>- **🍨 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]**
📌第9周:猫狗识别-2📌
- 难度:夯实基础⭐⭐
- 语言:Python3、TensorFlow2
🍺 要求:
- 找到并处理第8周的程序问题(本文给出了答案)
🍻 拔高(可选):
- 请尝试增加数据增强部分内容以提高准确率
- 可以使用哪些方式进行数据增强?(下一周给出了答案)
🔎 探索(难度有点大)
- 本文中的代码存在较大赘余,请对代码进行精简
🚀我的环境:
- 语言环境:Python3.11.7
- 编译器:jupyter notebook
- 深度学习框架:TensorFlow2.13.0
一、前期工作
1. 设置GPU
如果使用的是CPU可以注释掉这部分的代码。
import tensorflow as tf
gpus=tf.config.list_physical_devices("GPU")
if gpus:
tf.config.experimental.set_memory_growth(gpus[0],True)
tf.config.set_visible_devices([gpus[0]],"GPU")
#打印显卡信息,确认GPU可用
print(gpus)
2. 导入数据
import warnings
warnings.filterwarnings('ignore')
import pathlib
data_dir="D:\THE MNIST DATABASE\T8"
data_dir=pathlib.Path(data_dir)
image_count=len(list(data_dir.glob('*/*')))
print("图片总数为:",image_count)
运行结果:
图片总数为: 3400
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
加载训练集:
train_ds=tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="training",
seed=12,image_size=(224,224),
batch_size=8
)
运行结果:
Found 3400 files belonging to 2 classes.
Using 2720 files for training.
加载验证集:
val_ds=tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="validation",
seed=12,
image_size=(224,224),
batch_size=8
)
运行结果:
Found 3400 files belonging to 2 classes.
Using 680 files for validation.
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)
运行结果:
['cat', 'dog']
2. 再次检查数据
for image_batch,labels_batch in train_ds:
print(image_batch.shape)
print(labels_batch.shape)
break
运行结果:
(8, 224, 224, 3)
(8,)
Image_batch
是形状的张量(8, 224, 224, 3)。这是一批形状224x224x3的8张图片(最后一维指的是彩色通道RGB)。
Label_batch
是形状(8,)的张量,这些标签对应8张图片
3. 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,其详细介绍可以参考我前两篇文章,里面都有讲解。
- cache() :将数据集缓存到内存当中,加速运行
AUTOTUNE=tf.data.AUTOTUNE
def preprocess_image(image,label):
return (image/255.0,label)
#归一化处理
train_ds=train_ds.map(preprocess_image,num_parallel_calls=AUTOTUNE)
val_ds=val_ds.map(preprocess_image,num_parallel_calls=AUTOTUNE)
train_ds=train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds=val_ds.cache().prefetch(buffer_size=AUTOTUNE)
如果报 AttributeError: module 'tensorflow._api.v2.data' has no attribute 'AUTOTUNE'
错误,就将 AUTOTUNE = tf.data.AUTOTUNE
更换为 AUTOTUNE = tf.data.experimental.AUTOTUNE
,这个错误是由于版本问题引起的。
4. 可视化数据
import matplotlib.pyplot as plt
plt.figure(figsize=(15,10))
for images,labels in train_ds.take(1):
for i in range(8):
ax=plt.subplot(5,8,i+1)
plt.imshow(images[i])
plt.title(class_names[labels[i]])
plt.axis("off")
运行结果:
三、构建VG-16网络
VGG优缺点分析:
- VGG优点
VGG的结构非常简洁,整个网络都使用了同样大小的卷积核尺寸(3x3)和最大池化尺寸(2x2)。
- VGG缺点
1)训练时间过长,调参难度大。2)需要的存储容量大,不利于部署。例如存储VGG-16权重值文件的大小为500多MB,不利于安装到嵌入式系统中
结构说明:
●13个卷积层(Convolutional Layer),分别用blockX_convX表示
●3个全连接层(Fully connected Layer),分别用fcX与predictions表示
●5个池化层(Pool layer),分别用blockX_pool表示
VGG-16包含了16个隐藏层(13个卷积层和3个全连接层),故称为VGG-16
