第T5周：运动鞋品牌识别-CSDN博客

本文链接：https://blog.csdn.net/weixin_62819126/article/details/137210804

>- **🍨 本文为[🔗365天深度学习训练营](https://mp.weixin.qq.com/s/rbOOmire8OocQ90QM78DRA) 中的学习记录博客**
>- **🍖 原作者：[K同学啊 | 接辅导、项目定制](https://mtyjkh.blog.csdn.net/)**

🏡 我的环境：

语言环境：Python3.7

编译器：jupyter lab

深度学习环境：Pytorch

一、前期准备

1. 设置GPU

如果设备上支持GPU就使用GPU,否则使用CPU

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 = "./dataset/t5/"

data_dir = pathlib.Path(data_dir)

3. 查看数据

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[0]))

二、数据预处理

1. 加载数据

使用image_dataset_from_directory方法将磁盘中的数据加载到tf.data.Dataset中

tf.keras.preprocessing.image_dataset_from_directory()：是 TensorFlow 的 Keras 模块中的一个函数，用于从目录中创建一个图像数据集（dataset）。这个函数可以以更方便的方式加载图像数据，用于训练和评估神经网络模型。测试集与验证集的关系：

验证集并没有参与训练过程梯度下降过程的，狭义上来讲是没有参与模型的参数训练更新的。
但是广义上来讲，验证集存在的意义确实参与了一个“人工调参”的过程，我们根据每一个epoch训练之后模型在valid data上的表现来决定是否需要训练进行early stop，或者根据这个过程模型的性能变化来调整模型的超参数，如学习率，batch_size等等。
因此，我们也可以认为，验证集也参与了训练，但是并没有使得模型去overfit验证集

如果准备尝试 categorical_crossentropy损失函数，下面的代码遇到变动哈，变动细节将在下一周博客内公布。

batch_size = 32
img_height = 224
img_width = 224

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "./dataset/t5/train/",
    seed=123,
    image_size=(img_height,img_width),
    batch_size=batch_size)

Found 502 files belonging to 2 classes.

val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "./dataset/t5/test/",
    seed=123,
    image_size=(img_height,img_width),
    batch_size=batch_size)

Found 76 files belonging to 2 classes.

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
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):
        ax = plt.subplot(5,10,i+1)
        plt.imshow(images[i].numpy().astype("uint8"))
        plt.title(class_names[labels[i]])
        
        plt.axis("off")

3. 再次检查数据

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

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

Image_batch是形状的张量（32,224,224,3）。这是一批形状224x224x3的32张图片（最后一维指的是彩色通道RGB）。

Label_batch是形状（32，）的张量，这些标签对应32张图片

4. 配置数据集

AUTOTUNE = tf.data.experimental.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

卷积神经网络（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。不需要输入batch size。color_channels 为 (R,G,B) 分别对应 RGB 的三个颜色通道（color channel）。在此示例中，我们的 CNN 输入的形状是 (224, 224, 3)即彩色图像。我们需要在声明第一层时将形状赋值给参数input_shape。

model = models.Sequential([
    layers.experimental.preprocessing.Rescaling(1./255,input_shape=(img_height,img_width,3)),
    layers.Conv2D(16,(3,3),activation='relu',input_shape=(img_height,img_width,3)),
    layers.AveragePooling2D((2,2)),
    layers.Conv2D(32,(3,3),activation='relu'),
    layers.AveragePooling2D((2,2)),
    layers.Dropout(0.3),
    layers.Conv2D(64,(3,3),activation='relu'),
    layers.Dropout(0.3),

    layers.Flatten(),
    layers.Dense(128,activation='relu'),
    layers.Dense(len(class_names))
])

model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
rescaling (Rescaling)        (None, 224, 224, 3)       0         
_________________________________________________________________
conv2d (Conv2D)              (None, 222, 222, 16)      448       
_________________________________________________________________
average_pooling2d (AveragePo (None, 111, 111, 16)      0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 109, 109, 32)      4640      
_________________________________________________________________
average_pooling2d_1 (Average (None, 54, 54, 32)        0         
_________________________________________________________________
dropout (Dropout)            (None, 54, 54, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 52, 52, 64)        18496     
_________________________________________________________________
dropout_1 (Dropout)          (None, 52, 52, 64)        0         
_________________________________________________________________
flatten (Flatten)            (None, 173056)            0         
_________________________________________________________________
dense (Dense)                (None, 128)               22151296  
_________________________________________________________________
dense_1 (Dense)              (None, 2)                 258       
=================================================================
Total params: 22,175,138
Trainable params: 22,175,138
Non-trainable params: 0
_________________________________________________________________

