tensorflow-卷积神经网络-图像分类入门demo

猫狗识别

• 数据预处理：图像数据处理，准备训练和验证数据集
• 卷积网络模型：构建网络架构
• 过拟合问题：观察训练和验证效果，针对过拟合问题提出解决方法
• 数据增强：图像数据增强方法与效果
• 迁移学习：深度学习必备训练策略

导入工具包

import os
import warnings
warnings.filterwarnings("ignore")
import tensorflow as tf
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.preprocessing.image import ImageDataGenerator

指定好数据路径（训练和验证）

# 数据所在文件夹
base_dir = './data/cats_and_dogs'
train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')

# 训练集
train_cats_dir = os.path.join(train_dir, 'cats')
train_dogs_dir = os.path.join(train_dir, 'dogs')

# 验证集
validation_cats_dir = os.path.join(validation_dir, 'cats')
validation_dogs_dir = os.path.join(validation_dir, 'dogs')

构建卷积神经网络模型

• 几层都可以，大家可以随意玩
• 如果用CPU训练，可以把输入设置的更小一些，一般输入大小更主要的决定了训练速度

• model = tf.keras.models.Sequential([
      #如果训练慢，可以把数据设置的更小一些
      tf.keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=(64, 64, 3)),
      tf.keras.layers.MaxPooling2D(2, 2),
  
      tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
      tf.keras.layers.MaxPooling2D(2,2),
  
      tf.keras.layers.Conv2D(128, (3,3), activation='relu'),
      tf.keras.layers.MaxPooling2D(2,2),
      
      #为全连接层准备
      tf.keras.layers.Flatten(),
      
      tf.keras.layers.Dense(512, activation='relu'),
      # 二分类sigmoid就够了
      tf.keras.layers.Dense(1, activation='sigmoid')
  ])

  model.summary()

  

• 配置训练器

  model.compile(loss='binary_crossentropy',
                optimizer=Adam(lr=1e-4),
                metrics=['acc'])

  数据预处理

• 读进来的数据会被自动转换成tensor(float32)格式，分别准备训练和验证
• 图像数据归一化（0-1）区间

  train_datagen = ImageDataGenerator(rescale=1./255)
  test_datagen = ImageDataGenerator(rescale=1./255)

  train_generator = train_datagen.flow_from_directory(
          train_dir,  # 文件夹路径
          target_size=(64, 64),  # 指定resize成的大小
          batch_size=20,
          # 如果one-hot就是categorical，二分类用binary就可以
          class_mode='binary')
  
  validation_generator = test_datagen.flow_from_directory(
          validation_dir,
          target_size=(64, 64),
          batch_size=20,
          class_mode='binary')

  训练网络模型

• 直接fit也可以，但是通常咱们不能把所有数据全部放入内存，fit_generator相当于一个生成器，动态产生所需的batch数据
• steps_per_epoch相当给定一个停止条件，因为生成器会不断产生batch数据，说白了就是它不知道一个epoch里需要执行多少个step

  history = model.fit_generator(
        train_generator,
        steps_per_epoch=100,  # 2000 images = batch_size * steps
        epochs=20,
        validation_data=validation_generator,
        validation_steps=50,  # 1000 images = batch_size * steps
        verbose=2)

  Epoch 1/20
  100/100 - 7s - loss: 0.6892 - acc: 0.5325 - val_loss: 0.6705 - val_acc: 0.5970
  Epoch 2/20
  100/100 - 6s - loss: 0.6595 - acc: 0.6055 - val_loss: 0.6346 - val_acc: 0.6470
  Epoch 3/20
  100/100 - 6s - loss: 0.6350 - acc: 0.6515 - val_loss: 0.6358 - val_acc: 0.6320
  Epoch 4/20
  100/100 - 7s - loss: 0.5936 - acc: 0.6865 - val_loss: 0.5906 - val_acc: 0.6780
  Epoch 5/20
  100/100 - 7s - loss: 0.5530 - acc: 0.7170 - val_loss: 0.5978 - val_acc: 0.6670
  Epoch 6/20
  100/100 - 8s - loss: 0.5179 - acc: 0.7490 - val_loss: 0.5484 - val_acc: 0.7140
  Epoch 7/20
  100/100 - 8s - loss: 0.4854 - acc: 0.7725 - val_loss: 0.5686 - val_acc: 0.7080
  Epoch 8/20
  100/100 - 8s - loss: 0.4595 - acc: 0.7905 - val_loss: 0.5452 - val_acc: 0.7150
  Epoch 9/20
  100/100 - 8s - loss: 0.4406 - acc: 0.7885 - val_loss: 0.5453 - val_acc: 0.7210
  Epoch 10/20
  100/100 - 7s - loss: 0.4109 - acc: 0.8170 - val_loss: 0.5317 - val_acc: 0.7270
  Epoch 11/20
  100/100 - 8s - loss: 0.3892 - acc: 0.8285 - val_loss: 0.5384 - val_acc: 0.7220
  Epoch 12/20
  100/100 - 8s - loss: 0.3542 - acc: 0.8570 - val_loss: 0.5480 - val_acc: 0.7180
  Epoch 13/20
  100/100 - 8s - loss: 0.3421 - acc: 0.8580 - val_loss: 0.5355 - val_acc: 0.7420
  Epoch 14/20
  100/100 - 8s - loss: 0.3217 - acc: 0.8665 - val_loss: 0.5572 - val_acc: 0.7340
  Epoch 15/20
  100/100 - 8s - loss: 0.2931 - acc: 0.8805 - val_loss: 0.5545 - val_acc: 0.7400
  Epoch 16/20
  100/100 - 8s - loss: 0.2739 - acc: 0.8870 - val_loss: 0.5540 - val_acc: 0.7360
  Epoch 17/20
  100/100 - 8s - loss: 0.2535 - acc: 0.9040 - val_loss: 0.5564 - val_acc: 0.7380
  Epoch 18/20
  100/100 - 8s - loss: 0.2257 - acc: 0.9245 - val_loss: 0.5710 - val_acc: 0.7420
  Epoch 19/20
  100/100 - 8s - loss: 0.2084 - acc: 0.9350 - val_loss: 0.5734 - val_acc: 0.7460
  Epoch 20/20
  100/100 - 8s - loss: 0.2258 - acc: 0.9130 - val_loss: 0.5897 - val_acc: 0.7300
  

  效果展示

  import matplotlib.pyplot as plt
  acc = history.history['acc']
  val_acc = history.history['val_acc']
  loss = history.history['loss']
  val_loss = history.history['val_loss']
  
  epochs = range(len(acc))
  
  plt.plot(epochs, acc, 'bo', label='Training accuracy')
  plt.plot(epochs, val_acc, 'b', label='Validation accuracy')
  plt.title('Training and validation accuracy')
  
  plt.figure()
  
  plt.plot(epochs, loss, 'bo', label='Training Loss')
  plt.plot(epochs, val_loss, 'b', label='Validation Loss')
  plt.title('Training and validation loss')
  plt.legend()
  
  plt.show()

  

•