🍨 本文为[🔗365天深度学习训练营(https://mp.weixin.qq.com/s/Nb93582M_5usednAKp_Jtw) 中的学习记录博客
🍦 参考文章: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)
🍖 原作者:[K同学啊 | 接辅导、项目定制](https://mtyjkh.blog.csdn.net/)
🚀 文章来源:[K同学的学习圈子](https://www.yuque.com/mingtian-fkmxf/zxwb45)
环境准备
- python 3.8
- Tensorflow2.x
- pycharm编辑工具
数据集划分与数据预处理
train = tf.keras.preprocessing.image_dataset_from_directory(
self.DataPath,
validation_split=.2,
subset="training",
seed=42,
image_size=self.img_size,
batch_size=self.batch_size
)
val = tf.keras.preprocessing.image_dataset_from_directory(
self.DataPath,
validation_split=.2,
subset="validation",
seed=42,
image_size=self.img_size,
batch_size=self.batch_size
)
AUTOTUNE = tf.data.experimental.AUTOTUNE
self.train = train.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
self.val = val.cache().prefetch(buffer_size=AUTOTUNE)
# AUTOTUNE = tf.data.experimental.AUTOTUNE
# self.train = data_train.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
# self.val = data_val.cache().prefetch(buffer_size=AUTOTUNE)
return train, val
flow_from_directory :
- directory: 目标文件夹路径,对于每一个类,该文件夹都要包含一个子文件夹.子文件夹中任何JPG、PNG、BNP、PPM的图片都会被生成器使用.详情请查看此脚本
- target_size: 整数tuple,默认为(256, 256). 图像将被resize成该尺寸
- color_mode: 颜色模式,为"grayscale",“rgb"之一,默认为"rgb”.代表这些图片是否会被转换为单通道或三通道的图片.
- classes: 可选参数,为子文件夹的列表,如[‘dogs’,‘cats’]默认为None. 若未提供,则该类别列表将从directory下的子文件夹名称/结构自动推断。每一个子文件夹都会被认为是一个新的类。(类别的顺序将按照字母表顺序映射到标签值)。通过属性class_indices可获得文件夹名与类的序号的对应字典。
- class_mode: “categorical”, “binary”, "sparse"或None之一. 默认为"categorical. 该参数决定了返回的标签数组的形式, "categorical"会返回2D的one-hot编码标签,"binary"返回1D的二值标签."sparse"返回1D的整数标签,如果为None则不返回任何标签, 生成器将仅仅生成batch数据, 这种情况在使用model.predict_generator()和model.evaluate_generator()等函数时会用到.
- batch_size: batch数据的大小,默认32
- shuffle: 是否打乱数据,默认为True
- seed: 可选参数,打乱数据和进行变换时的随机数种子
- save_to_dir: None或字符串,该参数能让你将提升后的图片保存起来,用以可视化
- save_prefix:字符串,保存提升后图片时使用的前缀, 仅当设置了save_to_dir时生效
- save_format:“png"或"jpeg"之一,指定保存图片的数据格式,默认"jpeg”
- flollow_links: 是否访问子文件夹中的软链接
VGG16知识了解
- VGG网络采用连续的小卷积核(3x3)和池化层构建深度神经网络,网络深度可以达到16层或19层。
- 每个模型分成了5个Block,所有卷积层采用的都是3x3大小的卷积核 。
- 3x3卷积核可以大幅度减少网络的参数数量,从而减少过拟合的风险。
- 多个3x3卷积核串联起来可以形成一个感受野更大的卷积核,而且这个组合具有更强的非线性能力。
- 网络搭建:
- 输入:224x224的RGB 彩色图像;
- block1:包含2个 [64x3x3] 的卷积层;
- block2:包含2个 [128x3x3] 的卷积层;
- block3:包含3个 [256x3x3] 的卷积层;
- block4:包含3个 [512x3x3] 的卷积层;
- block5:包含3个 [512x3x3] 的卷积层;
- 接着有3个全连接层;
- 一个分类输出层,经过 SoftMax 输出 1000个类的后验概率
网络结构图
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
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',name='block1_conv1')(input_tensor)
x = Conv2D(64, (3,3), activation='relu', padding='same',name='block1_conv2')(x)
x = MaxPooling2D((2,2), strides=(2,2), name = 'block1_pool')(x)
# 2nd block
x = Conv2D(128, (3,3), activation='relu', padding='same',name='block2_conv1')(x)
x = Conv2D(128, (3,3), activation='relu', padding='same',name='block2_conv2')(x)
x = MaxPooling2D((2,2), strides=(2,2), name = 'block2_pool')(x)
# 3rd block
x = Conv2D(256, (3,3), activation='relu', padding='same',name='block3_conv1')(x)
x = Conv2D(256, (3,3), activation='relu', padding='same',name='block3_conv2')(x)
x = Conv2D(256, (3,3), activation='relu', padding='same',name='block3_conv3')(x)
x = MaxPooling2D((2,2), strides=(2,2), name = 'block3_pool')(x)
# 4th block
x = Conv2D(512, (3,3), activation='relu', padding='same',name='block4_conv1')(x)
x = Conv2D(512, (3,3), activation='relu', padding='same',name='block4_conv2')(x)
x = Conv2D(512, (3,3), activation='relu', padding='same',name='block4_conv3')(x)
x = MaxPooling2D((2,2), strides=(2,2), name = 'block4_pool')(x)
# 5th block
x = Conv2D(512, (3,3), activation='relu', padding='same',name='block5_conv1')(x)
x = Conv2D(512, (3,3), activation='relu', padding='same',name='block5_conv2')(x)
x = Conv2D(512, (3,3), activation='relu', padding='same',name='block5_conv3')(x)
x = MaxPooling2D((2,2), strides=(2,2), name = 'block5_pool')(x)
# full connection
x = Flatten()(x)
x = Dense(4096, activation='relu', name='fc1')(x)
x = Dense(4096, activation='relu', name='fc2')(x)
output_tensor = Dense(nb_classes, activation='softmax', name='predictions')(x)
model = Model(input_tensor, output_tensor)
return model
model=VGG16(len(class_names), (img_width, img_height, 3))
model.summary()
