- Python版本： Python3.x
- 运行平台： Windows
- IDE： jupyter / Colab / pycharm
- 转载请标明出处：https://blog.csdn.net/Tian121381
- 资料下载，提取码：1ibt
一、前言
前面我们学习了什么是过拟合，以及解决过拟合的一个方法----图像增强。这周我们将会学习一个新的方法解决过拟合。同时，我们也会学一个新的技巧，有时我们需要几周的时间去训练一个大型模型，每次改进我们可能会重新对模型进行训练，或者我们看到网络上开源的大型模型时，想直接使用，或做一定的更改后使用，这时应该怎么办呢？这就涉及到我们这次要讲的迁移学习了，让我们一起看一下吧！
二、迁移学习实例
获取授权码，加载谷歌网盘（ Colab / pycharm可省略）
from google.colab import drive
drive.mount('/content/drive')
导入相关包
import os
from tensorflow.keras import layers #挑选层
from tensorflow.keras import Model
from tensorflow.keras.applications.inception_v3 import InceptionV3 #导入Inception网络，此模型有1000各类别140万图片
数据处理
导入数据，解压到指定文件夹。
#导入相关包
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import os
import zipfile
#数据所在位置,解压
local_zip = '/content/drive/My Drive/Colab Notebooks/DateSet/cats_and_dogs_filtered.zip'
zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('/content/drive/My Drive/Colab Notebooks/tmp/')
zip_ref.close()
定义根目录，方便后期调用。
base_dir = '/content/drive/My Drive/Colab Notebooks/tmp/cats_and_dogs_filtered'
train_dir = os.path.join( base_dir, 'train')
validation_dir = os.path.join( base_dir, 'validation')
train_cats_dir = os.path.join(train_dir, 'cats')
train_dogs_dir = os.path.join(train_dir, 'dogs')
validation_cats_dir = os.path.join(validation_dir, 'cats')
validation_dogs_dir = os.path.join(validation_dir, 'dogs')
train_cat_fnames = os.listdir(train_cats_dir)
train_dog_fnames = os.listdir(train_dogs_dir)
#训练集猫狗个数
train_dog_num = len(os.listdir(train_dogs_dir))
train_cat_num = len(os.listdir(train_cats_dir))
print(train_cat_num)
print(train_dog_num)
结果：
1000
1000
训练集猫狗各1000个。
将图像数据定义图像生成器。为以后训练做打算。
#添加元素到图像生成器
train_datagen = ImageDataGenerator(rescale = 1./255.,
rotation_range = 40,
width_shift_range = 0.2,
height_shift_range = 0.2,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip = True)
# 验证集图像生成器
test_datagen = ImageDataGenerator( rescale = 1.0/255. )
# 定义图像位置，20个一组传入
train_generator = train_datagen.flow_from_directory(train_dir,
batch_size = 20,
class_mode = 'binary',
target_size = (150, 150))
validation_generator = test_datagen.flow_from_directory( validation_dir,
batch_size = 20,
class_mode = 'binary',
target_size = (150, 150))
结果：
Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.
可以看到训练集有2000个图像，共两个类。验证集有1000个图像，共两个类。
数据处理完成，该处理模型了。
模型构造—1
此项目我们要使用之前训练过的模型，具体操作如下。
#预先训练的权重信息
local_weights_file = '/content/drive/My Drive/Colab Notebooks/DateSet/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'
#定义先驱模型
"""
InceptionV3()
参数：
include_top：是否保留顶层的全连接网络
weights：None代表随机初始化，即不加载预训练权重。'imagenet’代表加载预训练权重
input_tensor：可填入Keras tensor作为模型的图像输出tensor
input_shape：可选，仅当include_top=False有效，应为长为3的tuple，指明输入图片的shape，图片的宽高必须大于197，如(200,200,3)
pooling：当include_top=False时，该参数指定了池化方式。None代表不池化，最后一个卷积层的输出为4D张量。‘avg’代表全局平均池化，‘max’代表全局最大值池化。
classes：可选，图片分类的类别数，仅当include_top=True并且不加载预训练权重时可用。
"""
pre_trained_model = InceptionV3(input_shape = (150,150,3),
include_top = False,# 是否保留顶层的3个全连接网络
weights = None #None代表随机初始化，即不加载预训练权重。'imagenet’代表加载预训练权重
)
#加载模型权重
pre_trained_model.load_weights(local_weights_file)
#特征值提取
#仅仅用其做特征提取 不需要更新权值
for layer in pre_trained_model.layers:
print(layer.name)
layer.trainable = False
last_layer = pre_trained_model.get_layer('mixed7') #设置最后一层
print('last layer output shape: ', last_layer.output_shape)
last_output = last_layer.output
加载完预先模型，开始编写自己的模型
from tensorflow.keras.optimizers import RMSprop
x = layers.Flatten()(last_output)
#print(x.shape)
x = layers.Dense(1024, activation='relu')(x)
#print(x.shape)
x = layers.Dense (1, activation='sigmoid')(x)
#Model抽象类创建模型，
model = Model( pre_trained_model.input, x)
model.compile(optimizer = RMSprop(lr=0.0001),
loss = 'binary_crossentropy',
metrics = ['acc'])
print(model.summary())
模型全部完成了，看一下我们的模型结构吧。（长图预警）
结果：
Model: "model_1"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 150, 150, 3) 0
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 74, 74, 32) 864 input_1[0][0]
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 74, 74, 32) 96 conv2d[0][0]
__________________________________________________________________________________________________
activation (Activation) (None, 74, 74, 32) 0 batch_normalization[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 72, 72, 32) 9216 activation[0][0]
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 72, 72, 32) 96 conv2d_1[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 72, 72, 32) 0 batch_normalization_1[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 72, 72, 64) 18432 activation_1[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 72, 72, 64) 192 conv2d_2[0][0]
__________________________________________________________________________________________________
activation_2 (Activation) (None, 72, 72, 64) 0 batch_normalization_2[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 35, 35, 64) 0 activation_2[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 35, 35, 80) 5120 max_pooling2d[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 35, 35, 80) 240 conv2d_3[0][0]
__________________________________________________________________________________________________
activation_3 (Activation) (None, 35, 35, 80) 0 batch_normalization_3[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 33, 33, 192) 138240 activation_3[0][0]
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 33, 33, 192) 576 conv2d_4[0][0]
__________________________________________________________________________________________________
activation_4 (Activation) (None, 33, 33, 192) 0 batch_normalization_4[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 16, 16, 192) 0 activation_4[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 16, 16, 64) 12288 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 16, 16, 64) 192 conv2d_8[0][0]