四、训练模型

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。

1.设置动态学习率

📮 ExponentialDecay函数：

tf.keras.optimizers.schedules.ExponentialDecay是 TensorFlow 中的一个学习率衰减策略，用于在训练神经网络时动态地降低学习率。学习率衰减是一种常用的技巧，可以帮助优化算法更有效地收敛到全局最小值，从而提高模型的性能。

🔎 主要参数：

initial_learning_rate（初始学习率）：初始学习率大小。
decay_steps（衰减步数）：学习率衰减的步数。在经过 decay_steps 步后，学习率将按照指数函数衰减。例如，如果 decay_steps 设置为 10，则每10步衰减一次。
decay_rate（衰减率）：学习率的衰减率。它决定了学习率如何衰减。通常，取值在 0 到 1 之间。
staircase（阶梯式衰减）：一个布尔值，控制学习率的衰减方式。如果设置为 True，则学习率在每个 decay_steps 步之后直接减小，形成阶梯状下降。如果设置为 False，则学习率将连续衰减。

initial_learning_rate = 0.001

ls_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate,
    decay_steps=10,
    decay_rate=0.92,
    staircase=True)

optimizer = tf.keras.optimizers.Adam(learning_rate=ls_schedule)

model.compile(optimizer=optimizer,
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

注：这里设置的动态学习率为：指数衰减型（ExponentialDecay）。在每一个epoch开始前，学习率（learning_rate）都将会重置为初始学习率（initial_learning_rate），然后再重新开始衰减。计算公式如下：

学习率大与学习率小的优缺点分析:

学习率大

优点：

加快学习速率。
有助于跳出局部最优值。

缺点：

导致模型训练不收敛。
单单使用大学习率容易导致模型不精确。

学习率小

优点：

有助于模型收敛、模型细化。
提高模型精度。

缺点：

很难跳出局部最优值。
收敛缓慢。

2.早停与保存最佳模型参数

EarlyStopping()参数说明：

monitor: 被监测的数据。

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，则使用在训练的最后一步获得的模型权重。

from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping

epochs = 50

checkpointer = ModelCheckpoint("best_model_t5.h5",
                              monitor='val_accuracy',
                              verbose=1,
                              save_best_only=True,
                              save_weight_only=True)

earlystopper = EarlyStopping(monitor="val_accuracy",
                            min_delta=0.001,
                            patience=20,
                            verbose=1)