from tensorflow.keras import layers,models,Input
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Conv2D,MaxPooling2D,Dense,Flatten,Dropout
def vgg16(nb_classes,input_shape):
input_tensor=Input(shape=input_shape)
#1st block
x=Conv2D(64,(3,3),activation='relu',padding='same')(input_tensor)
x=Conv2D(64,(3,3),activation='relu',padding='same')(x)
x=MaxPooling2D((2,2),strides=(2,2))(x)
#2nd block
x=Conv2D(128,(3,3),activation='relu',padding='same')(x)
x=Conv2D(128,(3,3),activation='relu',padding='same')(x)
x=MaxPooling2D((2,2),strides=(2,2))(x)
#3rd block
x=Conv2D(256,(3,3),activation='relu',padding='same')(x)
x=Conv2D(256,(3,3),activation='relu',padding='same')(x)
x=Conv2D(256,(3,3),activation='relu',padding='same')(x)
x=MaxPooling2D((2,2),strides=(2,2))(x)
#4th block
x=Conv2D(512,(3,3),activation='relu',padding='same')(x)
x=Conv2D(512,(3,3),activation='relu',padding='same')(x)
x=Conv2D(512,(3,3),activation='relu',padding='same')(x)
x=MaxPooling2D((2,2),strides=(2,2))(x)
#5th block
x=Conv2D(512,(3,3),activation='relu',padding='same')(x)
x=Conv2D(512,(3,3),activation='relu',padding='same')(x)
x=Conv2D(512,(3,3),activation='relu',padding='same')(x)
x=MaxPooling2D((2,2),strides=(2,2))(x)
#full connection
x=Flatten()(x)
x=Dense(4096,activation='relu')(x)
x=Dense(4096,activation='relu')(x)
output_tensor=Dense(nb_classes,activation='softmax')(x)
model=Model(input_tensor,output_tensor)
return model
model=vgg16(1000,(224,224,3))
model.summary()
首先,尝试模型在没有轻量化的情况下进行运行,留待后面查看运行结果
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 224, 224, 3)] 0
conv2d (Conv2D) (None, 224, 224, 64) 1792
conv2d_1 (Conv2D) (None, 224, 224, 64) 36928
max_pooling2d (MaxPooling2 (None, 112, 112, 64) 0
D)
conv2d_2 (Conv2D) (None, 112, 112, 128) 73856
conv2d_3 (Conv2D) (None, 112, 112, 128) 147584
max_pooling2d_1 (MaxPoolin (None, 56, 56, 128) 0
g2D)
conv2d_4 (Conv2D) (None, 56, 56, 256) 295168
conv2d_5 (Conv2D) (None, 56, 56, 256) 590080
conv2d_6 (Conv2D) (None, 56, 56, 256) 590080
max_pooling2d_2 (MaxPoolin (None, 28, 28, 256) 0
g2D)
conv2d_7 (Conv2D) (None, 28, 28, 512) 1180160
conv2d_8 (Conv2D) (None, 28, 28, 512) 2359808
conv2d_9 (Conv2D) (None, 28, 28, 512) 2359808
max_pooling2d_3 (MaxPoolin (None, 14, 14, 512) 0
g2D)
conv2d_10 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_11 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_12 (Conv2D) (None, 14, 14, 512) 2359808
max_pooling2d_4 (MaxPoolin (None, 7, 7, 512) 0
g2D)
flatten (Flatten) (None, 25088) 0
dense (Dense) (None, 4096) 102764544
dense_1 (Dense) (None, 4096) 16781312
dense_2 (Dense) (None, 1000) 4097000
=================================================================
Total params: 138357544 (527.79 MB)
Trainable params: 138357544 (527.79 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
四、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。