__________________________________________________________________________________________________
activation_8 (Activation) (None, 16, 16, 64) 0 batch_normalization_8[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 16, 16, 48) 9216 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 16, 16, 96) 55296 activation_8[0][0]
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 16, 16, 48) 144 conv2d_6[0][0]
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 16, 16, 96) 288 conv2d_9[0][0]
__________________________________________________________________________________________________
activation_6 (Activation) (None, 16, 16, 48) 0 batch_normalization_6[0][0]
__________________________________________________________________________________________________
activation_9 (Activation) (None, 16, 16, 96) 0 batch_normalization_9[0][0]
__________________________________________________________________________________________________
average_pooling2d (AveragePooli (None, 16, 16, 192) 0 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 16, 16, 64) 12288 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 16, 16, 64) 76800 activation_6[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 16, 16, 96) 82944 activation_9[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 16, 16, 32) 6144 average_pooling2d[0][0]
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 16, 16, 64) 192 conv2d_5[0][0]
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 16, 16, 64) 192 conv2d_7[0][0]
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 16, 16, 96) 288 conv2d_10[0][0]
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 16, 16, 32) 96 conv2d_11[0][0]
__________________________________________________________________________________________________
activation_5 (Activation) (None, 16, 16, 64) 0 batch_normalization_5[0][0]
__________________________________________________________________________________________________
activation_7 (Activation) (None, 16, 16, 64) 0 batch_normalization_7[0][0]
__________________________________________________________________________________________________
activation_10 (Activation) (None, 16, 16, 96) 0 batch_normalization_10[0][0]
__________________________________________________________________________________________________
activation_11 (Activation) (None, 16, 16, 32) 0 batch_normalization_11[0][0]
__________________________________________________________________________________________________
mixed0 (Concatenate) (None, 16, 16, 256) 0 activation_5[0][0]
activation_7[0][0]
activation_10[0][0]
activation_11[0][0]
__________________________________________________________________________________________________
conv2d_15 (Conv2D) (None, 16, 16, 64) 16384 mixed0[0][0]
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 16, 16, 64) 192 conv2d_15[0][0]
__________________________________________________________________________________________________
activation_15 (Activation) (None, 16, 16, 64) 0 batch_normalization_15[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 16, 16, 48) 12288 mixed0[0][0]
__________________________________________________________________________________________________
conv2d_16 (Conv2D) (None, 16, 16, 96) 55296 activation_15[0][0]
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 16, 16, 48) 144 conv2d_13[0][0]
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 16, 16, 96) 288 conv2d_16[0][0]
__________________________________________________________________________________________________
activation_13 (Activation) (None, 16, 16, 48) 0 batch_normalization_13[0][0]
__________________________________________________________________________________________________
activation_16 (Activation) (None, 16, 16, 96) 0 batch_normalization_16[0][0]
__________________________________________________________________________________________________
average_pooling2d_1 (AveragePoo (None, 16, 16, 256) 0 mixed0[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 16, 16, 64) 16384 mixed0[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 16, 16, 64) 76800 activation_13[0][0]
__________________________________________________________________________________________________
conv2d_17 (Conv2D) (None, 16, 16, 96) 82944 activation_16[0][0]
__________________________________________________________________________________________________
conv2d_18 (Conv2D) (None, 16, 16, 64) 16384 average_pooling2d_1[0][0]
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 16, 16, 64) 192 conv2d_12[0][0]
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 16, 16, 64) 192 conv2d_14[0][0]
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 16, 16, 96) 288 conv2d_17[0][0]
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 16, 16, 64) 192 conv2d_18[0][0]
__________________________________________________________________________________________________
activation_12 (Activation) (None, 16, 16, 64) 0 batch_normalization_12[0][0]
__________________________________________________________________________________________________
activation_14 (Activation) (None, 16, 16, 64) 0 batch_normalization_14[0][0]
__________________________________________________________________________________________________
activation_17 (Activation) (None, 16, 16, 96) 0 batch_normalization_17[0][0]
__________________________________________________________________________________________________
activation_18 (Activation) (None, 16, 16, 64) 0 batch_normalization_18[0][0]
__________________________________________________________________________________________________
mixed1 (Concatenate) (None, 16, 16, 288) 0 activation_12[0][0]
activation_14[0][0]
activation_17[0][0]
activation_18[0][0]
__________________________________________________________________________________________________
conv2d_22 (Conv2D) (None, 16, 16, 64) 18432 mixed1[0][0]