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 224, 224, 3)] 0
_________________________________________________________________
block1_conv1 (Conv2D) (None, 224, 224, 64) 1792
_________________________________________________________________
block1_conv2 (Conv2D) (None, 224, 224, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 112, 112, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 112, 112, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 112, 112, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 56, 56, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 56, 56, 256) 295168
_________________________________________________________________
block3_conv2 (Conv2D) (None, 56, 56, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 56, 56, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 28, 28, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, 28, 28, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 28, 28, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 28, 28, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, 14, 14, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_conv2 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_pool (MaxPooling2D) (None, 7, 7, 512) 0
_________________________________________________________________
flatten (Flatten) (None, 25088) 0
_________________________________________________________________
fc1 (Dense) (None, 4096) 102764544
_________________________________________________________________
fc2 (Dense) (None, 4096) 16781312
_________________________________________________________________
predictions (Dense) (None, 4) 16388
=================================================================
Total params: 134,276,932
Trainable params: 134,276,932
Non-trainable params: 0
_________________________________________________________________
模型编译
- 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。
# 设置初始学习率
initial_learning_rate = 5e-3
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'])
模型训练
def fit_model(self):
train, val = self.load_data()
# self.save_show()
model = self.VGG16(len(self.class_names), (*self.img_size, 3))
model.summary()
# 设置初始学习率
initial_learning_rate = 5e-3
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'])
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
# 保存最佳模型参数
checkpointer = ModelCheckpoint(self.save_model(),
monitor='val_accuracy',
verbose=1,
save_best_only=True,
save_weights_only=True)
# 设置早停
#########################################################################################################
# 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。
#########################################################################################################
earlystopper = EarlyStopping(monitor='val_accuracy',
min_delta=0.001,
patience=30,
verbose=1)
self.history = model.fit(
train,
epochs=self.epochs,
validation_data=val,
callbacks=[checkpointer, earlystopper])
istrue = self.save_evaluate()
Epoch 1/100
60/60 [==============================] - ETA: 0s - loss: 1.4888 - accuracy: 0.2531
Epoch 00001: val_accuracy improved from -inf to 0.23333, saving model to ./checkpoints/data\data-kfd.h5
60/60 [==============================] - 268s 4s/step - loss: 1.4888 - accuracy: 0.2531 - val_loss: 1.5103 - val_accuracy: 0.2333
Epoch 2/100
60/60 [==============================] - ETA: 0s - loss: 1.4895 - accuracy: 0.2542
Epoch 00002: val_accuracy did not improve from 0.23333
60/60 [==============================] - 267s 4s/step - loss: 1.4895 - accuracy: 0.2542 - val_loss: 1.5103 - val_accuracy: 0.2333
可视化训练结果
def plot_loss(x, history):
plt.plot(x, history['val_loss'], label='val', marker='o')
plt.plot(x, history['train_loss'], label='train', marker='o')
plt.title('Loss per epoch')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(), plt.grid()
plt.savefig('./logs/loss.png')
plt.show()
def plot_acc(x, history):
plt.plot(x, history['train_acc'], label='train_acc', marker='x')
plt.plot(x, history['val_acc'], label='val_acc', marker='x')
plt.title('Acc per epoch')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(), plt.grid()
plt.savefig('./logs/acc.png')
plt.show()
总结
- 模型结构使用3x3大小的卷积核与2x2池化,w参数减少,一定程度上防止过拟合。
- 通过增加深度和宽度能有效地提升性能。
- 卷积可代替全连接,测试阶段可适应各种尺寸的图片。