tensor.shape[-1]se = GlobalAveragePooling2D()(input_tensor) # Squeezese = Dense(channels // ratio, activation='relu')(se)se = Dense(channels, activation='sigmoid')(se) # Excitationse = Reshape((1, 1, channels))(se)return Multiply()([input_tensor, se]) # Scale# 创建模型
def create_model(input_shape=(224, 224, 3)):# 加载预训练的 ResNet50 和 DenseNet121resnet_base = ResNet50(weights='imagenet', include_top=False, input_shape=input_shape)densenet_base = DenseNet121(weights='imagenet', include_top=False, input_shape=input_shape)# 冻结卷积层for layer in resnet_base.layers:layer.trainable = Falsefor layer in densenet_base.layers:layer.trainable = False# 输入层inputs = layers.Input(shape=input_shape)# ResNet 分支x1 = resnet_base(inputs)x1 = BatchNormalization()(x1) # 添加 BNx1 = se_block(x1) # 添加 SE-Netx1 = GlobalAveragePooling2D()(x1)# DenseNet 分支x2 = densenet_base(inputs)x2 = BatchNormalization()(x2) # 添加 BNx2 = se_block(x2) # 添加 SE-Netx2 = GlobalAveragePooling2D()(x2)# 拼接特征x = Concatenate()([x1, x2])x = Dense(256, activation='relu')(x)x = BatchNormalization()(x) # 添加 BNx = Dropout(0.5)(x)outputs = Dense(num_classes, activation='sigmoid')(x)# 构建模型model = models.Model(inputs, outputs)return model# 创建并编译模型
num_classes = 2 # 二分类
model = create_model()
model.summary()Model: "model" __________________________________________________________________________________________________Layer (type) Output Shape Param # Connected to ==================================================================================================input_3 (InputLayer) [(None, 224, 224, 3 0 [] )] resnet50 (Functional) (None, 7, 7, 2048) 23587712 ['input_3[0][0]'] densenet121 (Functional) (None, 7, 7, 1024) 7037504 ['input_3[0][0]'] batch_normalization (BatchNorm (None, 7, 7, 2048) 8192 ['resnet50[0][0]'] alization) batch_normalization_1 (BatchNo (None, 7, 7, 1024) 4096 ['densenet121[0][0]'] rmalization) global_average_pooling2d (Glob (None, 2048) 0 ['batch_normalization[0][0]'] alAveragePooling2D) global_average_pooling2d_2 (Gl (None, 1024) 0 ['batch_normalization_1[0][0]'] obalAveragePooling2D) dense (Dense) (None, 128) 262272 ['global_average_pooling2d[0][0]'] dense_2 (Dense) (None, 64) 65600 ['global_average_pooling2d_2[0][0]'] dense_1 (Dense) (None, 2048) 264192 ['dense[0][0]'] dense_3 (Dense) (None, 1024) 66560 ['dense_2[0][0]'] reshape (Reshape) (None, 1, 1, 2048) 0 ['dense_1[0][0]'] reshape_1 (Reshape) (None, 1, 1, 1024) 0 ['dense_3[0][0]'] multiply (Multiply) (None, 7, 7, 2048) 0 ['batch_normalization[0][0]', 'reshape[0][0]'] multiply_1 (Multiply) (None, 7, 7, 1024) 0 ['batch_normalization_1[0][0]', 'reshape_1[0][0]'] global_average_pooling2d_1 (Gl (None, 2048) 0 ['multiply[0][0]'] obalAveragePooling2D) global_average_pooling2d_3 (Gl (None, 1024) 0 ['multiply_1[0][0]'] obalAveragePooling2D) concatenate (Concatenate) (None, 3072) 0 ['global_average_pooling2d_1[0][0]', 'global_average_pooling2d_3[0][0]'] dense_4 (Dense) (None, 256) 786688 ['concatenate[0][0]'] batch_normalization_2 (BatchNo (None, 256) 1024 ['dense_4[0][0]'] rmalization) dropout (Dropout) (None, 256) 0 ['batch_normalization_2[0][0]'] dense_5 (Dense) (None, 2) 514 ['dropout[0][0]'] ================================================================================================== Total params: 32,084,354 Trainable params: 1,452,482 Non-trainable params: 30,631,872 __________________________________________________________________________________________________
# 设置优化器
opt = tf.keras.optimizers.Adam(learning_rate=1e-7)model.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])from keras.callbacks import EarlyStopping
# 设置早停法
early_stopping = EarlyStopping(monitor='val_loss',patience=3,verbose=1,restore_best_weights=True
)
epochs = 10history = model.fit(train_ds,validation_data=val_ds,epochs=epochs,callbacks=[early_stopping]
)Epoch 1/10 147/147 [==============================] - 1032s 7s/step - loss: 0.4909 - accuracy: 0.8249 - val_loss: 0.4680 - val_accuracy: 0.8478 Epoch 2/10 147/147 [==============================] - 1031s 7s/step - loss: 0.3099 - accuracy: 0.8759 - val_loss: 0.3266 - val_accuracy: 0.8836 Epoch 3/10 147/147 [==============================] - 1040s 7s/step - loss: 0.2522 - accuracy: 0.9029 - val_loss: 0.2955 - val_accuracy: 0.8876 Epoch 4/10 147/147 [==============================] - 1048s 7s/step - loss: 0.2063 - accuracy: 0.9173 - val_loss: 0.2651 - val_accuracy: 0.8970 Epoch 5/10 147/147 [==============================] - 1048s 7s/step - loss: 0.1705 - accuracy: 0.9338 - val_loss: 0.2778 - val_accuracy: 0.9002 Epoch 6/10 147/147 [==============================] - 1026s 7s/step - loss: 0.1379 - accuracy: 0.9455 - val_loss: 0.2927 - val_accuracy: 0.8953 Epoch 7/10 147/147 [==============================] - ETA: 0s - loss: 0.1062 - accuracy: 0.9579Restoring model weights from the end of the best epoch: 4. 147/147 [==============================] - 1019s 7s/step - loss: 0.1062 - accuracy: 0.9579 - val_loss: 0.2981 - val_accuracy: 0.9045 Epoch 7: early stopping
# 获取实际训练轮数
actual_epochs = len(history.history['accuracy'])acc = history.history['accuracy']
val_acc = history.history['val_accuracy']loss = history.history['loss']
val_loss = history.history['val_loss']epochs_range = range(actual_epochs)plt.figure(figsize=(12, 4))# 绘制准确率
plt.subplot(1, 2, 1)
plt.plot(epochs_range, acc, label='Training Accuracy')
plt.plot(epochs_range, val_acc, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')# 绘制损失
plt.subplot(1, 2, 2)
plt.plot(epochs_range, loss, label='Training Loss')
plt.plot(epochs_range, val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.savefig('准确率.png')
plt.show()