__________________________________________________________________________________________________
batch_normalization_22 (BatchNo (None, 16, 16, 64) 192 conv2d_22[0][0]
__________________________________________________________________________________________________
activation_22 (Activation) (None, 16, 16, 64) 0 batch_normalization_22[0][0]
__________________________________________________________________________________________________
conv2d_20 (Conv2D) (None, 16, 16, 48) 13824 mixed1[0][0]
__________________________________________________________________________________________________
conv2d_23 (Conv2D) (None, 16, 16, 96) 55296 activation_22[0][0]
__________________________________________________________________________________________________
batch_normalization_20 (BatchNo (None, 16, 16, 48) 144 conv2d_20[0][0]
__________________________________________________________________________________________________
batch_normalization_23 (BatchNo (None, 16, 16, 96) 288 conv2d_23[0][0]
__________________________________________________________________________________________________
activation_20 (Activation) (None, 16, 16, 48) 0 batch_normalization_20[0][0]
__________________________________________________________________________________________________
activation_23 (Activation) (None, 16, 16, 96) 0 batch_normalization_23[0][0]
__________________________________________________________________________________________________
average_pooling2d_2 (AveragePoo (None, 16, 16, 288) 0 mixed1[0][0]
__________________________________________________________________________________________________
conv2d_19 (Conv2D) (None, 16, 16, 64) 18432 mixed1[0][0]
__________________________________________________________________________________________________
conv2d_21 (Conv2D) (None, 16, 16, 64) 76800 activation_20[0][0]
__________________________________________________________________________________________________
conv2d_24 (Conv2D) (None, 16, 16, 96) 82944 activation_23[0][0]
__________________________________________________________________________________________________
conv2d_25 (Conv2D) (None, 16, 16, 64) 18432 average_pooling2d_2[0][0]
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 16, 16, 64) 192 conv2d_19[0][0]
__________________________________________________________________________________________________
batch_normalization_21 (BatchNo (None, 16, 16, 64) 192 conv2d_21[0][0]
__________________________________________________________________________________________________
batch_normalization_24 (BatchNo (None, 16, 16, 96) 288 conv2d_24[0][0]
__________________________________________________________________________________________________
batch_normalization_25 (BatchNo (None, 16, 16, 64) 192 conv2d_25[0][0]
__________________________________________________________________________________________________
activation_19 (Activation) (None, 16, 16, 64) 0 batch_normalization_19[0][0]
__________________________________________________________________________________________________
activation_21 (Activation) (None, 16, 16, 64) 0 batch_normalization_21[0][0]
__________________________________________________________________________________________________
activation_24 (Activation) (None, 16, 16, 96) 0 batch_normalization_24[0][0]
__________________________________________________________________________________________________
activation_25 (Activation) (None, 16, 16, 64) 0 batch_normalization_25[0][0]
__________________________________________________________________________________________________
mixed2 (Concatenate) (None, 16, 16, 288) 0 activation_19[0][0]
activation_21[0][0]
activation_24[0][0]
activation_25[0][0]
__________________________________________________________________________________________________
conv2d_27 (Conv2D) (None, 16, 16, 64) 18432 mixed2[0][0]
__________________________________________________________________________________________________
batch_normalization_27 (BatchNo (None, 16, 16, 64) 192 conv2d_27[0][0]
__________________________________________________________________________________________________
activation_27 (Activation) (None, 16, 16, 64) 0 batch_normalization_27[0][0]
__________________________________________________________________________________________________
conv2d_28 (Conv2D) (None, 16, 16, 96) 55296 activation_27[0][0]
__________________________________________________________________________________________________
batch_normalization_28 (BatchNo (None, 16, 16, 96) 288 conv2d_28[0][0]
__________________________________________________________________________________________________
activation_28 (Activation) (None, 16, 16, 96) 0 batch_normalization_28[0][0]
__________________________________________________________________________________________________
conv2d_26 (Conv2D) (None, 7, 7, 384) 995328 mixed2[0][0]
__________________________________________________________________________________________________
conv2d_29 (Conv2D) (None, 7, 7, 96) 82944 activation_28[0][0]
__________________________________________________________________________________________________
batch_normalization_26 (BatchNo (None, 7, 7, 384) 1152 conv2d_26[0][0]
__________________________________________________________________________________________________
batch_normalization_29 (BatchNo (None, 7, 7, 96) 288 conv2d_29[0][0]
__________________________________________________________________________________________________
activation_26 (Activation) (None, 7, 7, 384) 0 batch_normalization_26[0][0]
__________________________________________________________________________________________________
activation_29 (Activation) (None, 7, 7, 96) 0 batch_normalization_29[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 7, 7, 288) 0 mixed2[0][0]
__________________________________________________________________________________________________
mixed3 (Concatenate) (None, 7, 7, 768) 0 activation_26[0][0]
