本次所用數據來自ImageNet，使用預訓練好的數據來預測一個新的數據集：貓狗圖片分類。這里，使用VGG模型，這個模型內置在Keras中，直接導入就可以了。

from keras.applications import VGG16conv_base = VGG16(weights='imagenet',include_top=False,input_shape=(150, 150, 3))

說一下這三個參數：

• weights：指定模型初始化權重檢查點
• include_top：指定模型最后是否包含密集連接分類器。默認情況下，這個密集連接分類器對應于ImageNet的1000個類別。因為我們打算使用自己的分類器（只有兩個類別：cat和dog），所以不用包含。
• input_shape：輸入到網絡中的圖像張量（可選參數），如果不傳入這個參數，那么網絡可以處理任意形狀的輸入

看一下VGG16網絡的詳細構架：

conv_base.summary() _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_1 (InputLayer) (None, 150, 150, 3) 0 _________________________________________________________________ block1_conv1 (Conv2D) (None, 150, 150, 64) 1792 _________________________________________________________________ block1_conv2 (Conv2D) (None, 150, 150, 64) 36928 _________________________________________________________________ block1_pool (MaxPooling2D) (None, 75, 75, 64) 0 _________________________________________________________________ block2_conv1 (Conv2D) (None, 75, 75, 128) 73856 _________________________________________________________________ block2_conv2 (Conv2D) (None, 75, 75, 128) 147584 _________________________________________________________________ block2_pool (MaxPooling2D) (None, 37, 37, 128) 0 _________________________________________________________________ block3_conv1 (Conv2D) (None, 37, 37, 256) 295168 _________________________________________________________________ block3_conv2 (Conv2D) (None, 37, 37, 256) 590080 _________________________________________________________________ block3_conv3 (Conv2D) (None, 37, 37, 256) 590080 _________________________________________________________________ block3_pool (MaxPooling2D) (None, 18, 18, 256) 0 _________________________________________________________________ block4_conv1 (Conv2D) (None, 18, 18, 512) 1180160 _________________________________________________________________ block4_conv2 (Conv2D) (None, 18, 18, 512) 2359808 _________________________________________________________________ block4_conv3 (Conv2D) (None, 18, 18, 512) 2359808 _________________________________________________________________ block4_pool (MaxPooling2D) (None, 9, 9, 512) 0 _________________________________________________________________ block5_conv1 (Conv2D) (None, 9, 9, 512) 2359808 _________________________________________________________________ block5_conv2 (Conv2D) (None, 9, 9, 512) 2359808 _________________________________________________________________ block5_conv3 (Conv2D) (None, 9, 9, 512) 2359808 _________________________________________________________________ block5_pool (MaxPooling2D) (None, 4, 4, 512) 0 ================================================================= Total params: 14,714,688 Trainable params: 14,714,688 Non-trainable params: 0

最后這個特征圖形狀為（4， 4， 512），我們在這個特征上面添加一個密集連接分類器。

---

不使用數據增強的快速特征提取（計算代價低）

首先，運行ImageDataGenerator實例，將圖像及其標簽提取為Numpy數組，調用conv_base模型的predict方法從這些圖像的中提取特征。

import os import numpy as np from keras.preprocessing.image import ImageDataGeneratorbase_dir = '/Users/fchollet/Downloads/cats_and_dogs_small'train_dir = os.path.join(base_dir, 'train') validation_dir = os.path.join(base_dir, 'validation') test_dir = os.path.join(base_dir, 'test')datagen = ImageDataGenerator(rescale=1./255) batch_size = 20def extract_features(directory, sample_count):features = np.zeros(shape=(sample_count, 4, 4, 512))labels = np.zeros(shape=(sample_count))generator = datagen.flow_from_directory(directory,target_size=(150, 150),batch_size=batch_size,class_mode='binary')i = 0for inputs_batch, labels_batch in generator:features_batch = conv_base.predict(inputs_batch)features[i * batch_size : (i + 1) * batch_size] = features_batchlabels[i * batch_size : (i + 1) * batch_size] = labels_batchi += 1if i * batch_size >= sample_count:break # 這些生成器在循環(huán)中不斷生成數據，所以你必須在讀完所有圖像之后終止循環(huán)return features, labelstrain_features, train_labels = extract_features(train_dir, 2000) validation_features, validation_labels = extract_features(validation_dir, 1000) test_features, test_labels = extract_features(test_dir, 1000)

目前，提取的特征形狀為（samples， 4， 4， 512），我們要將其輸入到密集連接分類器中去，所以必須首先對其形狀展平為（samples ，8192）

train_features = np.reshape(train_features, (2000, 4 * 4 * 512)) validation_features = np.reshape(validation_features, (1000, 4 * 4 * 512)) test_features = np.reshape(test_features, (1000, 4 * 4 * 512))

