- 🍨 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
电脑环境：
语言环境：Python 3.8.0
编译器：Jupyter Notebook
深度学习环境：tensorflow 2.15.0
一、前期工作
1.设置GPU（如果使用的是CPU可以忽略这步）
from tensorflow import keras
from 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")
2. 导入数据
data_dir = "./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)
输出：图片总数为： 2142
打开一张图片：
Monkeypox = list(data_dir.glob('Monkeypox/*.jpg'))
PIL.Image.open(str(Monkeypox[1]))
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。
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。
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.csdn.net/article/details/117018789
"""
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="training",
seed=123,
image_size=(img_height, img_width),
batch_size=batch_size)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="validation",
seed=123,
image_size=(img_height, img_width),
batch_size=batch_size)
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)
输出：
[‘Monkeypox’, ‘Others’]
2、数据可视化
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 image_batch, labels_batch in train_ds:
print(image_batch.shape)
print(labels_batch.shape)
break
输出：
(32, 224, 224, 3)
(32,)
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
num_classes = 2
"""
关于卷积核的计算不懂的可以参考文章：https://blog.csdn.net/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。
在上一篇文章花朵识别中，训练准确率与验证准确率相差巨大就是由于模型过拟合导致的
关于Dropout层的更多介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/115826689
"""
model = models.Sequential([
layers.experimental.preprocessing.Rescaling(1./255, input_shape=(img_height, img_width, 3)), # 归一化
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(num_classes) # 输出层，输出预期结果
])
model.summary() # 打印网络结构
四、编译
# 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'])
五、训练模型
关于ModelCheckpoint的详细介绍可参考文章ModelCheckpoint 讲解【TensorFlow2入门手册】
from tensorflow.keras.callbacks import ModelCheckpoint
epochs = 50
checkpointer = ModelCheckpoint('best_model.h5', # 保存最好模型的路径
monitor='val_accuracy', # 需要监视的指标
verbose=1, # 信息展示模式，0/1
save_best_only=True, # 当设置为True时，监测指标有改进时才会保存当前的模型
save_weights_only=True) # 当设置为True时，则只保存模型权重，否则将保存整个模型（模型结构，配置信息等）
history = model.fit(train_ds,
validation_data=val_ds,
epochs=epochs,
callbacks=[checkpointer])
输出：
Epoch 1/50
53/54 [============================>.] - ETA: 0s - loss: 0.7034 - accuracy: 0.5612
Epoch 1: val_accuracy improved from -inf to 0.58879, saving model to /content/drive/MyDrive/app/T4/best_model.h5
54/54 [==============================] - 280s 976ms/step - loss: 0.7029 - accuracy: 0.5607 - val_loss: 0.6588 - val_accuracy: 0.5888
.................................
Epoch 47/50
53/54 [============================>.] - ETA: 0s - loss: 0.0545 - accuracy: 0.9875
Epoch 47: val_accuracy did not improve from 0.88318
54/54 [==============================] - 2s 45ms/step - loss: 0.0549 - accuracy: 0.9872 - val_loss: 0.4760 - val_accuracy: 0.8762
Epoch 48/50
53/54 [============================>.] - ETA: 0s - loss: 0.0525 - accuracy: 0.9857
Epoch 48: val_accuracy did not improve from 0.88318
54/54 [==============================] - 3s 48ms/step - loss: 0.0526 - accuracy: 0.9860 - val_loss: 0.4829 - val_accuracy: 0.8808
Epoch 49/50
54/54 [==============================] - ETA: 0s - loss: 0.0606 - accuracy: 0.9802
Epoch 49: val_accuracy improved from 0.88318 to 0.88551, saving model to /content/drive/MyDrive/app/T4/best_model.h5
54/54 [==============================] - 4s 76ms/step - loss: 0.0606 - accuracy: 0.9802 - val_loss: 0.5093 - val_accuracy: 0.8855
Epoch 50/50
53/54 [============================>.] - ETA: 0s - loss: 0.0614 - accuracy: 0.9786
Epoch 50: val_accuracy did not improve from 0.88551