activation_29[0][0]
max_pooling2d_2[0][0]
__________________________________________________________________________________________________
conv2d_34 (Conv2D) (None, 7, 7, 128) 98304 mixed3[0][0]
__________________________________________________________________________________________________
batch_normalization_34 (BatchNo (None, 7, 7, 128) 384 conv2d_34[0][0]
__________________________________________________________________________________________________
activation_34 (Activation) (None, 7, 7, 128) 0 batch_normalization_34[0][0]
__________________________________________________________________________________________________
conv2d_35 (Conv2D) (None, 7, 7, 128) 114688 activation_34[0][0]
__________________________________________________________________________________________________
batch_normalization_35 (BatchNo (None, 7, 7, 128) 384 conv2d_35[0][0]
__________________________________________________________________________________________________
activation_35 (Activation) (None, 7, 7, 128) 0 batch_normalization_35[0][0]
__________________________________________________________________________________________________
conv2d_31 (Conv2D) (None, 7, 7, 128) 98304 mixed3[0][0]
__________________________________________________________________________________________________
conv2d_36 (Conv2D) (None, 7, 7, 128) 114688 activation_35[0][0]
__________________________________________________________________________________________________
batch_normalization_31 (BatchNo (None, 7, 7, 128) 384 conv2d_31[0][0]
__________________________________________________________________________________________________
batch_normalization_36 (BatchNo (None, 7, 7, 128) 384 conv2d_36[0][0]
__________________________________________________________________________________________________
activation_31 (Activation) (None, 7, 7, 128) 0 batch_normalization_31[0][0]
__________________________________________________________________________________________________
activation_36 (Activation) (None, 7, 7, 128) 0 batch_normalization_36[0][0]
__________________________________________________________________________________________________
conv2d_32 (Conv2D) (None, 7, 7, 128) 114688 activation_31[0][0]
__________________________________________________________________________________________________
conv2d_37 (Conv2D) (None, 7, 7, 128) 114688 activation_36[0][0]
__________________________________________________________________________________________________
batch_normalization_32 (BatchNo (None, 7, 7, 128) 384 conv2d_32[0][0]
__________________________________________________________________________________________________
batch_normalization_37 (BatchNo (None, 7, 7, 128) 384 conv2d_37[0][0]
__________________________________________________________________________________________________
activation_32 (Activation) (None, 7, 7, 128) 0 batch_normalization_32[0][0]
__________________________________________________________________________________________________
activation_37 (Activation) (None, 7, 7, 128) 0 batch_normalization_37[0][0]
__________________________________________________________________________________________________
average_pooling2d_3 (AveragePoo (None, 7, 7, 768) 0 mixed3[0][0]
__________________________________________________________________________________________________
conv2d_30 (Conv2D) (None, 7, 7, 192) 147456 mixed3[0][0]
__________________________________________________________________________________________________
conv2d_33 (Conv2D) (None, 7, 7, 192) 172032 activation_32[0][0]
__________________________________________________________________________________________________
conv2d_38 (Conv2D) (None, 7, 7, 192) 172032 activation_37[0][0]
__________________________________________________________________________________________________
conv2d_39 (Conv2D) (None, 7, 7, 192) 147456 average_pooling2d_3[0][0]
__________________________________________________________________________________________________
batch_normalization_30 (BatchNo (None, 7, 7, 192) 576 conv2d_30[0][0]
__________________________________________________________________________________________________
batch_normalization_33 (BatchNo (None, 7, 7, 192) 576 conv2d_33[0][0]
__________________________________________________________________________________________________
batch_normalization_38 (BatchNo (None, 7, 7, 192) 576 conv2d_38[0][0]
__________________________________________________________________________________________________
batch_normalization_39 (BatchNo (None, 7, 7, 192) 576 conv2d_39[0][0]
__________________________________________________________________________________________________
activation_30 (Activation) (None, 7, 7, 192) 0 batch_normalization_30[0][0]
__________________________________________________________________________________________________
activation_33 (Activation) (None, 7, 7, 192) 0 batch_normalization_33[0][0]
__________________________________________________________________________________________________
activation_38 (Activation) (None, 7, 7, 192) 0 batch_normalization_38[0][0]
__________________________________________________________________________________________________
activation_39 (Activation) (None, 7, 7, 192) 0 batch_normalization_39[0][0]
__________________________________________________________________________________________________
mixed4 (Concatenate) (None, 7, 7, 768) 0 activation_30[0][0]
activation_33[0][0]
activation_38[0][0]
activation_39[0][0]
__________________________________________________________________________________________________
conv2d_44 (Conv2D) (None, 7, 7, 160) 122880 mixed4[0][0]
__________________________________________________________________________________________________
batch_normalization_44 (BatchNo (None, 7, 7, 160) 480 conv2d_44[0][0]
