Keras貓狗大戰七：resnet50預訓練模型遷移學習優化，動態調整學習率，精度提高到96.2%

在https://www.cnblogs.com/zhengbiqing/p/11780161.html中直接在resnet網絡的卷積層后添加一層分類層，得到一個最簡單的遷移學習模型，得到的結果為95.3%。

這里對最后的分類網絡做些優化：用GlobalAveragePooling2D替換Flatten、增加一個密集連接層（同時添加BN、Activation、Dropout）：

conv_base = ResNet50(weights='imagenet', include_top=False, input_shape=(150, 150, 3))
for layers in conv_base.layers[:]:
    layers.trainable = False
    
x = conv_base.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dropout(0.3)(x)
predictions = Dense(1, activation='sigmoid')(x)
model = Model(inputs=conv_base.input, outputs=predictions)

另外采用動態學習率，並且打印顯示出學習率：

optimizer = optimizers.RMSprop(lr=1e-3)

def get_lr_metric(optimizer):
    def lr(y_true, y_pred):
        return optimizer.lr

    return lr

lr_metric = get_lr_metric(optimizer)

model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['acc',lr_metric])

當模型的val_loss訓練多輪不再下降時，提前結束訓練：

from keras.callbacks import ReduceLROnPlateau,EarlyStopping

early_stop = EarlyStopping(monitor='val_loss', patience=13)
reduce_lr = ReduceLROnPlateau(monitor='val_loss', patience=7, mode='auto', factor=0.2)
callbacks = [early_stop,reduce_lr]

history = model.fit_generator(
      train_generator,
      steps_per_epoch=train_generator.samples//batch_size,
      epochs=100,
      validation_data=validation_generator,
      validation_steps=validation_generator.samples//batch_size,
      callbacks = callbacks)

共訓練了61epochs，學習率從0.001下降到1.6e-6：

Epoch 1/100
281/281 [==============================] - 141s 503ms/step - loss: 0.3322 - acc: 0.8589 - lr: 0.0010 - val_loss: 0.2344 - val_acc: 0.9277 - val_lr: 0.0010
Epoch 2/100
281/281 [==============================] - 79s 279ms/step - loss: 0.2591 - acc: 0.8862 - lr: 0.0010 - val_loss: 0.2331 - val_acc: 0.9288 - val_lr: 0.0010
Epoch 3/100
281/281 [==============================] - 78s 279ms/step - loss: 0.2405 - acc: 0.8959 - lr: 0.0010 - val_loss: 0.2292 - val_acc: 0.9303 - val_lr: 0.0010
......

281/281 [==============================] - 77s 275ms/step - loss: 0.1532 - acc: 0.9407 - lr: 1.6000e-06 - val_loss: 0.1871 - val_acc: 0.9412 - val_lr: 1.6000e-06
Epoch 60/100
281/281 [==============================] - 78s 276ms/step - loss: 0.1492 - acc: 0.9396 - lr: 1.6000e-06 - val_loss: 0.1687 - val_acc: 0.9450 - val_lr: 1.6000e-06
Epoch 61/100
281/281 [==============================] - 77s 276ms/step - loss: 0.1468 - acc: 0.9414 - lr: 1.6000e-06 - val_loss: 0.1825 - val_acc: 0.9454 - val_lr: 1.6000e-06

加載模型：

optimizer = optimizers.RMSprop(lr=1e-3)

def get_lr_metric(optimizer):
    def lr(y_true, y_pred):
        return optimizer.lr

    return lr

lr_metric = get_lr_metric(optimizer)
model = load_model(model_file, custom_objects={'lr':lr_metric})

修改混淆矩陣函數，以打印每個類別的精確度：

def plot_sonfusion_matrix(cm, classes, normalize=False, title='Confusion matrix', cmap=plt.cm.Blues):
    plt.imshow(cm, interpolation='nearest', cmap=cmap)
    plt.title(title)
    plt.colorbar()
    tick_marks = np.arange(len(classes))
    plt.xticks(tick_marks, classes, rotation=45)
    plt.yticks([-0.5,1.5], classes)

    print(cm)
    ok_num = 0
    for k in range(cm.shape[0]):
        print(cm[k,k]/np.sum(cm[k,:]))
        ok_num += cm[k,k]
        
    print(ok_num/np.sum(cm))
        
    if normalize:
        cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]

    thresh = cm.max() / 2.0
    for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):
        plt.text(j, i, cm[i, j], horizontalalignment='center', color='white' if cm[i, j] > thresh else 'black')

    plt.tight_layout()
    plt.ylabel('True label')
    plt.xlabel('Predict label')

測試結果為：

[[1200   50]
 [  45 1205]]
0.96
0.964
0.962
貓的准確度為96%，狗的為96.4%，總的准確度為96.2%。混淆矩陣圖：