>- **🍨 本文为[🔗365天深度学习训练营](https://mp.weixin.qq.com/s/xLjALoOD8HPZcH563En8bQ) 中的学习记录博客**
>- **🍦 参考文章：365天深度学习训练营-第5周：运动鞋品牌识别（训练营内部成员可读）**
>- **🍖 原作者：[K同学啊](https://mp.weixin.qq.com/s/xLjALoOD8HPZcH563En8bQ)**
# 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 = "D:\\deeplearning\\day05\\DataSet\\46-data\\"
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)
roses = list(data_dir.glob('train/nike/*.jpg'))
PIL.Image.open(str(roses[0]))
# 二、数据预处理
batch_size = 32
img_height = 224
img_width = 224
"""
关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.youkuaiyun.com/article/details/117018789
"""
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
"D:\\deeplearning\\day05\\DataSet\\46-data\\train",
seed=123,
image_size=(img_height, img_width),
batch_size=batch_size)
"""
关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.youkuaiyun.com/article/details/117018789
"""
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
"D:\\deeplearning\\day05\\DataSet\\46-data\\test",
seed=123,
image_size=(img_height, img_width),
batch_size=batch_size)
'''
输出class_names
'''
class_names = train_ds.class_names
print(class_names)
'''
输出图片的尺寸
'''
for image_batch, labels_batch in train_ds:
print(image_batch.shape)
print(labels_batch.shape)
break
# 4.配置数据集
'''
tensorflow中数据集Dataset中shuffle(buffer_size)方法详解
参考文章--->https://zhuanlan.zhihu.com/p/42417456
shuffle方法用来打乱数据集中的数据顺序。
buffer_size参数设置是数据集首先会选取数据的前buffer_size个数据项，填充buffer
然后从中随机挑选一条数据，并让原数据集中一条数据补充进来，然后再从buffer中选择下一条数据输出
以此类推...
shuffle的作用是为了防止数据过拟合的重要手段，不当的buffer size会让shuffle毫无意义
'''
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
"""
关于卷积核的计算不懂的可以参考文章：https://blog.youkuaiyun.com/qq_38251616/article/details/114278995
layers.Dropout(0.4) 作用是防止过拟合，提高模型的泛化能力。
关于Dropout层的更多介绍可以参考文章：https://mtyjkh.blog.youkuaiyun.com/article/details/115826689
"""
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)), # 卷积层1，卷积核3*3
layers.AveragePooling2D((2, 2)), # 池化层1，2*2采样
layers.Conv2D(32, (3, 3), activation='relu'), # 卷积层2，卷积核3*3
layers.AveragePooling2D((2, 2)), # 池化层2，2*2采样
layers.Dropout(0.3),
layers.Conv2D(64, (3, 3), activation='relu'), # 卷积层3，卷积核3*3
layers.Dropout(0.3),
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
layers.Dense(len(class_names)) # 输出层，输出预期结果
])
model.summary() # 打印网络结构
# 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
# 设置初始学习率
initial_learning_rate = 0.0001
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate,
decay_steps=30, # 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(learning_rate=lr_schedule)
model.compile(optimizer=optimizer,
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
'''
学习率大与学习率小的优缺点分析:
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精确。
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。
注意：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：
learning_rate = initial_learning_rate * decay_rate ^ (step / decay_steps)
'''
'''
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
# 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
history = model.fit(train_ds,
validation_data=val_ds,
epochs=epochs,
callbacks=[checkpointer, earlystopper])
# 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("./45-data/Monkeypox/M06_01_04.jpg") #这里选择你需要预测的图片
img = Image.open("D:\\deeplearning\\day05\\DataSet\\46-data\\test\\nike\\2.jpg") #这里选择你需要预测的图片
image = tf.image.resize(img, [img_height, img_width])
img_array = tf.expand_dims(image, 0) #/255.0 # 记得做归一化处理（与训练集处理方式保持一致）
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)])
图片总数为： 578
Found 502 files belonging to 2 classes.
Found 76 files belonging to 2 classes.