__________________________________________________________________________________________________
activation_44 (Activation) (None, 7, 7, 160) 0 batch_normalization_44[0][0]
__________________________________________________________________________________________________
conv2d_45 (Conv2D) (None, 7, 7, 160) 179200 activation_44[0][0]
__________________________________________________________________________________________________
batch_normalization_45 (BatchNo (None, 7, 7, 160) 480 conv2d_45[0][0]
__________________________________________________________________________________________________
activation_45 (Activation) (None, 7, 7, 160) 0 batch_normalization_45[0][0]
__________________________________________________________________________________________________
conv2d_41 (Conv2D) (None, 7, 7, 160) 122880 mixed4[0][0]
__________________________________________________________________________________________________
conv2d_46 (Conv2D) (None, 7, 7, 160) 179200 activation_45[0][0]
__________________________________________________________________________________________________
batch_normalization_41 (BatchNo (None, 7, 7, 160) 480 conv2d_41[0][0]
__________________________________________________________________________________________________
batch_normalization_46 (BatchNo (None, 7, 7, 160) 480 conv2d_46[0][0]
__________________________________________________________________________________________________
activation_41 (Activation) (None, 7, 7, 160) 0 batch_normalization_41[0][0]
__________________________________________________________________________________________________
activation_46 (Activation) (None, 7, 7, 160) 0 batch_normalization_46[0][0]
__________________________________________________________________________________________________
conv2d_42 (Conv2D) (None, 7, 7, 160) 179200 activation_41[0][0]
__________________________________________________________________________________________________
conv2d_47 (Conv2D) (None, 7, 7, 160) 179200 activation_46[0][0]
__________________________________________________________________________________________________
batch_normalization_42 (BatchNo (None, 7, 7, 160) 480 conv2d_42[0][0]
__________________________________________________________________________________________________
batch_normalization_47 (BatchNo (None, 7, 7, 160) 480 conv2d_47[0][0]
__________________________________________________________________________________________________
activation_42 (Activation) (None, 7, 7, 160) 0 batch_normalization_42[0][0]
__________________________________________________________________________________________________
activation_47 (Activation) (None, 7, 7, 160) 0 batch_normalization_47[0][0]
__________________________________________________________________________________________________
average_pooling2d_4 (AveragePoo (None, 7, 7, 768) 0 mixed4[0][0]
__________________________________________________________________________________________________
conv2d_40 (Conv2D) (None, 7, 7, 192) 147456 mixed4[0][0]
__________________________________________________________________________________________________
conv2d_43 (Conv2D) (None, 7, 7, 192) 215040 activation_42[0][0]
__________________________________________________________________________________________________
conv2d_48 (Conv2D) (None, 7, 7, 192) 215040 activation_47[0][0]
__________________________________________________________________________________________________
conv2d_49 (Conv2D) (None, 7, 7, 192) 147456 average_pooling2d_4[0][0]
__________________________________________________________________________________________________
batch_normalization_40 (BatchNo (None, 7, 7, 192) 576 conv2d_40[0][0]
__________________________________________________________________________________________________
batch_normalization_43 (BatchNo (None, 7, 7, 192) 576 conv2d_43[0][0]
__________________________________________________________________________________________________
batch_normalization_48 (BatchNo (None, 7, 7, 192) 576 conv2d_48[0][0]
__________________________________________________________________________________________________
batch_normalization_49 (BatchNo (None, 7, 7, 192) 576 conv2d_49[0][0]
__________________________________________________________________________________________________
activation_40 (Activation) (None, 7, 7, 192) 0 batch_normalization_40[0][0]
__________________________________________________________________________________________________
activation_43 (Activation) (None, 7, 7, 192) 0 batch_normalization_43[0][0]
__________________________________________________________________________________________________
activation_48 (Activation) (None, 7, 7, 192) 0 batch_normalization_48[0][0]
__________________________________________________________________________________________________
activation_49 (Activation) (None, 7, 7, 192) 0 batch_normalization_49[0][0]
__________________________________________________________________________________________________
mixed5 (Concatenate) (None, 7, 7, 768) 0 activation_40[0][0]
activation_43[0][0]
activation_48[0][0]
activation_49[0][0]
__________________________________________________________________________________________________
conv2d_54 (Conv2D) (None, 7, 7, 160) 122880 mixed5[0][0]
__________________________________________________________________________________________________
batch_normalization_54 (BatchNo (None, 7, 7, 160) 480 conv2d_54[0][0]
__________________________________________________________________________________________________
activation_54 (Activation) (None, 7, 7, 160) 0 batch_normalization_54[0][0]
__________________________________________________________________________________________________
conv2d_55 (Conv2D) (None, 7, 7, 160) 179200 activation_54[0][0]
__________________________________________________________________________________________________
batch_normalization_55 (BatchNo (None, 7, 7, 160) 480 conv2d_55[0][0]