下面定義一個密集連接分類器，并在剛剛保存好的數據和標簽上訓練分類器：

from keras import models from keras import layers from keras import optimizersmodel = models.Sequential() model.add(layers.Dense(256, activation='relu', input_dim=4 * 4 * 512)) model.add(layers.Dropout(0.5)) model.add(layers.Dense(1, activation='sigmoid'))model.compile(optimizer=optimizers.RMSprop(lr=2e-5),loss='binary_crossentropy',metrics=['acc'])history = model.fit(train_features, train_labels,epochs=30,batch_size=20,validation_data=(validation_features, validation_labels))

訓練速度非常快，因為只需要處理兩個Dense層。下面看一下訓練過程中的損失曲線和精度曲線：

import matplotlib.pyplot as pltacc = 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 acc') plt.plot(epochs, val_acc, 'b', label='Validation acc') plt.title('Training and validation accuracy') plt.legend()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()

從圖中可以看出，驗證精度達到了約90%，比之前從一開始就訓練小型模型效果要好很多，但是從圖中也可以看出，雖然dropout比率比較大，但模型從一開始就出現了過擬合。這是因為本方法中沒有使用數據增強，而數據增強對防止小型圖片數據集過擬合非常重要。

---

使用數據增強的特征提取（計算代價高）

這種方法速度更慢，計算代價更高，但是可以在訓練期間使用數據增強。這種方法是：擴展conv_base模型，然后在輸入數據上端到端的運行模型。（這種方法計算代價很高，必須在GPU上運行）

承接我們之前定義的網絡模型

from keras import models from keras import layersmodel = models.Sequential() model.add(conv_base) model.add(layers.Flatten()) model.add(layers.Dense(256, activation='relu')) model.add(layers.Dense(1, activation='sigmoid')) model.summary() _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= vgg16 (Model) (None, 4, 4, 512) 14714688 _________________________________________________________________ flatten_1 (Flatten) (None, 8192) 0 _________________________________________________________________ dense_3 (Dense) (None, 256) 2097408 _________________________________________________________________ dense_4 (Dense) (None, 1) 257 ================================================================= Total params: 16,812,353 Trainable params: 16,812,353 Non-trainable params: 0

我們可以看到，VGG16的卷積基一共有14714688個參數，其上添加的分類器一共有200萬個參數，非常多。

在編譯和訓練模型之前，需要凍結卷積基。凍結一個或多個層是指在訓練過程中保持其權重不變。如果不這么做，那么卷積基之前學到的表示將會在訓練過程中被修改。因為其上添加的Dense是隨機初始化的，所以非常打的權重更新會在網絡中進行傳播，對之前學到的表示造成很大破壞。

在Keras中，凍結網絡的方法是將其trainable屬性設置為False

print('This is the number of trainable weights ''before freezing the conv base:', len(model.trainable_weights))

This is the number of trainable weights before freezing the conv base: 30

conv_base.trainable = False print('This is the number of trainable weights ''after freezing the conv base:', len(model.trainable_weights))

This is the number of trainable weights after freezing the conv base: 4

如此設置之后，只有添加的兩個Dense層的權重才會被訓練，總共有4個權重張量，每層2個（主權重矩陣和偏置向量），注意的是，如果想修改權重屬性trainable，那么應該修改好屬性之后再編譯模型。

下面，我們可以訓練模型了，并使用數據增強的辦法：

from keras.preprocessing.image import ImageDataGeneratortrain_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,fill_mode='nearest')# Note that the validation data should not be augmented! test_datagen = ImageDataGenerator(rescale=1./255)train_generator = train_datagen.flow_from_directory(# This is the target directorytrain_dir,# All images will be resized to 150x150target_size=(150, 150),batch_size=20,# Since we use binary_crossentropy loss, we need binary labelsclass_mode='binary')validation_generator = test_datagen.flow_from_directory(validation_dir,target_size=(150, 150),batch_size=20,class_mode='binary')model.compile(loss='binary_crossentropy',optimizer=optimizers.RMSprop(lr=2e-5),metrics=['acc'])history = model.fit_generator(train_generator,steps_per_epoch=100,epochs=30,validation_data=validation_generator,validation_steps=50,verbose=2) model.save('cats_and_dogs_small_3.h5')

我們再來看看驗證精度：

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 acc') plt.plot(epochs, val_acc, 'b', label='Validation acc') plt.title('Training and validation accuracy') plt.legend()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()

驗證精度到了將近96%，而且減少了過擬合。

---

微調模型

我們下面使用模型微調，進一步提高模型的性能。模型微調的步驟如下：

• （1）在已經訓練好的基網絡（base network）上添加自定義網絡
• （2）凍結基網絡
• （3）訓練所添加的部分
• （4）解凍基網絡的一些層
• （5）聯合訓練解凍的這些層和添加的部分