['adidas', 'nike']
(32, 224, 224, 3)
(32,)
Model: "sequential_5"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
rescaling_5 (Rescaling) (None, 224, 224, 3) 0
_________________________________________________________________
conv2d_15 (Conv2D) (None, 222, 222, 16) 448
_________________________________________________________________
average_pooling2d_10 (Averag (None, 111, 111, 16) 0
_________________________________________________________________
conv2d_16 (Conv2D) (None, 109, 109, 32) 4640
_________________________________________________________________
average_pooling2d_11 (Averag (None, 54, 54, 32) 0
_________________________________________________________________
dropout_10 (Dropout) (None, 54, 54, 32) 0
_________________________________________________________________
conv2d_17 (Conv2D) (None, 52, 52, 64) 18496
_________________________________________________________________
dropout_11 (Dropout) (None, 52, 52, 64) 0
_________________________________________________________________
flatten_5 (Flatten) (None, 173056) 0
_________________________________________________________________
dense_10 (Dense) (None, 128) 22151296
_________________________________________________________________
dense_11 (Dense) (None, 2) 258
=================================================================
Total params: 22,175,138
Trainable params: 22,175,138
Non-trainable params: 0
_________________________________________________________________
Epoch 1/50
16/16 [==============================] - 1s 23ms/step - loss: 0.9877 - accuracy: 0.4661 - val_loss: 0.7043 - val_accuracy: 0.5000
Epoch 00001: val_accuracy improved from -inf to 0.50000, saving model to best_model.h5
Epoch 2/50
16/16 [==============================] - 0s 18ms/step - loss: 0.6999 - accuracy: 0.4781 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 00002: val_accuracy did not improve from 0.50000
Epoch 3/50
16/16 [==============================] - 0s 18ms/step - loss: 0.6937 - accuracy: 0.5020 - val_loss: 0.6922 - val_accuracy: 0.5000
Epoch 00003: val_accuracy did not improve from 0.50000
Epoch 4/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6919 - accuracy: 0.5438 - val_loss: 0.6908 - val_accuracy: 0.5263
Epoch 00004: val_accuracy improved from 0.50000 to 0.52632, saving model to best_model.h5
Epoch 5/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6910 - accuracy: 0.5199 - val_loss: 0.6894 - val_accuracy: 0.6053
Epoch 00005: val_accuracy improved from 0.52632 to 0.60526, saving model to best_model.h5
Epoch 6/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6888 - accuracy: 0.5598 - val_loss: 0.6880 - val_accuracy: 0.6447
Epoch 00006: val_accuracy improved from 0.60526 to 0.64474, saving model to best_model.h5
Epoch 7/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6877 - accuracy: 0.6414 - val_loss: 0.6868 - val_accuracy: 0.6184
Epoch 00007: val_accuracy did not improve from 0.64474
Epoch 8/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6867 - accuracy: 0.5737 - val_loss: 0.6870 - val_accuracy: 0.5395
Epoch 00008: val_accuracy did not improve from 0.64474
Epoch 9/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6831 - accuracy: 0.6096 - val_loss: 0.6830 - val_accuracy: 0.6711
Epoch 00009: val_accuracy improved from 0.64474 to 0.67105, saving model to best_model.h5
Epoch 10/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6838 - accuracy: 0.6016 - val_loss: 0.6815 - val_accuracy: 0.6711
Epoch 00010: val_accuracy did not improve from 0.67105
Epoch 11/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6809 - accuracy: 0.6135 - val_loss: 0.6790 - val_accuracy: 0.6711
Epoch 00011: val_accuracy did not improve from 0.67105
Epoch 12/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6767 - accuracy: 0.7191 - val_loss: 0.6757 - val_accuracy: 0.6711
Epoch 00012: val_accuracy did not improve from 0.67105
Epoch 13/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6755 - accuracy: 0.6355 - val_loss: 0.6739 - val_accuracy: 0.6711
Epoch 00013: val_accuracy did not improve from 0.67105
Epoch 14/50
16/16 [==============================] - 0s 17ms/step - loss: 0.6704 - accuracy: 0.6813 - val_loss: 0.6677 - val_accuracy: 0.6579
Epoch 00014: val_accuracy did not improve from 0.67105
Epoch 15/50
16/16 [==============================] - 0s 17ms/step - loss: 0.6666 - accuracy: 0.6713 - val_loss: 0.6689 - val_accuracy: 0.6447
Epoch 00015: val_accuracy did not improve from 0.67105
Epoch 16/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6623 - accuracy: 0.6295 - val_loss: 0.6594 - val_accuracy: 0.6711