__________________________________________________________________________________________________
activation_55 (Activation) (None, 7, 7, 160) 0 batch_normalization_55[0][0]
__________________________________________________________________________________________________
conv2d_51 (Conv2D) (None, 7, 7, 160) 122880 mixed5[0][0]
__________________________________________________________________________________________________
conv2d_56 (Conv2D) (None, 7, 7, 160) 179200 activation_55[0][0]
__________________________________________________________________________________________________
batch_normalization_51 (BatchNo (None, 7, 7, 160) 480 conv2d_51[0][0]
__________________________________________________________________________________________________
batch_normalization_56 (BatchNo (None, 7, 7, 160) 480 conv2d_56[0][0]
__________________________________________________________________________________________________
activation_51 (Activation) (None, 7, 7, 160) 0 batch_normalization_51[0][0]
__________________________________________________________________________________________________
activation_56 (Activation) (None, 7, 7, 160) 0 batch_normalization_56[0][0]
__________________________________________________________________________________________________
conv2d_52 (Conv2D) (None, 7, 7, 160) 179200 activation_51[0][0]
__________________________________________________________________________________________________
conv2d_57 (Conv2D) (None, 7, 7, 160) 179200 activation_56[0][0]
__________________________________________________________________________________________________
batch_normalization_52 (BatchNo (None, 7, 7, 160) 480 conv2d_52[0][0]
__________________________________________________________________________________________________
batch_normalization_57 (BatchNo (None, 7, 7, 160) 480 conv2d_57[0][0]
__________________________________________________________________________________________________
activation_52 (Activation) (None, 7, 7, 160) 0 batch_normalization_52[0][0]
__________________________________________________________________________________________________
activation_57 (Activation) (None, 7, 7, 160) 0 batch_normalization_57[0][0]
__________________________________________________________________________________________________
average_pooling2d_5 (AveragePoo (None, 7, 7, 768) 0 mixed5[0][0]
__________________________________________________________________________________________________
conv2d_50 (Conv2D) (None, 7, 7, 192) 147456 mixed5[0][0]
__________________________________________________________________________________________________
conv2d_53 (Conv2D) (None, 7, 7, 192) 215040 activation_52[0][0]
__________________________________________________________________________________________________
conv2d_58 (Conv2D) (None, 7, 7, 192) 215040 activation_57[0][0]
__________________________________________________________________________________________________
conv2d_59 (Conv2D) (None, 7, 7, 192) 147456 average_pooling2d_5[0][0]
__________________________________________________________________________________________________
batch_normalization_50 (BatchNo (None, 7, 7, 192) 576 conv2d_50[0][0]
__________________________________________________________________________________________________
batch_normalization_53 (BatchNo (None, 7, 7, 192) 576 conv2d_53[0][0]
__________________________________________________________________________________________________
batch_normalization_58 (BatchNo (None, 7, 7, 192) 576 conv2d_58[0][0]
__________________________________________________________________________________________________
batch_normalization_59 (BatchNo (None, 7, 7, 192) 576 conv2d_59[0][0]
__________________________________________________________________________________________________
activation_50 (Activation) (None, 7, 7, 192) 0 batch_normalization_50[0][0]
__________________________________________________________________________________________________
activation_53 (Activation) (None, 7, 7, 192) 0 batch_normalization_53[0][0]
__________________________________________________________________________________________________
activation_58 (Activation) (None, 7, 7, 192) 0 batch_normalization_58[0][0]
__________________________________________________________________________________________________
activation_59 (Activation) (None, 7, 7, 192) 0 batch_normalization_59[0][0]
__________________________________________________________________________________________________
mixed6 (Concatenate) (None, 7, 7, 768) 0 activation_50[0][0]
activation_53[0][0]
activation_58[0][0]
activation_59[0][0]
__________________________________________________________________________________________________
conv2d_64 (Conv2D) (None, 7, 7, 192) 147456 mixed6[0][0]
__________________________________________________________________________________________________
batch_normalization_64 (BatchNo (None, 7, 7, 192) 576 conv2d_64[0][0]
__________________________________________________________________________________________________
activation_64 (Activation) (None, 7, 7, 192) 0 batch_normalization_64[0][0]
__________________________________________________________________________________________________
conv2d_65 (Conv2D) (None, 7, 7, 192) 258048 activation_64[0][0]
__________________________________________________________________________________________________
batch_normalization_65 (BatchNo (None, 7, 7, 192) 576 conv2d_65[0][0]
__________________________________________________________________________________________________
activation_65 (Activation) (None, 7, 7, 192) 0 batch_normalization_65[0][0]
__________________________________________________________________________________________________
conv2d_61 (Conv2D) (None, 7, 7, 192) 147456 mixed6[0][0]