- 评价函数(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。
model.compile(optimizer="adam",
loss='sparse_categorical_crossentropy',
metrics=['accuracy']
)
五、训练模型
修改了在T8中出现的问题,在T8中,每一轮次的训练后,都会将history的值赋给loss和accuracy,故每次记录的结果都是相同的。在本次模型中,将每轮loss和accuracy的值添加到一个列表中(即loss.append(history)),保存了所有的历史记录。但本次显示的结果是所有loss和accuracy的平均值(即np.mean())
from tqdm import tqdm
import tensorflow.keras.backend as K
import numpy as np
epochs=10
lr=1e-4
#记录训练数据,方便后面的分析
history_train_loss=[]
history_train_accuracy=[]
history_val_loss=[]
history_val_accuracy=[]
for epoch in range(epochs):
train_total=len(train_ds)
val_total=len(val_ds)
"""
total:预期迭代数目
ncols:控制进度条宽度
mininterval:进度更新最小间隔,以秒为单位(默认值:0.1)
"""
with tqdm(total=train_total,desc=f'Epoch {epoch+1}/{epochs}',mininterval=1,ncols=100) as pbar:
lr=lr*0.92
K.set_value(model.optimizer.lr,lr)
#注:此处与T8不同
train_loss=[]
train_accuracy=[]
for image,label in train_ds:
"""
训练模型,简单理解train_on_batch就是:它是比model.fit()%更高级的一个用法
"""
#这里生成的是每一个batch的acc与loss
history=model.train_on_batch(image,label)
#注:此处与T8不同
train_loss.append(history[0])
train_accuracy.append(history[1])
pbar.set_postfix({"train_loss":"%.4f"%history[0],
"train_accuracy":"%.4f"%history[1],
"lr":K.get_value(model.optimizer.lr)})
pbar.update(1)
history_train_loss.append(np.mean(train_loss))
history_train_accuracy.append(np.mean(train_accuracy))
print('开始验证!')
with tqdm(total=val_total,desc=f'Epoch {epoch+1}/{epochs}',mininterval=0.3,ncols=100) as pbar:
#注:此处与T8不同
val_loss=[]
val_accuracy=[]
for image,label in val_ds:
#这里生成的是每一个batch的acc与loss
history=model.test_on_batch(image,label)
#注:此处与T8不同
val_loss.append(history[0])
val_accuracy.append(history[1])
pbar.set_postfix({"val_loss":"%.4f"%history[0],
"val_accuracy":"%.4f"%history[1]})
pbar.update(1)
history_val_loss.append(np.mean(val_loss))
history_val_accuracy.append(np.mean(val_accuracy))
print('结束验证!')
print("验证loss为:%.4f"%np.mean(val_loss))
print("验证准确率为:%.4f"%np.mean(val_accuracy))
运行结果:
Epoch 1/10: 100%|█| 340/340 [23:53<00:00, 4.22s/it, train_loss=0.4389, train_accuracy=0.8750, lr=9.
开始验证!
Epoch 1/10: 100%|█████████████| 85/85 [00:52<00:00, 1.63it/s, val_loss=0.6702, val_accuracy=0.7500]
结束验证!
验证loss为:0.5361
验证准确率为:0.7265
Epoch 2/10: 100%|█| 340/340 [23:41<00:00, 4.18s/it, train_loss=0.1310, train_accuracy=1.0000, lr=8.
开始验证!
Epoch 2/10: 100%|█████████████| 85/85 [00:50<00:00, 1.69it/s, val_loss=0.1669, val_accuracy=1.0000]
结束验证!
验证loss为:0.2707
验证准确率为:0.8838
Epoch 3/10: 100%|█| 340/340 [23:43<00:00, 4.19s/it, train_loss=0.0046, train_accuracy=1.0000, lr=7.
开始验证!
Epoch 3/10: 100%|█████████████| 85/85 [00:54<00:00, 1.56it/s, val_loss=0.0030, val_accuracy=1.0000]
结束验证!
验证loss为:0.0445
验证准确率为:0.9809
Epoch 4/10: 100%|█| 340/340 [26:36<00:00, 4.69s/it, train_loss=0.0817, train_accuracy=1.0000, lr=7.