在做特征提取的時候已經完成了前三個步驟。我們繼續(xù)第四個步驟，先解凍conv_base，然后凍結其中的部分層。

_________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_1 (InputLayer) (None, 150, 150, 3) 0 _________________________________________________________________ block1_conv1 (Conv2D) (None, 150, 150, 64) 1792 _________________________________________________________________ block1_conv2 (Conv2D) (None, 150, 150, 64) 36928 _________________________________________________________________ block1_pool (MaxPooling2D) (None, 75, 75, 64) 0 _________________________________________________________________ block2_conv1 (Conv2D) (None, 75, 75, 128) 73856 _________________________________________________________________ block2_conv2 (Conv2D) (None, 75, 75, 128) 147584 _________________________________________________________________ block2_pool (MaxPooling2D) (None, 37, 37, 128) 0 _________________________________________________________________ block3_conv1 (Conv2D) (None, 37, 37, 256) 295168 _________________________________________________________________ block3_conv2 (Conv2D) (None, 37, 37, 256) 590080 _________________________________________________________________ block3_conv3 (Conv2D) (None, 37, 37, 256) 590080 _________________________________________________________________ block3_pool (MaxPooling2D) (None, 18, 18, 256) 0 _________________________________________________________________ block4_conv1 (Conv2D) (None, 18, 18, 512) 1180160 _________________________________________________________________ block4_conv2 (Conv2D) (None, 18, 18, 512) 2359808 _________________________________________________________________ block4_conv3 (Conv2D) (None, 18, 18, 512) 2359808 _________________________________________________________________ block4_pool (MaxPooling2D) (None, 9, 9, 512) 0 _________________________________________________________________ block5_conv1 (Conv2D) (None, 9, 9, 512) 2359808 _________________________________________________________________ block5_conv2 (Conv2D) (None, 9, 9, 512) 2359808 _________________________________________________________________ block5_conv3 (Conv2D) (None, 9, 9, 512) 2359808 _________________________________________________________________ block5_pool (MaxPooling2D) (None, 4, 4, 512) 0 ================================================================= Total params: 14,714,688 Trainable params: 14,714,688 Non-trainable params: 0

再回顧一下這些層，我們將微調最后三個卷積層，也就是說，知道block4_pool的所有層都應該被凍結，后面三層來進行訓練。

要知道，訓練的參數越多，過擬合的風險越大。卷積基有1500萬個參數，所以你在小型數據集上訓練這么多參數是有風險的。因此，這種情況下最好的策略是僅微調卷積基最后三兩層。

conv_base.trainable = Trueset_trainable = False for layer in conv_base.layers:if layer.name == 'block5_conv1':set_trainable = Trueif set_trainable:layer.trainable = Trueelse:layer.trainable = False

現在可以微調網絡了，我們將使用學習率非常小的RMSProp優(yōu)化器來實現。之所以讓學習率很小，是因為對于微調網絡的三層表示，我們希望其變化范圍不要太大，太大的權重可能會破壞這些表示。

model.compile(loss='binary_crossentropy',optimizer=optimizers.RMSprop(lr=1e-5),metrics=['acc'])history = model.fit_generator(train_generator,steps_per_epoch=100,epochs=100,validation_data=validation_generator,validation_steps=50) model.save('cats_and_dogs_small_4.h5')

下面，繪制曲線看看效果：

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 acc') plt.plot(epochs, val_acc, 'b', label='Validation acc') plt.title('Training and validation accuracy') plt.legend()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()

這些曲線看起來包含噪音。為了讓圖像更具有可讀性，可以讓每個損失精度替換為指數移動平均，從而讓曲線變得更加平滑，下面用一個簡單實用函數來實現：

def smooth_curve(points, factor=0.8):smoothed_points = []for point in points:if smoothed_points:previous = smoothed_points[-1]smoothed_points.append(previous * factor + point * (1 - factor))else:smoothed_points.append(point)return smoothed_pointsplt.plot(epochs,smooth_curve(acc), 'bo', label='Smoothed training acc') plt.plot(epochs,smooth_curve(val_acc), 'b', label='Smoothed validation acc') plt.title('Training and validation accuracy') plt.legend()plt.figure()plt.plot(epochs,smooth_curve(loss), 'bo', label='Smoothed training loss') plt.plot(epochs,smooth_curve(val_loss), 'b', label='Smoothed validation loss') plt.title('Training and validation loss') plt.legend()plt.show()

通過指數移動平均，驗證曲線變得更清楚了。可以看到，精度提高了1%，約從96%提高到了97%。

下面，在測試集上評估一下這個模型

test_generator = test_datagen.flow_from_directory(test_dir,target_size=(150, 150),batch_size=20,class_mode='binary')test_loss, test_acc = model.evaluate_generator(test_generator, steps=50) print('test acc:', test_acc)

Found 1000 images belonging to 2 classes.
test acc: 0.967999992371

得到了差不多97%的測試精度，在關于這個數據集的原始Kaggle競賽中，這個結果是最佳結果之一。
值得注意的是，我們只是用了一小部分訓練數據（約10%）就得到了這個結果。訓練20000個樣本和訓練2000個樣本還是有很大差別的。

更多精彩內容，歡迎關注我的微信公眾號：數據瞎分析

總結

以上是生活随笔為你收集整理的使用预训练的卷积神经网络（猫狗图片分类）的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。