__________________________________________________________________________________________________
conv2d_66 (Conv2D) (None, 7, 7, 192) 258048 activation_65[0][0]
__________________________________________________________________________________________________
batch_normalization_61 (BatchNo (None, 7, 7, 192) 576 conv2d_61[0][0]
__________________________________________________________________________________________________
batch_normalization_66 (BatchNo (None, 7, 7, 192) 576 conv2d_66[0][0]
__________________________________________________________________________________________________
activation_61 (Activation) (None, 7, 7, 192) 0 batch_normalization_61[0][0]
__________________________________________________________________________________________________
activation_66 (Activation) (None, 7, 7, 192) 0 batch_normalization_66[0][0]
__________________________________________________________________________________________________
conv2d_62 (Conv2D) (None, 7, 7, 192) 258048 activation_61[0][0]
__________________________________________________________________________________________________
conv2d_67 (Conv2D) (None, 7, 7, 192) 258048 activation_66[0][0]
__________________________________________________________________________________________________
batch_normalization_62 (BatchNo (None, 7, 7, 192) 576 conv2d_62[0][0]
__________________________________________________________________________________________________
batch_normalization_67 (BatchNo (None, 7, 7, 192) 576 conv2d_67[0][0]
__________________________________________________________________________________________________
activation_62 (Activation) (None, 7, 7, 192) 0 batch_normalization_62[0][0]
__________________________________________________________________________________________________
activation_67 (Activation) (None, 7, 7, 192) 0 batch_normalization_67[0][0]
__________________________________________________________________________________________________
average_pooling2d_6 (AveragePoo (None, 7, 7, 768) 0 mixed6[0][0]
__________________________________________________________________________________________________
conv2d_60 (Conv2D) (None, 7, 7, 192) 147456 mixed6[0][0]
__________________________________________________________________________________________________
conv2d_63 (Conv2D) (None, 7, 7, 192) 258048 activation_62[0][0]
__________________________________________________________________________________________________
conv2d_68 (Conv2D) (None, 7, 7, 192) 258048 activation_67[0][0]
__________________________________________________________________________________________________
conv2d_69 (Conv2D) (None, 7, 7, 192) 147456 average_pooling2d_6[0][0]
__________________________________________________________________________________________________
batch_normalization_60 (BatchNo (None, 7, 7, 192) 576 conv2d_60[0][0]
__________________________________________________________________________________________________
batch_normalization_63 (BatchNo (None, 7, 7, 192) 576 conv2d_63[0][0]
__________________________________________________________________________________________________
batch_normalization_68 (BatchNo (None, 7, 7, 192) 576 conv2d_68[0][0]
__________________________________________________________________________________________________
batch_normalization_69 (BatchNo (None, 7, 7, 192) 576 conv2d_69[0][0]
__________________________________________________________________________________________________
activation_60 (Activation) (None, 7, 7, 192) 0 batch_normalization_60[0][0]
__________________________________________________________________________________________________
activation_63 (Activation) (None, 7, 7, 192) 0 batch_normalization_63[0][0]
__________________________________________________________________________________________________
activation_68 (Activation) (None, 7, 7, 192) 0 batch_normalization_68[0][0]
__________________________________________________________________________________________________
activation_69 (Activation) (None, 7, 7, 192) 0 batch_normalization_69[0][0]
__________________________________________________________________________________________________
mixed7 (Concatenate) (None, 7, 7, 768) 0 activation_60[0][0]
activation_63[0][0]
activation_68[0][0]
activation_69[0][0]
__________________________________________________________________________________________________
flatten_1 (Flatten) (None, 37632) 0 mixed7[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) (None, 1024) 38536192 flatten_1[0][0]
__________________________________________________________________________________________________
dense_3 (Dense) (None, 1) 1025 dense_2[0][0]
==================================================================================================
Total params: 47,512,481
Trainable params: 38,537,217
Non-trainable params: 8,975,264
__________________________________________________________________________________________________
None
这么深的网络，要自己跑得跑多久呀，想想都可怕，幸亏有开源的模型参数。我们只需导入，然后按层自己修改即可。
训练—1
模型既然已经写好了，那就开始训练吧！
history = model.fit_generator(
train_generator,
validation_data = validation_generator,
steps_per_epoch = 100,
epochs = 100,
validation_steps = 50)
结果：
查看训练集和验证集准确率的关系。
通过上面，我们了解了迁移学习，但当我们再次训练猫狗照片时，由上图可知，虽然我们进行了图像增强，训练集的准确率在升高，但验证集的准确率却在不断下降，我们还是造成了过拟合，现在，就来看看另一种解决过拟合的方法dropout（随机节点删除），它的思想是，因为网络中最终具有相似的权重，并且可能会相互影响，导致过拟合。对于一个大型的模型，这肯定是一个麻烦，所以我们可以通过随机删除一些节点，这样可以使相邻节点之间的影响减少，消除过拟合。下面就是在代码中实现的具体过程。
模型构造—2
x = layers.Flatten()(last_output)
#print(x.shape)
x = layers.Dense(1024, activation='relu')(x)
#print(x.shape)
#随机节点删除 20%
x = layers.Dropout(0.2)(x)
x = layers.Dense (1, activation='sigmoid')(x)
model = Model( pre_trained_model.input, x)
model.compile(optimizer = RMSprop(lr=0.0001),
loss = 'binary_crossentropy',
metrics = ['acc'])
print(model.summary())
看一下全连接层的结构。
结果：
训练—2
开始训练新模型
history_drop = model.fit_generator(
train_generator,
validation_data = validation_generator,