3. 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: 1.4111 - accuracy: 0.5319
Epoch 00001: val_accuracy improved from -inf to 0.50000, saving model to best_model_t5.h5
16/16 [==============================] - 14s 890ms/step - loss: 1.4111 - accuracy: 0.5319 - val_loss: 0.6894 - val_accuracy: 0.5000
Epoch 2/50
16/16 [==============================] - ETA: 0s - loss: 0.6865 - accuracy: 0.5239
Epoch 00002: val_accuracy did not improve from 0.50000
16/16 [==============================] - 16s 1s/step - loss: 0.6865 - accuracy: 0.5239 - val_loss: 0.6684 - val_accuracy: 0.5000
Epoch 3/50
16/16 [==============================] - ETA: 0s - loss: 0.6672 - accuracy: 0.5956
Epoch 00003: val_accuracy improved from 0.50000 to 0.60526, saving model to best_model_t5.h5
16/16 [==============================] - 25s 2s/step - loss: 0.6672 - accuracy: 0.5956 - val_loss: 0.6512 - val_accuracy: 0.6053
Epoch 4/50
16/16 [==============================] - ETA: 0s - loss: 0.6376 - accuracy: 0.6295
Epoch 00004: val_accuracy improved from 0.60526 to 0.65789, saving model to best_model_t5.h5
16/16 [==============================] - 18s 1s/step - loss: 0.6376 - accuracy: 0.6295 - val_loss: 0.6261 - val_accuracy: 0.6579
Epoch 5/50
16/16 [==============================] - ETA: 0s - loss: 0.5799 - accuracy: 0.6892
Epoch 00005: val_accuracy improved from 0.65789 to 0.71053, saving model to best_model_t5.h5
16/16 [==============================] - 17s 1s/step - loss: 0.5799 - accuracy: 0.6892 - val_loss: 0.5931 - val_accuracy: 0.7105
Epoch 6/50
16/16 [==============================] - ETA: 0s - loss: 0.5091 - accuracy: 0.7291
Epoch 00006: val_accuracy improved from 0.71053 to 0.73684, saving model to best_model_t5.h5
16/16 [==============================] - 18s 1s/step - loss: 0.5091 - accuracy: 0.7291 - val_loss: 0.5773 - val_accuracy: 0.7368
Epoch 7/50
16/16 [==============================] - ETA: 0s - loss: 0.4237 - accuracy: 0.8048
Epoch 00007: val_accuracy did not improve from 0.73684
16/16 [==============================] - 18s 1s/step - loss: 0.4237 - accuracy: 0.8048 - val_loss: 0.6148 - val_accuracy: 0.7368
Epoch 8/50
16/16 [==============================] - ETA: 0s - loss: 0.3595 - accuracy: 0.8386
Epoch 00008: val_accuracy improved from 0.73684 to 0.77632, saving model to best_model_t5.h5
16/16 [==============================] - 18s 1s/step - loss: 0.3595 - accuracy: 0.8386 - val_loss: 0.5377 - val_accuracy: 0.7763
Epoch 9/50
16/16 [==============================] - ETA: 0s - loss: 0.3043 - accuracy: 0.8625
Epoch 00009: val_accuracy did not improve from 0.77632
16/16 [==============================] - 17s 1s/step - loss: 0.3043 - accuracy: 0.8625 - val_loss: 0.5637 - val_accuracy: 0.7632
Epoch 10/50
16/16 [==============================] - ETA: 0s - loss: 0.2440 - accuracy: 0.8904
Epoch 00010: val_accuracy improved from 0.77632 to 0.80263, saving model to best_model_t5.h5
16/16 [==============================] - 17s 1s/step - loss: 0.2440 - accuracy: 0.8904 - val_loss: 0.5427 - val_accuracy: 0.8026
Epoch 11/50
16/16 [==============================] - ETA: 0s - loss: 0.2028 - accuracy: 0.9203
Epoch 00011: val_accuracy did not improve from 0.80263
16/16 [==============================] - 16s 1s/step - loss: 0.2028 - accuracy: 0.9203 - val_loss: 0.6144 - val_accuracy: 0.7763
Epoch 12/50
16/16 [==============================] - ETA: 0s - loss: 0.1765 - accuracy: 0.9442
Epoch 00012: val_accuracy improved from 0.80263 to 0.82895, saving model to best_model_t5.h5
16/16 [==============================] - 17s 1s/step - loss: 0.1765 - accuracy: 0.9442 - val_loss: 0.5424 - val_accuracy: 0.8289
