365天深度学习训练营 第5周：运动鞋品牌识别

加载包

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

导入数据

cd  'week 5'/

/home/test/Modelwhale/deep learning/week 5

%ls

[0m[01;34m46-data[0m/

data_dir = "./46-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

roses = list(data_dir.glob('train/nike/*.jpg'))
PIL.Image.open(str(roses[1]))

加载数据集的图片

使用keras.preprocessing从磁盘上加载这些图像。

定义加载图片的一些参数，包括：批量大小、图像高度、图像宽度

batch_size = 32
img_height = 224
img_width = 224

已经分好train的数据

"""
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(
    "./46-data/train/",
    seed=123,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 502 files belonging to 2 classes.

将80％的图像用于训练

# 将80％的图像用于训练
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
 "./46-data/train/",
validation_split=0.2,
subset="validation",# subset="training", ###将80％的图像用于训练 将20％的图像用于验证
seed=123,
image_size=(img_height, img_width),
batch_size=batch_size)

Found 502 files belonging to 2 classes.
Using 100 files for validation.

验证数据

"""
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
"""
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "./46-data/test/",
    seed=123,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 76 files belonging to 2 classes.

打印数据集中鞋的类别名称

#我们可以通过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']

可视化数据

查看一下训练数据集中的20张图像

import matplotlib.pyplot as plt

plt.figure(figsize=(20, 5))

for images, labels in train_ds.take(1):
    for i in range(20):
        ax = plt.subplot(2, 10, i + 1)

        plt.imshow(images[i].numpy().astype("uint8"))
        plt.title(class_names[labels[i]])
        
        plt.axis("off")

再次检查数据

图像形状传递这些数据集来训练模型model.fit，可以手动遍历数据集并检索成批图像

for image_batch, labels_batch in train_ds:
    print(image_batch.shape)
    print(labels_batch.shape)
    break

(32, 224, 224, 3)
(32,)

image_batch是图片形状的张量(32, 224, 224, 3)。32是指批量大小；

224，224分别表示图像的高度、宽度，

3是颜色通道RGB。32张图片组成一个批次。

label_batch是形状的张量(32,)，对应32张图片的标签。

数据预处理

像素标准化

将像素的值标准化至0到1的区间内：

normalization_layer = layers.experimental.preprocessing.Rescaling(1./255)

为什么是除以255呢？由于图片的像素范围是0-255，我们把它变成0-1的范围，于是每张图像（训练集、测试集）都除以255。

标准化数据# 调用map将其应用于数据集：

normalized_ds = train_ds.map(lambda x, y: (normalization_layer(x), y))
image_batch, labels_batch = next(iter(normalized_ds))
first_image = image_batch[0]
# Notice the pixels values are now in `[0,1]`.
print(np.min(first_image), np.max(first_image))

0.0 1.0

配置数据集

shuffle() ：打乱数据，关于此函数的详细介绍可以参考：https://zhuanlan.zhihu.com/p/42417456

prefetch() ：预取数据，加速运行

cache() ：将数据集缓存到内存当中，加速运行

AUTOTUNE = tf.data.AUTOTUNE

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

构建CNN网络

常见卷积神经网络（CNN），主要由几个 卷积层Conv2D 和 池化层MaxPooling2D 层组成。卷积层与池化层的叠加实现对输入数据的特征提取，最后连接全连接层实现分类。

特征提取——卷积层与池化层
实现分类——全连接层

CNN 的输入是张量 (Tensor) 形式的 (image_height, image_width, color_channels)，包含了图像高度、宽度及颜色信息。通常图像使用 RGB 色彩模式，color_channels 为 (R,G,B) 分别对应 RGB 的三个颜色通道，即：image_height 和 image_width 根据图像的像素高度、宽度决定color_channels是3，对应RGB的3通道。

鞋数据集中的图片，形状是 (240, 240, 3)，我们可以在声明第一层时将形状赋值给参数 input_shape 。num_classes = 5

model = models.Sequential([
    layers.experimental.preprocessing.Rescaling(1./255, input_shape=(img_height, img_width, 3)),
    