开始验证!
Epoch 4/10: 100%|█████████████| 85/85 [00:56<00:00, 1.49it/s, val_loss=0.0072, val_accuracy=1.0000]
结束验证!
验证loss为:0.0314
验证准确率为:0.9897
Epoch 5/10: 100%|█| 340/340 [26:48<00:00, 4.73s/it, train_loss=0.1259, train_accuracy=0.8750, lr=6.
开始验证!
Epoch 5/10: 100%|█████████████| 85/85 [00:55<00:00, 1.52it/s, val_loss=0.0593, val_accuracy=1.0000]
结束验证!
验证loss为:0.0458
验证准确率为:0.9824
Epoch 6/10: 100%|█| 340/340 [26:47<00:00, 4.73s/it, train_loss=0.0073, train_accuracy=1.0000, lr=6.
开始验证!
Epoch 6/10: 100%|█████████████| 85/85 [00:54<00:00, 1.57it/s, val_loss=0.1767, val_accuracy=0.8750]
结束验证!
验证loss为:0.0585
验证准确率为:0.9838
Epoch 7/10: 100%|█| 340/340 [24:45<00:00, 4.37s/it, train_loss=0.0024, train_accuracy=1.0000, lr=5.
开始验证!
Epoch 7/10: 100%|█████████████| 85/85 [00:52<00:00, 1.62it/s, val_loss=0.0038, val_accuracy=1.0000]
结束验证!
验证loss为:0.0235
验证准确率为:0.9897
Epoch 8/10: 100%|█| 340/340 [24:12<00:00, 4.27s/it, train_loss=0.0000, train_accuracy=1.0000, lr=5.
开始验证!
Epoch 8/10: 100%|█████████████| 85/85 [00:51<00:00, 1.66it/s, val_loss=0.0840, val_accuracy=1.0000]
结束验证!
验证loss为:0.0835
验证准确率为:0.9706
Epoch 9/10: 100%|█| 340/340 [23:52<00:00, 4.21s/it, train_loss=0.0021, train_accuracy=1.0000, lr=4.
开始验证!
Epoch 9/10: 100%|█████████████| 85/85 [00:50<00:00, 1.68it/s, val_loss=0.0367, val_accuracy=1.0000]
结束验证!
验证loss为:0.0238
验证准确率为:0.9912
Epoch 10/10: 100%|█| 340/340 [23:44<00:00, 4.19s/it, train_loss=0.0121, train_accuracy=1.0000, lr=4
开始验证!
Epoch 10/10: 100%|████████████| 85/85 [00:50<00:00, 1.68it/s, val_loss=0.0000, val_accuracy=1.0000]
结束验证!
验证loss为:0.0096
验证准确率为:0.9971
六、模型评估
epochs_range = range(epochs)
plt.figure(figsize=(14, 4))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, history_train_accuracy, label='Training Accuracy')
plt.plot(epochs_range, history_val_accuracy, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')
plt.subplot(1, 2, 2)
plt.plot(epochs_range, history_train_loss, label='Training Loss')
plt.plot(epochs_range, history_val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()
运行结果:
七、模型轻量化
由于网络原因实在无法安装上GPU版本的TensorFlow,从T6开始,模型跑起来就像蜗牛一样,本轮次的模型我完全按照vgg16的原版模型进行搭建,故而参数数量较多,模型总体大小为500MB左右。在此,我试图将模型进行轻量化,减少参数数量,削减模型大小,减少运行时间。故而在全连接层进行修改,但并不知道运行结果能否达到满意状态。修改如下:
#full connection
x=Flatten()(x)
x=Dense(1024,activation='relu')(x)
x=Dense(108,activation='relu')(x)
output_tensor=Dense(nb_classes,activation='softmax')(x)
全连接层的参数从4096减少到1024,第二层更是减少到108。查看模型大小后结果如下:
Total params: 40625524 (154.97 MB)
Trainable params: 40625524 (154.97 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
可以看出,模型参数少了2/3左右,大小变为155MB左右。运行模型后,结果如下:
Epoch 1/10: 100%|█| 340/340 [18:11<00:00, 3.21s/it, train_loss=0.3996, train_accuracy=1.0000, lr=9.