Epoch 00016: val_accuracy did not improve from 0.67105
Epoch 17/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6604 - accuracy: 0.5976 - val_loss: 0.6583 - val_accuracy: 0.6579
Epoch 00017: val_accuracy did not improve from 0.67105
Epoch 18/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6530 - accuracy: 0.6793 - val_loss: 0.6582 - val_accuracy: 0.6579
Epoch 00018: val_accuracy did not improve from 0.67105
Epoch 19/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6474 - accuracy: 0.6454 - val_loss: 0.6493 - val_accuracy: 0.6842
Epoch 00019: val_accuracy improved from 0.67105 to 0.68421, saving model to best_model.h5
Epoch 20/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6455 - accuracy: 0.6594 - val_loss: 0.6539 - val_accuracy: 0.6579
Epoch 00020: val_accuracy did not improve from 0.68421
Epoch 21/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6408 - accuracy: 0.6713 - val_loss: 0.6482 - val_accuracy: 0.6579
Epoch 00021: val_accuracy did not improve from 0.68421
Epoch 22/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6353 - accuracy: 0.6653 - val_loss: 0.6426 - val_accuracy: 0.6711
Epoch 00022: val_accuracy did not improve from 0.68421
Epoch 23/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6321 - accuracy: 0.6793 - val_loss: 0.6443 - val_accuracy: 0.6711
Epoch 00023: val_accuracy did not improve from 0.68421
Epoch 24/50
16/16 [==============================] - 0s 17ms/step - loss: 0.6264 - accuracy: 0.6733 - val_loss: 0.6432 - val_accuracy: 0.6711
Epoch 00024: val_accuracy did not improve from 0.68421
Epoch 25/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6243 - accuracy: 0.6773 - val_loss: 0.6321 - val_accuracy: 0.6711
Epoch 00025: val_accuracy did not improve from 0.68421
Epoch 26/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6245 - accuracy: 0.6753 - val_loss: 0.6375 - val_accuracy: 0.6842
Epoch 00026: val_accuracy did not improve from 0.68421
Epoch 27/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6224 - accuracy: 0.6753 - val_loss: 0.6321 - val_accuracy: 0.6842
Epoch 00027: val_accuracy did not improve from 0.68421
Epoch 28/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6133 - accuracy: 0.6972 - val_loss: 0.6440 - val_accuracy: 0.6711
Epoch 00028: val_accuracy did not improve from 0.68421
Epoch 29/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6145 - accuracy: 0.6853 - val_loss: 0.6380 - val_accuracy: 0.6711
Epoch 00029: val_accuracy did not improve from 0.68421
Epoch 30/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6147 - accuracy: 0.6873 - val_loss: 0.6274 - val_accuracy: 0.6842
Epoch 00030: val_accuracy did not improve from 0.68421
Epoch 31/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6077 - accuracy: 0.6912 - val_loss: 0.6324 - val_accuracy: 0.6842
Epoch 00031: val_accuracy did not improve from 0.68421
Epoch 32/50
16/16 [==============================] - 0s 16ms/step - loss: 0.6021 - accuracy: 0.6992 - val_loss: 0.6214 - val_accuracy: 0.6579
Epoch 00032: val_accuracy did not improve from 0.68421
Epoch 33/50
16/16 [==============================] - 0s 17ms/step - loss: 0.6010 - accuracy: 0.7131 - val_loss: 0.6341 - val_accuracy: 0.6842
Epoch 00033: val_accuracy did not improve from 0.68421
Epoch 34/50
16/16 [==============================] - 0s 16ms/step - loss: 0.5984 - accuracy: 0.6972 - val_loss: 0.6257 - val_accuracy: 0.6842
Epoch 00034: val_accuracy did not improve from 0.68421
Epoch 35/50
16/16 [==============================] - 0s 17ms/step - loss: 0.5986 - accuracy: 0.7171 - val_loss: 0.6217 - val_accuracy: 0.6842
Epoch 00035: val_accuracy did not improve from 0.68421
Epoch 36/50
16/16 [==============================] - 0s 16ms/step - loss: 0.5937 - accuracy: 0.7052 - val_loss: 0.6233 - val_accuracy: 0.6842
Epoch 00036: val_accuracy did not improve from 0.68421
Epoch 37/50
16/16 [==============================] - 0s 17ms/step - loss: 0.5940 - accuracy: 0.7171 - val_loss: 0.6222 - val_accuracy: 0.6842
Epoch 00037: val_accuracy did not improve from 0.68421
Epoch 38/50
16/16 [==============================] - 0s 16ms/step - loss: 0.5892 - accuracy: 0.7211 - val_loss: 0.6281 - val_accuracy: 0.6842
Epoch 00038: val_accuracy did not improve from 0.68421
Epoch 39/50
16/16 [==============================] - 0s 16ms/step - loss: 0.5890 - accuracy: 0.7171 - val_loss: 0.6212 - val_accuracy: 0.6842
Epoch 00039: val_accuracy did not improve from 0.68421
Epoch 00039: early stopping
预测结果为： nike
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