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(img_height, img_width, 3)), # 卷积层1，卷积核3*3  
    layers.AveragePooling2D((2, 2)),               # 池化层1，2*2采样
    # layers.Conv2D(64, (3, 3), activation='relu'),  # 卷积层2，卷积核3*3
    layers.Dropout(0.4),  #作用是防止过拟合，提高模型的泛化能力。
    layers.AveragePooling2D((2, 2)),               # 池化层2，2*2采样
    layers.Conv2D(64, (3, 3), activation='relu'),  # 卷积层3，卷积核3*3
    layers.Dropout(0.4),  
    
    layers.Flatten(),                       # Flatten层，连接卷积层与全连接层
    layers.Dense(240, activation='relu'),   # 全连接层，特征进一步提取
    layers.Dense(len(class_names))          # 输出层，输出预期结果
])

编译

在准备对模型进行训练之前，还需要再对其进行一些设置。以下内容是在模型的编译步骤中添加的：

损失函数（loss）：用于衡量模型在训练期间的准确率。

优化器（optimizer）：决定模型如何根据其看到的数据和自身的损失函数进行更新。

指标（metrics）：用于监控训练和测试步骤。以下示例使用了准确率，即被正确分类的图像的比率。

# 设置优化器
opt = tf.keras.optimizers.Adam(learning_rate=1e-4)

model.compile(optimizer=opt,
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
#查看一下网络模型
model.summary()

Model: "sequential_6"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 rescaling_9 (Rescaling)     (None, 224, 224, 3)       0         
                                                                 
 conv2d_13 (Conv2D)          (None, 222, 222, 32)      896       
                                                                 
 average_pooling2d_10 (Avera  (None, 111, 111, 32)     0         
 gePooling2D)                                                    
                                                                 
 dropout_11 (Dropout)        (None, 111, 111, 32)      0         
                                                                 
 average_pooling2d_11 (Avera  (None, 55, 55, 32)       0         
 gePooling2D)                                                    
                                                                 
 conv2d_14 (Conv2D)          (None, 53, 53, 64)        18496     
                                                                 
 dropout_12 (Dropout)        (None, 53, 53, 64)        0         
                                                                 
 flatten_6 (Flatten)         (None, 179776)            0         
                                                                 
 dense_11 (Dense)            (None, 240)               43146480  
                                                                 
 dense_12 (Dense)            (None, 2)                 482       
                                                                 
=================================================================
Total params: 43,166,354
Trainable params: 43,166,354
Non-trainable params: 0
_________________________________________________________________

训练模型

这里我们输入准备好的训练集数据（包括图像、对应的标签），测试集的数据（包括图像、对应的标签），模型一共训练50次。

epochs = 50

history = model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=epochs
)