54/54 [==============================] - 3s 51ms/step - loss: 0.0615 - accuracy: 0.9784 - val_loss: 0.4773 - val_accuracy: 0.8762
六、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(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()
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("./45-data/Monkeypox/M06_01_04.jpg") #这里选择你需要预测的图片
img = Image.open("./45-data/Others/NM01_01_00.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)
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)])
输出：
1/1 [==============================] - 0s 19ms/step
预测结果为： Others
预测正确。
七、优化
1、使用model.evaluate使用测试集评估模型
test_loss, test_acc = model.evaluate(val_ds, verbose=1) # verbose=0不显示进度条，1显示进度条
print('Test loss:', test_loss)
print('Test accuracy:', test_acc)
输出：
14/14 [==============================] - 0s 16ms/step - loss: 0.5093 - accuracy: 0.8855
Test loss: 0.5093046426773071
Test accuracy: 0.8855140209197998
在上边Loss与Accuracy图中可以看出，模型存在过拟合的问题，出现过拟合的问题解决办法有减小网络的大小、添加正则化项、添加dropout层等。
2、网络结构优化
经过多次训练和优化，最终效果最好的网络结构如下：包括5个卷积层，5个池化层，1个dropout层。
num_classes = 2
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
layers.AveragePooling2D((2, 2)), # 池化层1
layers.Conv2D(32, (3, 3), activation='relu'), # 卷积层2
layers.AveragePooling2D((2, 2)), # 池化层2
layers.Conv2D(32, (3, 3), activation='relu'), # 卷积层3
layers.AveragePooling2D((2, 2)), # 池化层3
layers.Conv2D(32, (3, 3), activation='relu'), # 卷积层4
layers.AveragePooling2D((2, 2)), # 池化层4
layers.Conv2D(64, (3, 3), activation='relu'), # 卷积层5
layers.AveragePooling2D((2, 2)), # 池化层5
layers.Dropout(0.4),
layers.Flatten(),
layers.Dense(128, activation='relu'),
layers.Dense(num_classes)
])
训练输出结果：
Epoch 1/50
54/54 [==============================] - ETA: 0s - loss: 0.6823 - accuracy: 0.5443
Epoch 1: val_accuracy improved from -inf to 0.56308, saving model to /content/drive/MyDrive/app/T4/best_model2.h5
54/54 [==============================] - 5s 52ms/step - loss: 0.6823 - accuracy: 0.5443 - val_loss: 0.6771 - val_accuracy: 0.5631
。。。。。。。。。。。。。。。。。。。。。。。。。。。。
Epoch 47/50
53/54 [============================>.] - ETA: 0s - loss: 0.3012 - accuracy: 0.8757
Epoch 47: val_accuracy improved from 0.85981 to 0.87383, saving model to /content/drive/MyDrive/app/T4/best_model2.h5
54/54 [==============================] - 3s 52ms/step - loss: 0.3042 - accuracy: 0.8734 - val_loss: 0.3395 - val_accuracy: 0.8738
Epoch 48/50
53/54 [============================>.] - ETA: 0s - loss: 0.3027 - accuracy: 0.8829
Epoch 48: val_accuracy did not improve from 0.87383
54/54 [==============================] - 2s 38ms/step - loss: 0.3056 - accuracy: 0.8810 - val_loss: 0.3355 - val_accuracy: 0.8738
Epoch 49/50
53/54 [============================>.] - ETA: 0s - loss: 0.3036 - accuracy: 0.8805
Epoch 49: val_accuracy did not improve from 0.87383
54/54 [==============================] - 2s 38ms/step - loss: 0.3021 - accuracy: 0.8816 - val_loss: 0.3387 - val_accuracy: 0.8621
Epoch 50/50
53/54 [============================>.] - ETA: 0s - loss: 0.2950 - accuracy: 0.8912
Epoch 50: val_accuracy improved from 0.87383 to 0.89019, saving model to /content/drive/MyDrive/app/T4/best_model2.h5
54/54 [==============================] - 2s 39ms/step - loss: 0.2945 - accuracy: 0.8909 - val_loss: 0.3172 - val_accuracy: 0.8902
3、 Loss与Accuracy图
从图中可以看出模型过拟合的影响大大减小，我没有在网络中增加正则化项，模型还有提升的空间。
4、使用model.evaluate评估优化后的模型
test_loss, test_acc = model.evaluate(val_ds, verbose=1) # verbose=0不显示进度条，1显示进度条
print('Test loss:', test_loss)
print('Test accuracy:', test_acc)
输出：
14/14 [==============================] - 0s 13ms/step - loss: 0.3172 - accuracy: 0.8902
Test loss: 0.31724825501441956
Test accuracy: 0.8901869058609009
测试集的loss大大减小，且acc提高了。
八、总结
对于神经网络学习图片特征过程中，过拟合现象很容易发生，耐心调参即可对模型进行优化。