深度学习 第七周 tensorflow实现咖啡豆识别

• 🍨 本文为🔗365天深度学习训练营 内部限免文章（版权归 K同学啊 所有）
• ** 参考文章地址：[🔗TensorFlow入门实战｜365天深度学习训练营-第7周：咖啡豆识别（训练营内部成员可读） ]**
• 🍖 作者：K同学啊
  

一、本周学习内容:

1、自己搭建VGG16网络

vgg16网络代码

def vgg_net(num_layer,filter):
    blk = Sequential()
    for _ in range(num_layer):
        blk.add(Conv2D(filters=filter,kernel_size=3,padding='same',activation='relu'))
    blk.add(MaxPool2D(pool_size=(2,2),strides=2))
    return blk

conv_arch = [(2,64),(2,128),(3,256),(3,512),(3,512)]
model = Sequential()
for num_layer,filter in conv_arch:
    model.add(vgg_net(num_layer,filter))
model.add(Sequential([
    Flatten(),
    Dense(4096,activation='relu'),
    Dropout(0.5),
    Dense(4096,activation='relu'),
    Dropout(0.5),
    Dense(len(class_num),activation='softmax')  # 输出类别数自行更改
]))

x = tf.random.uniform((1,224,224,3))
y = model(x)
print(model.summary())

2、第一次训练

Epoch 1: val_acc improved from -inf to 0.47500, saving model to models_weights.h5
30/30 [==============================] - 28s 622ms/step - loss: 0.5809 - acc: 0.2823 - val_loss: 0.5517 - val_acc: 0.4750
……
我设置了早停 它经过来 193次迭代后结束
Epoch 195: val_acc did not improve from 0.99583
30/30 [==============================] - 8s 261ms/step - loss: 6.0461e-05 - acc: 1.0000 - val_loss: 0.0194 - val_acc: 0.9917
最优验证集达到了 0.99583
准备开始继续训练  即将第一次训练结束的最优模型加载入网络  其他不做修改  继续训练网络
将model = load_model("网络模型") #替代自己搭建的vgg16网络

3、继续训练

完蛋 忘记修改继续训练的模型的保存名称了
第一次迭代 验证集掉到 0.98333
终于在 第102次迭代时验证集到达了1.0000

Epoch 1: val_acc improved from -inf to 0.98333, saving model to models_weights.h5
30/30 [==============================] - 31s 828ms/step - loss: 0.0176 - acc: 0.9885 - val_loss: 0.0197 - val_acc: 0.9833
……
Epoch 102: val_acc improved from 0.99167 to 1.00000, saving model to models_weights1.h5
30/30 [==============================] - 16s 537ms/step - loss: 0.0166 - acc: 0.9906 - val_loss: 0.0107 - val_acc: 1.0000
Epoch 103/1000
30/30 [==============================] - ETA: 0s - loss: 0.0112 - acc: 0.9917
Epoch 103: val_acc did not improve from 1.00000
30/30 [==============================] - 8s 244ms/step - loss: 0.0112 - acc: 0.9917 - val_loss: 0.0080 - val_acc: 1.0000

二、前言

我们的数据图片共有1200 张,四个类别,类别即为文件夹名。
类别包括:[‘Dark’, ‘Green’, ‘Light’, ‘Medium’]
在我们使用tensorflow.keras.preprocessing.image_dataset_from_directory()函数读取后会弹出下列提示
Found 1200 files belonging to 4 classes.
Using 240 files for validation.
[‘Dark’, ‘Green’, ‘Light’, ‘Medium’]

三、电脑环境

电脑系统：Windows 10
语言环境：Python 3.8.8
编译器：Pycharm 2021.1.3
深度学习环境：TensorFlow 2.8.0,keras 2.8.0
显卡及显存：RTX 3070 8G

四、前期准备

1、导入相关依赖项

from keras.models import Sequential
from keras.layers import *
from tensorflow import keras
from keras.callbacks import ModelCheckpoint
import tensorflow as tf
import matplotlib.pyplot as plt

2、设置GPU(我下载的tensorflow-gpu 默认使用GPU)

只使用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")

使用cpu和gpu
os.environ[“CUDA_VISIBLE_DEVICES”] = “-1”