Epoch 1/50
16/16 [==============================] - 6s 363ms/step - loss: 1.1402 - accuracy: 0.5060 - val_loss: 0.6908 - val_accuracy: 0.5395
Epoch 2/50
16/16 [==============================] - 6s 379ms/step - loss: 0.7293 - accuracy: 0.5259 - val_loss: 0.6890 - val_accuracy: 0.5000
Epoch 3/50
16/16 [==============================] - 6s 367ms/step - loss: 0.6907 - accuracy: 0.5418 - val_loss: 0.7166 - val_accuracy: 0.5263
Epoch 4/50
16/16 [==============================] - 6s 358ms/step - loss: 0.6560 - accuracy: 0.6056 - val_loss: 0.6345 - val_accuracy: 0.6842
Epoch 5/50
16/16 [==============================] - 6s 390ms/step - loss: 0.6332 - accuracy: 0.6235 - val_loss: 0.6147 - val_accuracy: 0.6974
Epoch 6/50
16/16 [==============================] - 6s 362ms/step - loss: 0.6318 - accuracy: 0.6255 - val_loss: 0.6538 - val_accuracy: 0.6316
Epoch 7/50
16/16 [==============================] - 6s 358ms/step - loss: 0.6034 - accuracy: 0.6853 - val_loss: 0.6752 - val_accuracy: 0.6184
Epoch 8/50
16/16 [==============================] - 6s 390ms/step - loss: 0.5652 - accuracy: 0.7410 - val_loss: 0.6619 - val_accuracy: 0.6579
Epoch 9/50
16/16 [==============================] - 6s 377ms/step - loss: 0.5435 - accuracy: 0.7590 - val_loss: 0.6664 - val_accuracy: 0.6447
Epoch 10/50
16/16 [==============================] - 6s 379ms/step - loss: 0.5294 - accuracy: 0.7749 - val_loss: 0.5913 - val_accuracy: 0.7105
Epoch 11/50
16/16 [==============================] - 6s 386ms/step - loss: 0.5008 - accuracy: 0.7749 - val_loss: 0.5973 - val_accuracy: 0.6842
Epoch 12/50
16/16 [==============================] - 6s 379ms/step - loss: 0.4997 - accuracy: 0.7809 - val_loss: 0.7795 - val_accuracy: 0.6053
Epoch 13/50
16/16 [==============================] - 6s 378ms/step - loss: 0.4849 - accuracy: 0.7610 - val_loss: 0.5082 - val_accuracy: 0.7500
Epoch 14/50
16/16 [==============================] - 6s 376ms/step - loss: 0.4875 - accuracy: 0.7789 - val_loss: 0.4860 - val_accuracy: 0.7632
Epoch 15/50
16/16 [==============================] - 6s 376ms/step - loss: 0.4394 - accuracy: 0.8167 - val_loss: 0.5199 - val_accuracy: 0.7237
Epoch 16/50
16/16 [==============================] - 6s 402ms/step - loss: 0.4182 - accuracy: 0.8267 - val_loss: 0.4844 - val_accuracy: 0.7500
Epoch 17/50
16/16 [==============================] - 6s 379ms/step - loss: 0.3856 - accuracy: 0.8506 - val_loss: 0.5402 - val_accuracy: 0.7237
Epoch 18/50
16/16 [==============================] - 6s 351ms/step - loss: 0.3959 - accuracy: 0.8367 - val_loss: 0.4774 - val_accuracy: 0.7500
Epoch 19/50
16/16 [==============================] - 6s 367ms/step - loss: 0.3617 - accuracy: 0.8805 - val_loss: 0.5969 - val_accuracy: 0.7237
Epoch 20/50
16/16 [==============================] - 6s 359ms/step - loss: 0.4321 - accuracy: 0.7749 - val_loss: 0.5671 - val_accuracy: 0.7237
Epoch 21/50
16/16 [==============================] - 6s 357ms/step - loss: 0.3595 - accuracy: 0.8625 - val_loss: 0.4786 - val_accuracy: 0.7500
Epoch 22/50
16/16 [==============================] - 6s 380ms/step - loss: 0.3567 - accuracy: 0.8705 - val_loss: 0.4366 - val_accuracy: 0.8158
Epoch 23/50
16/16 [==============================] - 6s 381ms/step - loss: 0.3046 - accuracy: 0.9084 - val_loss: 0.4577 - val_accuracy: 0.7632
Epoch 24/50
16/16 [==============================] - 6s 346ms/step - loss: 0.2881 - accuracy: 0.9064 - val_loss: 0.4394 - val_accuracy: 0.7895
Epoch 25/50
16/16 [==============================] - 6s 375ms/step - loss: 0.2985 - accuracy: 0.9104 - val_loss: 0.4532 - val_accuracy: 0.7763
Epoch 26/50
16/16 [==============================] - 6s 374ms/step - loss: 0.2740 - accuracy: 0.9163 - val_loss: 0.5409 - val_accuracy: 0.7500
Epoch 27/50
16/16 [==============================] - 6s 396ms/step - loss: 0.2610 - accuracy: 0.9143 - val_loss: 0.5458 - val_accuracy: 0.7368
