T5周：运动鞋品牌识别

• 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
• 🍖 原作者：K同学啊|接辅导、项目定制

Z. 心得感受+知识点补充

1. EarlyStopping()参数说明

• monitor: 被监测的数据（这里我们用的是val_acc，验证集的精确度）
• min_delta: 在被监测的数据中被认为是提升的最小变化，例如，小于min_delta的绝对变化会被认为没有提升
• patience: 没有进步的训练轮数，在这之后训练就会被停止
• verbose: 详细信息模式
• mode: {auto, min, max}其中之一。在min模式中，当被监测的数据停止下降，训练就会停止；在max模式中，当被监测的数据停止上升，训练就会停止；在auto模式中，方向会自动从被监测的数据的名字中判断出来
• baseline: 要监控的数量的基准值。如果模型没有显示基准的改善，训练将停止。
• estore_best_weights: 是否从具有监测数量的最佳值的时期恢复模型权重。如果为False，则是否在训练的最后一步获得的模型权重

2. 学习率大或小的优缺点分析

（1）学习率大：

• 优点：
  1、加快学习速率
  2、有助于跳出局部最优值
• 缺点：
  1、导致模型训练不收敛
  2、单单使用大学习率容易导致模型不准确

（2）学习率小：

• 优点：
  1、有助于模型收敛、模型细化
  2、提高模型精度
• 缺点：
  1、很难跳出局部最优值
  2、收敛缓慢

PS：这里设置的动态学习率为：指数衰减型（ExponentialDecay）。公式为：
learning_rate = initial_learning_rate * decay_rate ^ (step / decay_steps)

一、前期工作

1. 设置GPU

from tensorflow import keras
from tensorflow.keras import layers, models
import os, PIL, pathlib
import matplotlib.pyplot as plt
import tensorflow as tf

gpus = tf.config.list_physical_devices('GPU')

if gpus:
    gpu0 = gpus[0] #如果有多个GPU，仅使用第0个GPU
    tf.config.experimental.set_memory_growth(gpu0, True) #设置GPU显存用量按需使用
    tf.config.set_visible_devices([gpu0], 'GPU')
    
gpus

2. 导入数据

data_dir = './data/'

data_dir = pathlib.Path(data_dir)

image_count = len(list(data_dir.glob('*/*/*.jpg')))

print("图片总数为：", image_count)

图片总数为： 578

3. 查看数据

roses = list(data_dir.glob('train/nike/*.jpg'))
PIL.Image.open(str(roses[0]))

二、数据预处理

1. 加载数据

batch_size = 32
img_height = 224
img_width = 224

#使用image_dataset_from_directory方法将磁盘中的数据加载到tf.data.Dataset中
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "./data/train/",
    seed = 123,
    image_size=(img_height, img_width),
    batch_size=batch_size
)

Found 502 files belonging to 2 classes.

#使用image_dataset_from_directory方法将磁盘中的数据加载到tf.data.Dataset中
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "./data/test/",
    seed = 123,
    image_size=(img_height, img_width),
    batch_size=batch_size
)

Found 76 files belonging to 2 classes.

#可以通过class_names输出数据集的标签，标签将按字母顺序对应于目录名称
class_names = train_ds.class_names
print(class_names)

[‘adidas’, ‘nike’]

2. 可视化数据

plt.figure(figsize=(20,10))
for images, labels in train_ds.take(1):
    for i in range(20):
    # 将整个figure分成5行10列，绘制第i+1个子图
        ax = plt.subplot(5, 10, i+1)
    #图像展示，cmap为颜色图谱，"plt.cm.binary为matplotlib.cm中的色表"
        plt.imshow(images[i].numpy().astype('uint8'))
    #设置x轴标签显示为图片对应的数字
        plt.title(class_names[labels[i]])
        
        plt.axis('off')

3. 再次检查数据

plt.figure(figsize=(20,10))
for images, labels in train_ds.take(1):
    for i in range(20):
    # 将整个figure分成5行10列，绘制第i+1个子图
        ax = plt.subplot(5, 10, i+1)
    #图像展示，cmap为颜色图谱，"plt.cm.binary为matplotlib.cm中的色表"
        plt.imshow(images[i].numpy().astype('uint8'))
    #设置x轴标签显示为图片对应的数字
        plt.title(class_names[labels[i]])
        
        plt.axis('off')

4. 配置数据集

AUTOTUNE = tf.data.AUTOTUNE

train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size = AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

三、构建CNN网络模型

#创建并设置卷积神经网络

model = models.Sequential([
    layers.experimental.preprocessing.Rescaling(1./255, input_shape=(img_height, img_width, 3)),
    