steps_per_epoch = 100,
epochs = 100,
validation_steps = 50)
结果：
再次看一下训练集和验证集准确率的关系。
acc = history_drop.history['acc']
val_acc = history_drop.history['val_acc']
loss = history_drop.history['loss']
val_loss = history_drop.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, acc, 'r', label='Training accuracy')
plt.plot(epochs, val_acc, 'b', label='Validation accuracy')
plt.title('Training and validation accuracy')
plt.legend(loc=0)
plt.figure()
plt.show()
结果：
可以看到，训练集验证集的准确度都在不断上升，并无过拟合的特征。是个很好的模型。(☆▽☆)
过拟合
既然讲到了过拟合的原因，不妨看看，造成它的其它原因：
- 建模样本选取有误，如样本数量太少，选样方法错误，样本标签错误等，导致选取的样本数据不足以代表预定的分类规则
- 样本噪音干扰过大，使得机器将部分噪音认为是特征从而扰乱了预设的分类规则
- 假设的模型无法合理存在，或者说是假设成立的条件实际并不成立
- 参数太多，模型复杂度过高
- 对于决策树模型，如果我们对于其生长没有合理的限制，其自由生长有可能使节点只包含单纯的事件数据(event)或非事件数据(no
event)，使其虽然可以完美匹配（拟合）训练数据，但是无法适应其他数据集 - 对于神经网络模型：a)对样本数据可能存在分类决策面不唯一，随着学习的进行,，BP算法使权值可能收敛过于复杂的决策面；b)权值学习迭代次数足够多(Overtraining)，拟合了训练数据中的噪声和训练样例中没有代表性的特征
过拟合的解决方案
三、多分类的模型构造
我们上面讲到了迁移学习，解决过拟合的内容。但那些只是涉及到两个分类的情况，那多个分类的情况要怎么办呢？其实很简单，所以也放在了这章讲解。
获取授权码，加载谷歌网盘（ Colab / pycharm可省略）
from google.colab import drive
drive.mount('/content/drive')
这次，我们要以识别石头、剪刀、布三类来详细讲解一下多分类模型。
导入相关包
import os
import zipfile
数据处理
首先，下载数据集，解压到相关目录。（下载处在置顶的资料下载）
local_zip = '/content/drive/My Drive/Colab Notebooks/DateSet/rps.zip'
zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('/content/drive/My Drive/Colab Notebooks/tmp')
zip_ref.close()
local_zip = '/content/drive/My Drive/Colab Notebooks/DateSet/rps-test-set.zip'
zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('/content/drive/My Drive/Colab Notebooks/tmp')
zip_ref.close()
定义根目录。同时也查看一下各数据集的数据个数，及命名规则。
rock_dir = os.path.join('/content/drive/My Drive/Colab Notebooks/tmp/rps/rock')
papper_dir = os.path.join('/content/drive/My Drive/Colab Notebooks/tmp/rps/paper')
scissors_dir = os.path.join('/content/drive/My Drive/Colab Notebooks/tmp/rps/scissors')
#查看一下数据大小
print("rock numbers:",len(os.listdir(rock_dir)))
print("papper numbers:",len(os.listdir(papper_dir)))
print("scissors numbers:",len(os.listdir(scissors_dir)))
#查看前10个数据的命名规则
rock_files = os.listdir(rock_dir)
print(rock_files[:10])
papper_files = os.listdir(papper_dir)
print(papper_files[:10])
scissors_files = os.listdir(scissors_dir)
print(scissors_files[:10])
结果：
以图片的形式查看一下
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
pic_index = 2
next_rock = [os.path.join(rock_dir,fname) for fname in rock_files[pic_index-2:pic_index]]
next_papper = [os.path.join(papper_dir,fname) for fname in papper_files[pic_index-2:pic_index]]
next_scissors = [os.path.join(scissors_dir,fname) for fname in scissors_files[pic_index-2:pic_index]]
for i, img_path in enumerate(next_rock+next_papper+next_scissors):
print(img_path)
img = mpimg.imread(img_path)
plt.imshow(img)
plt.axis("off")
plt.show()
输出了几个例图
图像生成器构造
下面就开始构造图像生成器，准备开始训练
#导入相关包
import tensorflow as tf
import keras_preprocessing
from keras_preprocessing import image
from keras_preprocessing.image import ImageDataGenerator
TRIANING_DIR = '/content/drive/My Drive/Colab Notebooks/tmp/rps'
#训练集生成器-----图像增强
train_datagen = ImageDataGenerator(
rescale = 1/255,
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True, #镜像
fill_mode='nearest')
#验证集生成器
VALIDATION_DIR = '/content/drive/My Drive/Colab Notebooks/tmp/rps-test-set'
validation_datagen = ImageDataGenerator(rescale=1/255) #验证集只规范化
#构造生成器
train_generation = train_datagen.flow_from_directory(
TRIANING_DIR,
target_size = (150,150),
batch_size = 30,
class_mode = 'categorical' ) #多分类，模型换为categorical
validation_generation = validation_datagen.flow_from_directory(
VALIDATION_DIR,
target_size = (150,150),
batch_size = 30,
class_mode = 'categorical')
结果：
Found 2520 images belonging to 3 classes.
Found 372 images belonging to 3 classes.
数据处理完毕，开始通过上面处理后的数据进行训练。
模型构造
model = tf.keras.Sequential([
tf.keras.layers.Conv2D(64,(3,3),activation="relu",input_shape = (150,150,3)),
tf.keras.layers.MaxPool2D(2,2),
tf.keras.layers.Conv2D(64,(3,3),activation="relu"),
tf.keras.layers.MaxPool2D(2,2),
tf.keras.layers.Conv2D(128,(3,3),activation="relu"),
tf.keras.layers.MaxPool2D(2,2),
tf.keras.layers.Conv2D(128,(3,3),activation="relu"),
tf.keras.layers.MaxPool2D(2,2),
tf.keras.layers.Flatten(),
tf.keras.layers.Dropout(0.5), #随机删除节点
tf.keras.layers.Dense(512,activation="relu"),
tf.keras.layers.Dense(3,activation="softmax")])
#查看网络结构
model.summary()
model.compile(loss='categorical_crossentropy',optimizer='rmsprop',metrics=['accuracy'])
### 模型构造
history = model.fit_generator(train_generation,
steps_per_epoch=90,
epochs=25,
validation_data = validation_generation,
validation_steps = 10)
保存模型参数
#将训练好的模型保存下来，减少以后训练。这不就是上面我们用到过的.h5文件嘛o(￣▽￣)o
model.save("/content/drive/My Drive/Colab Notebooks/rps.h5")
绘制训练集和验证集准确率的关系。
import matplotlib.pyplot as plt
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, acc, 'r', label='Training accuracy')
plt.plot(epochs, val_acc, 'b', label='Validation accuracy')
plt.title('Training and validation accuracy')
plt.legend(loc=0)
plt.figure()
plt.show()
运行模型
运行程序测试一下，从未出现过的图片在模型中的输出结果。
import numpy as np
from google.colab import files
from keras.preprocessing import image
uploaded = files.upload()
for fn in uploaded.keys():
# predicting images
path = fn
img = image.load_img(path, target_size=(150, 150))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
images = np.vstack([x])
classes = model.predict(images, batch_size=10)
print(fn)
print(classes)
我输入的图像如下。
结果：
模型在处理时，已按文件夹的位置给了他们相应的标签。比如我这
是这个顺序的文件夹，那么就有了papper—1;rock—2,scissor—3。再看我们的结果。aaa是石头第二个为1；bbb是布，第一个为1；ccc是剪刀，第三个为1。
总结
好了，这一部分到这我们就完成了。我们在这部分主要讲解了在相对复杂的情况下（图像彩色，大小，形状不一）的卷积神经网络的使用，也讲解了使用图片/数据增强，Dropout(随机节点删除)的方法解决过拟合问题。还讲了使用他人模型，或自己之前模型时，可以直接载入，能节省大量时间的迁移学习。这些都是对图像的处理，因为图像是有像素值的，我们直接使用像素值当作数据来处理图像，那么文字呢？下一部分，我们就要讲解一下自然语言怎么被当作数据来进行处理！下章见！！
随后我会更新课后练习，附上参考代码。需要各位自练呦o(￣▽￣)o。
- 博主水平有限，如有错误，请不吝指教，谢谢！(❤ ω ❤)
- 喜欢的请点赞呀(★ ω ★)
835
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