3、加载数据集和展示

(1)、数据集加载

# 数据集加载
traindatadir = "./dataset"
train = keras.preprocessing.image_dataset_from_directory(directory=traindatadir,
                                                           validation_split=0.2,
                                                           subset='training',
                                                           seed=469,
                                                           label_mode='categorical',
                                                           image_size=(224,224),
                                                           batch_size=32)
val = keras.preprocessing.image_dataset_from_directory(directory=traindatadir,
                                                           validation_split=0.2,
                                                           subset='validation',
                                                           seed=469,
                                                           label_mode='categorical',
                                                           image_size=(224,224),
                                                           batch_size=32)


class_names = train.class_names
print(class_names)

(2)、数据展示

图片展示

# 图片展示
plt.figure(figsize=(20, 5))  # 创建一个画布，画布大小为宽20、高5(单位为英寸inch)
for images,labels in train.take(1):
    for i in range(20):
     # 将整个画布分成2行10列，绘制第i+1个子图。
       plt.subplot(2, 10, i+1)
       plt.imshow(images[i].numpy().astype("uint8"), cmap=plt.cm.binary)
       plt.title(class_names[np.argmax(labels[i])])  # 由于我们使用了one_hot 这行与之前不同
       plt.axis('off')

plt.show()  #使用pycharm的需要加入这行代码才能将图像显示出来
# train.take(i) 取出第i组图片数据和标签,即将所有图片打包后一个batch为32,

五、数据预处理

注意此时我们使用了热编码加载数据 标签格式变为:

# 数据预处理
tf.keras.layers.experimental.preprocessing.Rescaling(1./255,input_shape=(224,224,3))

六、搭建CNN网络

# 网络模型
def vgg_net(num_layer,filter):
    blk = Sequential()
    for _ in range(num_layer):
        blk.add(Conv2D(filters=filter,kernel_size=3,padding='same',activation='relu'))
    blk.add(MaxPool2D(pool_size=(2,2),strides=2))
    return blk

conv_arch = [(2,64),(2,128),(3,256),(3,512),(3,512)]
# model = Sequential([Conv2D(filters=64,kernel_size=3,padding='same',activation='relu',input_shape=(224,224,1))])
model = Sequential()
for num_layer,filter in conv_arch:
    model.add(vgg_net(num_layer,filter))
# print(model.summary())
model.add(Sequential([
    Flatten(),
    Dense(4096,activation='relu'),
    Dropout(0.5),
    Dense(4096,activation='relu'),
    Dropout(0.5),
    Dense(17,activation='softmax')
]))

x = tf.random.uniform((1,224,224,3))
y = model(x)
print('y',y)
print(model.summary())

# 设置保存最优模型
checkpointer = ModelCheckpoint(
    filepath='models_weights.h5', #模型保存位置
    monitor='val_acc',   # 监测参数
    verbose=1,      # 0为不输出日志信息,1为输出进度条记录,2为每个epoch输出一条记录
    save_best_only=True,  # 仅保存最优模型
    save_weights_only= True # 仅保存权重文件 注意如果仅保存权重模型,还需要自己搭建原模型对应的网络
)
# 设置早停
earlystopper = EarlyStopping(
    monitor="val_accuracy",  # 被检测的数据
    min_delta=0.001,    # 被检测数据中被认为的最小变化值
    patience=50,    # 训练没有进步的轮次
    verbose=1   # 日志
)
# 设置 动态学习率
initial_learing_rate = 0.001
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learing_rate,
    decay_steps=8,    # 训练多少个steps衰减一次(注意不是epochs) staircase = True 时有效
    decay_rate = 0.92, # 衰减的比例 用原本学习率乘以该数值获得衰减后的学习率
    staircase = True   # 如果为True, 学习率呈现阶梯型下降趋势.
)


model.compile(optimizer=keras.optimizers.SGD(learning_rate=lr_schedule),loss=loss=keras.losses.binary_crossentropy,metrics=['acc'])
history = model.fit(train,validation_data=val,epochs=1000,verbose=1,
                    callbacks=[checkpointer,earlystopper])
evaluate = model.evaluate(val)

print(evaluate)

七、绘制损失函数图像和准确度图像

绘制代码与前几周相同

# 画准确度图
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs_range = range(1000)
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()

以上就是我本周的学习内容