    
    #设置二维卷积层1， 设置16个3*3卷积核，激活函数设置为ReLU，input_shape参数将图层的输入形状设置为（32， 32， 3）
    layers.Conv2D(16, (3, 3), activation='relu', input_shape = (img_height, img_width, 3)),
    #池化层1, 2*2采样
    layers.AveragePooling2D((2, 2)),
    #设置二维卷积层2,32个3*3卷积核
    layers.Conv2D(32, (3, 3), activation='relu'),
    #池化层2, 2*2采样
    layers.AveragePooling2D((2, 2)),
    layers.Dropout(0.3),
    layers.Conv2D(64, (3, 3), activation = 'relu'), #卷积层3，卷积核3*3
    layers.Dropout(0.3), #让神经元以一定的概率停止工作，防止过拟合，提高模型的泛化能力
    
    layers.Flatten(), #Flatten层，连接卷积层与全连接层
    layers.Dense(128, activation='relu'), #全连接层，特征进一步提取，128位输出空间的维数
    layers.Dense(len(class_names)) #输出层，输出预期结果
])

model.summary()

Model: “sequential”

---

Layer (type) Output Shape Param #

rescaling_1 (Rescaling) (None, 224, 224, 3) 0

conv2d_3 (Conv2D) (None, 222, 222, 16) 448

average_pooling2d_2 (Averag (None, 111, 111, 16) 0
ePooling2D)

conv2d_4 (Conv2D) (None, 109, 109, 32) 4640

average_pooling2d_3 (Averag (None, 54, 54, 32) 0
ePooling2D)

dropout_2 (Dropout) (None, 54, 54, 32) 0

conv2d_5 (Conv2D) (None, 52, 52, 64) 18496

dropout_3 (Dropout) (None, 52, 52, 64) 0

flatten_1 (Flatten) (None, 173056) 0

dense_1 (Dense) (None, 128) 22151296

dense_2 (Dense) (None, 2) 258

=================================================================
Total params: 22,175,138
Trainable params: 22,175,138
Non-trainable params: 0

---

四、训练模型

1. 设置动态学习率

#设置初始学习率
initial_learning_rate = 0.1

lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate,
    decay_steps=10, #这里是指steps，不是指epochs
    decay_rate=0.92, #lr经过一次衰减会变成decay_rate*lr
    staircase=True
)

#将指数衰减学习率送入优化器
optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule)

model.compile(
    #设置Adam优化器
    optimizer = optimizer,
    #设置损失函数为交叉熵损失交叉熵函数
    #from_logits为True时，会将y_pred转换为概率（用softmax），否则不进行转换，通常True结果更稳定
    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=['accuracy']
)

2. 早停与保存最佳模型参数

from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping

epochs = 50

#保存最佳模型参数
checkpointer = ModelCheckpoint(
    "best_model.h5",
    monitor = 'val_accuracy',
    verbose = 1,
    save_best_only = True,
    save_weights_only = True
)

#设置早停
earlystopper = EarlyStopping(
    monitor='val_accuracy',
    min_delta=0.001,
    patience=20,
    verbose=1
)

3. 模型训练

#模型训练
history = model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=epochs,
    callbacks = [checkpointer,earlystopper]
)

Epoch 1/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 1: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 492ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 2/50
16/16 [==============================] - ETA: 0s - loss: 0.6933 - accuracy: 0.5000
Epoch 2: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 511ms/step - loss: 0.6933 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 3/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 3: val_accuracy did not improve from 0.50000
16/16 [==============================] - 7s 466ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 4/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 4: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 469ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 5/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 5: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 496ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 6/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 6: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 474ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 7/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 7: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 498ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 8/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.4681
Epoch 8: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 488ms/step - loss: 0.6932 - accuracy: 0.4681 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 9/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 9: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 489ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 10/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 10: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 502ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 11/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 11: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 497ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 12/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 12: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 473ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 13/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 13: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 479ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 14/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 14: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 481ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 15/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 15: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 479ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 16/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 16: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 476ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 17/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 17: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 474ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 18/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 18: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 475ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 19/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 19: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 477ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 20/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 20: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 481ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 21/50
16/16 [==============================] - ETA: 0s - loss: 0.6932 - accuracy: 0.5000
Epoch 21: val_accuracy did not improve from 0.50000
16/16 [==============================] - 8s 483ms/step - loss: 0.6932 - accuracy: 0.5000 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 21: early stopping

五、模型评估

1. Loss与Accuracy图

acc = history.history['accuracy']
val_acc = history.history['val_accuracy']

loss = history.history['loss']
val_loss = history.history['val_loss']

epochs_range = range(len(loss))

plt.figure(figsize=(12, 4))
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')

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.show()

2. 指定图片进行预测

#加载效果最好的模型权重
model.load_weights('best_model.h5')

from PIL import Image
import numpy as np

img = Image.open('./data/test/nike/1.jpg') #这里选择你需要预测的图片
img = np.array(img) #TensorFlow's tf.image.resize() function is expecting a tensor as input, but you're providing it with a PIL Image object.
img = tf.convert_to_tensor(img, dtype=tf.float32)
image = tf.image.resize(img, [img_height, img_width])

img_array = tf.expand_dims(image, 0)

predictions = model.predict(img_array) #这里选用你已经训练好的模型
print('预测结果为：', class_names[np.argmax(predictions)])

预测结果为： nike