Epoch 13/50
16/16 [==============================] - ETA: 0s - loss: 0.1360 - accuracy: 0.9562
Epoch 00013: val_accuracy did not improve from 0.82895
16/16 [==============================] - 20s 1s/step - loss: 0.1360 - accuracy: 0.9562 - val_loss: 0.5904 - val_accuracy: 0.7763
Epoch 14/50
16/16 [==============================] - ETA: 0s - loss: 0.1231 - accuracy: 0.9582
Epoch 00014: val_accuracy did not improve from 0.82895
16/16 [==============================] - 25s 2s/step - loss: 0.1231 - accuracy: 0.9582 - val_loss: 0.5937 - val_accuracy: 0.7895
Epoch 15/50
16/16 [==============================] - ETA: 0s - loss: 0.1068 - accuracy: 0.9681
Epoch 00015: val_accuracy did not improve from 0.82895
16/16 [==============================] - 36s 2s/step - loss: 0.1068 - accuracy: 0.9681 - val_loss: 0.5534 - val_accuracy: 0.8026
Epoch 16/50
16/16 [==============================] - ETA: 0s - loss: 0.0923 - accuracy: 0.9741
Epoch 00016: val_accuracy improved from 0.82895 to 0.84211, saving model to best_model_t5.h5
16/16 [==============================] - 35s 2s/step - loss: 0.0923 - accuracy: 0.9741 - val_loss: 0.5915 - val_accuracy: 0.8421
Epoch 17/50
16/16 [==============================] - ETA: 0s - loss: 0.0779 - accuracy: 0.9841
Epoch 00017: val_accuracy did not improve from 0.84211
16/16 [==============================] - 22s 1s/step - loss: 0.0779 - accuracy: 0.9841 - val_loss: 0.5828 - val_accuracy: 0.8289
Epoch 18/50
16/16 [==============================] - ETA: 0s - loss: 0.0747 - accuracy: 0.9801
Epoch 00018: val_accuracy did not improve from 0.84211
16/16 [==============================] - 23s 1s/step - loss: 0.0747 - accuracy: 0.9801 - val_loss: 0.6285 - val_accuracy: 0.8289
Epoch 19/50
16/16 [==============================] - ETA: 0s - loss: 0.0659 - accuracy: 0.9900
Epoch 00019: val_accuracy did not improve from 0.84211
16/16 [==============================] - 23s 1s/step - loss: 0.0659 - accuracy: 0.9900 - val_loss: 0.5332 - val_accuracy: 0.8158
Epoch 20/50
16/16 [==============================] - ETA: 0s - loss: 0.0668 - accuracy: 0.9861
Epoch 00020: val_accuracy did not improve from 0.84211
16/16 [==============================] - 22s 1s/step - loss: 0.0668 - accuracy: 0.9861 - val_loss: 0.6275 - val_accuracy: 0.8289
Epoch 21/50
16/16 [==============================] - ETA: 0s - loss: 0.0687 - accuracy: 0.9761
Epoch 00021: val_accuracy did not improve from 0.84211
16/16 [==============================] - 19s 1s/step - loss: 0.0687 - accuracy: 0.9761 - val_loss: 0.6639 - val_accuracy: 0.8421
Epoch 22/50
16/16 [==============================] - ETA: 0s - loss: 0.0598 - accuracy: 0.9880
Epoch 00022: val_accuracy did not improve from 0.84211
16/16 [==============================] - 22s 1s/step - loss: 0.0598 - accuracy: 0.9880 - val_loss: 0.6308 - val_accuracy: 0.8289
Epoch 23/50
16/16 [==============================] - ETA: 0s - loss: 0.0574 - accuracy: 0.9821
Epoch 00023: val_accuracy did not improve from 0.84211
16/16 [==============================] - 27s 2s/step - loss: 0.0574 - accuracy: 0.9821 - val_loss: 0.6534 - val_accuracy: 0.8289
Epoch 24/50
16/16 [==============================] - ETA: 0s - loss: 0.0547 - accuracy: 0.9861
Epoch 00024: val_accuracy did not improve from 0.84211
16/16 [==============================] - 23s 1s/step - loss: 0.0547 - accuracy: 0.9861 - val_loss: 0.5928 - val_accuracy: 0.8421
Epoch 25/50
16/16 [==============================] - ETA: 0s - loss: 0.0514 - accuracy: 0.9880
Epoch 00025: val_accuracy did not improve from 0.84211
16/16 [==============================] - 17s 1s/step - loss: 0.0514 - accuracy: 0.9880 - val_loss: 0.6205 - val_accuracy: 0.8289