Epoch 28/50
16/16 [==============================] - 6s 348ms/step - loss: 0.2700 - accuracy: 0.8964 - val_loss: 0.4511 - val_accuracy: 0.7895
Epoch 29/50
16/16 [==============================] - 6s 365ms/step - loss: 0.2473 - accuracy: 0.9263 - val_loss: 0.4392 - val_accuracy: 0.7895
Epoch 30/50
16/16 [==============================] - 6s 382ms/step - loss: 0.2387 - accuracy: 0.9124 - val_loss: 0.5414 - val_accuracy: 0.7500
Epoch 31/50
16/16 [==============================] - 6s 381ms/step - loss: 0.2331 - accuracy: 0.9363 - val_loss: 0.4227 - val_accuracy: 0.8026
Epoch 32/50
16/16 [==============================] - 6s 362ms/step - loss: 0.2423 - accuracy: 0.9183 - val_loss: 0.4048 - val_accuracy: 0.8289
Epoch 33/50
16/16 [==============================] - 6s 376ms/step - loss: 0.2874 - accuracy: 0.8825 - val_loss: 0.4788 - val_accuracy: 0.7763
Epoch 34/50
16/16 [==============================] - 6s 385ms/step - loss: 0.2401 - accuracy: 0.9243 - val_loss: 0.4193 - val_accuracy: 0.7895
Epoch 35/50
16/16 [==============================] - 6s 399ms/step - loss: 0.1868 - accuracy: 0.9582 - val_loss: 0.3914 - val_accuracy: 0.8553
Epoch 36/50
16/16 [==============================] - 6s 367ms/step - loss: 0.2207 - accuracy: 0.9283 - val_loss: 0.4541 - val_accuracy: 0.7763
Epoch 37/50
16/16 [==============================] - 7s 408ms/step - loss: 0.2007 - accuracy: 0.9402 - val_loss: 0.3839 - val_accuracy: 0.8158
Epoch 38/50
16/16 [==============================] - 6s 397ms/step - loss: 0.1836 - accuracy: 0.9482 - val_loss: 0.3863 - val_accuracy: 0.8158
Epoch 39/50
16/16 [==============================] - 6s 379ms/step - loss: 0.1795 - accuracy: 0.9482 - val_loss: 0.4069 - val_accuracy: 0.8158
Epoch 40/50
16/16 [==============================] - 6s 389ms/step - loss: 0.1735 - accuracy: 0.9542 - val_loss: 0.4048 - val_accuracy: 0.8158
Epoch 41/50
16/16 [==============================] - 6s 378ms/step - loss: 0.1685 - accuracy: 0.9462 - val_loss: 0.3930 - val_accuracy: 0.8289
Epoch 42/50
16/16 [==============================] - 6s 383ms/step - loss: 0.1740 - accuracy: 0.9562 - val_loss: 0.4742 - val_accuracy: 0.7763
Epoch 43/50
16/16 [==============================] - 6s 366ms/step - loss: 0.1767 - accuracy: 0.9382 - val_loss: 0.4110 - val_accuracy: 0.8026
Epoch 44/50
16/16 [==============================] - 6s 361ms/step - loss: 0.1620 - accuracy: 0.9462 - val_loss: 0.3925 - val_accuracy: 0.8158
Epoch 45/50
16/16 [==============================] - 6s 392ms/step - loss: 0.1323 - accuracy: 0.9661 - val_loss: 0.3901 - val_accuracy: 0.8289
Epoch 46/50
16/16 [==============================] - 6s 386ms/step - loss: 0.1262 - accuracy: 0.9721 - val_loss: 0.3915 - val_accuracy: 0.8421
Epoch 47/50
16/16 [==============================] - 6s 401ms/step - loss: 0.1431 - accuracy: 0.9542 - val_loss: 0.3868 - val_accuracy: 0.8421
Epoch 48/50
16/16 [==============================] - 6s 376ms/step - loss: 0.1394 - accuracy: 0.9622 - val_loss: 0.4206 - val_accuracy: 0.8158
Epoch 49/50
16/16 [==============================] - 6s 402ms/step - loss: 0.1225 - accuracy: 0.9721 - val_loss: 0.4001 - val_accuracy: 0.8289
Epoch 50/50
16/16 [==============================] - 7s 415ms/step - loss: 0.1260 - accuracy: 0.9622 - val_loss: 0.4035 - val_accuracy: 0.8289

通常loss越小越好，那什么是loss？

简单来说是 模型预测值 和 真实值 的相差的值，反映模型预测的结果和真实值的相差程度（ 模型预测值 - 真实值 ）；

通常准确度accuracy 越高，模型效果越好。评估模型在训练和验证集上创建损失和准确性图。

模型评估¶

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

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

epochs_range = range(epochs)

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()

从图中可以看出，训练精度和验证精度相差很大，模型仅在验证集上获得了约55％的精度。

训练精度随时间增长，而验证精度在训练过程中停滞在55％左右。

训练和验证准确性之间的准确性差异很明显，这是过拟合的标志。

可能过拟合出现的原因 ：当训练示例数量很少时，像这次的只有50多张图片，该模型有时会从训练示例中的噪音或不必要的细节中学习，从而模型在新示例上的性能产生负面影响。