开始验证!
Epoch 1/10: 100%|█████████████| 85/85 [00:50<00:00, 1.67it/s, val_loss=0.5278, val_accuracy=0.7500]
结束验证!
验证loss为:0.4936
验证准确率为:0.8000
Epoch 2/10: 100%|█| 340/340 [18:04<00:00, 3.19s/it, train_loss=0.2368, train_accuracy=0.8750, lr=8.
开始验证!
Epoch 2/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.0124, val_accuracy=1.0000]
结束验证!
验证loss为:0.1080
验证准确率为:0.9632
Epoch 3/10: 100%|█| 340/340 [18:02<00:00, 3.18s/it, train_loss=0.0012, train_accuracy=1.0000, lr=7.
开始验证!
Epoch 3/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.0022, val_accuracy=1.0000]
结束验证!
验证loss为:0.0615
验证准确率为:0.9779
Epoch 4/10: 100%|█| 340/340 [18:02<00:00, 3.18s/it, train_loss=0.0033, train_accuracy=1.0000, lr=7.
开始验证!
Epoch 4/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.1984, val_accuracy=0.8750]
结束验证!
验证loss为:0.0519
验证准确率为:0.9809
Epoch 5/10: 100%|█| 340/340 [18:01<00:00, 3.18s/it, train_loss=0.0092, train_accuracy=1.0000, lr=6.
开始验证!
Epoch 5/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.0003, val_accuracy=1.0000]
结束验证!
验证loss为:0.0551
验证准确率为:0.9853
Epoch 6/10: 100%|█| 340/340 [18:02<00:00, 3.18s/it, train_loss=0.0098, train_accuracy=1.0000, lr=6.
开始验证!
Epoch 6/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.0223, val_accuracy=1.0000]
结束验证!
验证loss为:0.0498
验证准确率为:0.9779
Epoch 7/10: 100%|█| 340/340 [18:01<00:00, 3.18s/it, train_loss=0.0164, train_accuracy=1.0000, lr=5.
开始验证!
Epoch 7/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.0005, val_accuracy=1.0000]
结束验证!
验证loss为:0.0324
验证准确率为:0.9868
Epoch 8/10: 100%|█| 340/340 [18:01<00:00, 3.18s/it, train_loss=0.0188, train_accuracy=1.0000, lr=5.
开始验证!
Epoch 8/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.0020, val_accuracy=1.0000]
结束验证!
验证loss为:0.0730
验证准确率为:0.9706
Epoch 9/10: 100%|█| 340/340 [18:03<00:00, 3.19s/it, train_loss=0.0212, train_accuracy=1.0000, lr=4.
开始验证!
Epoch 9/10: 100%|█████████████| 85/85 [00:48<00:00, 1.75it/s, val_loss=0.0001, val_accuracy=1.0000]
结束验证!
验证loss为:0.0463
验证准确率为:0.9868
Epoch 10/10: 100%|█| 340/340 [18:04<00:00, 3.19s/it, train_loss=0.0001, train_accuracy=1.0000, lr=4
开始验证!
Epoch 10/10: 100%|████████████| 85/85 [00:50<00:00, 1.68it/s, val_loss=0.0000, val_accuracy=1.0000]
结束验证!
验证loss为:0.0492
验证准确率为:0.9868
模型结果绘图如下:
八、心得体会
本项目中自行搭建了vgg16模型,运行结果比较满意,在第10轮的时候达到了99.71%。但是,由于自身原因,一直是在用CPU运行模型,耗时十分严重。改进模型进行轻量化后,模型的准确率稍有下降,但也能达到98.68%,而模型体积大幅度减少,从性价比来看改进后的模型性价比较高。当然,这知识对于我这种CPU跑模型的情况来说,如果是在GPU情况下当然要追求更高的准确率。