Epoch 26/50
16/16 [==============================] - ETA: 0s - loss: 0.0478 - accuracy: 0.9920
Epoch 00026: val_accuracy did not improve from 0.84211
16/16 [==============================] - 22s 1s/step - loss: 0.0478 - accuracy: 0.9920 - val_loss: 0.6880 - val_accuracy: 0.8421
Epoch 27/50
16/16 [==============================] - ETA: 0s - loss: 0.0498 - accuracy: 0.9920
Epoch 00027: val_accuracy did not improve from 0.84211
16/16 [==============================] - 19s 1s/step - loss: 0.0498 - accuracy: 0.9920 - val_loss: 0.6666 - val_accuracy: 0.8421
Epoch 28/50
16/16 [==============================] - ETA: 0s - loss: 0.0421 - accuracy: 0.9920
Epoch 00028: val_accuracy did not improve from 0.84211
16/16 [==============================] - 24s 1s/step - loss: 0.0421 - accuracy: 0.9920 - val_loss: 0.6582 - val_accuracy: 0.8289
Epoch 29/50
16/16 [==============================] - ETA: 0s - loss: 0.0457 - accuracy: 0.9841
Epoch 00029: val_accuracy did not improve from 0.84211
16/16 [==============================] - 25s 2s/step - loss: 0.0457 - accuracy: 0.9841 - val_loss: 0.6752 - val_accuracy: 0.8289
Epoch 30/50
16/16 [==============================] - ETA: 0s - loss: 0.0411 - accuracy: 0.9940
Epoch 00030: val_accuracy did not improve from 0.84211
16/16 [==============================] - 13s 830ms/step - loss: 0.0411 - accuracy: 0.9940 - val_loss: 0.6697 - val_accuracy: 0.8289
Epoch 31/50
16/16 [==============================] - ETA: 0s - loss: 0.0374 - accuracy: 0.9920
Epoch 00031: val_accuracy did not improve from 0.84211
16/16 [==============================] - 12s 745ms/step - loss: 0.0374 - accuracy: 0.9920 - val_loss: 0.6438 - val_accuracy: 0.8289
Epoch 32/50
16/16 [==============================] - ETA: 0s - loss: 0.0430 - accuracy: 0.9900
Epoch 00032: val_accuracy did not improve from 0.84211
16/16 [==============================] - 12s 754ms/step - loss: 0.0430 - accuracy: 0.9900 - val_loss: 0.6547 - val_accuracy: 0.8421
Epoch 33/50
16/16 [==============================] - ETA: 0s - loss: 0.0422 - accuracy: 0.9920
Epoch 00033: val_accuracy did not improve from 0.84211
16/16 [==============================] - 12s 745ms/step - loss: 0.0422 - accuracy: 0.9920 - val_loss: 0.6382 - val_accuracy: 0.8289
Epoch 34/50
16/16 [==============================] - ETA: 0s - loss: 0.0388 - accuracy: 0.9920
Epoch 00034: val_accuracy did not improve from 0.84211
16/16 [==============================] - 12s 760ms/step - loss: 0.0388 - accuracy: 0.9920 - val_loss: 0.6462 - val_accuracy: 0.8289
Epoch 35/50
16/16 [==============================] - ETA: 0s - loss: 0.0400 - accuracy: 0.9920
Epoch 00035: val_accuracy did not improve from 0.84211
16/16 [==============================] - 12s 757ms/step - loss: 0.0400 - accuracy: 0.9920 - val_loss: 0.6155 - val_accuracy: 0.8289
Epoch 36/50
16/16 [==============================] - ETA: 0s - loss: 0.0365 - accuracy: 0.9980
Epoch 00036: val_accuracy did not improve from 0.84211
16/16 [==============================] - 12s 744ms/step - loss: 0.0365 - accuracy: 0.9980 - val_loss: 0.6172 - val_accuracy: 0.8289
Epoch 00036: 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_t5.h5')

from PIL import Image
import numpy as np

# img = Image.open("./dataset/t4/Monkeypox/M06_01_04.jpg")  #这里选择你需要预测的图片
img = Image.open("./dataset/t5/test/nike/1.jpg")  #这里选择你需要预测的图片

# 将图像转换为NumPy数组
img_array = np.array(img)

# 使用TensorFlow的函数将NumPy数组转换为张量
img_tensor = tf.convert_to_tensor(img_array, dtype=tf.float32)
image = tf.image.resize(img_tensor, [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)])