一.代碼實現
1 # -*- coding: utf-8 -*- 2 """ 3 Created on Sat Feb 9 15:33:39 2019 4 5 @author: zhen 6 """ 7 8 from keras.applications.vgg16 import VGG16 9 10 from keras.layers import Flatten 11 from keras.layers import Dense 12 from keras.layers import Dropout 13 from keras.models import Model 14 from keras.optimizers import SGD 15 16 from keras.datasets import mnist 17 18 import cv2 19 import numpy as np 20 # 因初始設置需大量內存（至少24G），現設置為最小分辨率以降低內存的要求 21 model_vgg = VGG16(include_top=False, weights='imagenet', input_shape=(48, 48, 3)) 22 23 for layer in model_vgg.layers: 24 layer.trainable = False 25 model = Flatten(name='flatten')(model_vgg.output) # 扁平化 26 model = Dense(4096, activation='relu', name='fc1')(model) 27 model = Dense(4096, activation='relu', name='fc2')(model) 28 model = Dropout(0.5)(model) 29 model = Dense(10, activation='softmax')(model) 30 model_vgg_mnist = Model(inputs=model_vgg.input, outputs=model, name='vgg16') 31 32 model_vgg_mnist.summary() 33 34 # VGGNet初始推薦 35 model_vgg = VGG16(include_top=False, weights='imagenet', input_shape=(224, 224, 3)) 36 for layer in model_vgg.layers: 37 layer.trainable = False 38 39 model = Flatten()(model_vgg.output) 40 model = Dense(4096, activation='relu', name='fc1')(model) 41 model = Dense(4096, activation='relu', name='fc2')(model) 42 model = Dropout(0.5)(model) 43 model = Dense(10, activation='softmax', name='prediction')(model) 44 model_vgg_mnist_pretrain = Model(model_vgg.input, model, name='vgg16_pretrain') 45 46 model_vgg_mnist_pretrain.summary() 47 48 sgd = SGD(lr=0.05, decay=1e-5) # 隨機梯度下降 49 model_vgg_mnist.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy']) 50 51 (x_train, y_train), (x_test, y_test) = mnist.load_data("../test_data_home") 52 x_train, y_train = x_train[:1000], y_train[:1000] 53 x_test, y_test = x_test[:1000], y_test[:1000] 54 # GRAY兩通道轉換為RGB三通道 55 x_train = [cv2.cvtColor(cv2.resize(i, (48, 48)), cv2.COLOR_GRAY2RGB) for i in x_train] 56 x_train = np.concatenate([arr[np.newaxis] for arr in x_train]).astype('float32') 57 58 x_test = [cv2.cvtColor(cv2.resize(i, (48, 48)), cv2.COLOR_GRAY2RGB) for i in x_test] 59 x_test = np.concatenate([arr[np.newaxis] for arr in x_test]).astype('float32') 60 61 print(x_train.shape) 62 print(x_test.shape) 63 64 x_train = x_train / 255 65 x_test = x_test / 255 66 67 def tran_y(y): 68 y_ohe = np.zeros(10) 69 y_ohe[y] = 1 70 return y_ohe 71 72 y_train_ohe = np.array([tran_y(y_train[i]) for i in range(len(y_train))]) 73 y_test_ohe = np.array([tran_y(y_test[i]) for i in range(len(y_test))]) 74 75 model_vgg_mnist.fit(x_train, y_train_ohe, validation_data=(x_test, y_test_ohe), epochs=20, batch_size=100)
二.結果
_________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_9 (InputLayer) (None, 48, 48, 3) 0 _________________________________________________________________ block1_conv1 (Conv2D) (None, 48, 48, 64) 1792 _________________________________________________________________ block1_conv2 (Conv2D) (None, 48, 48, 64) 36928 _________________________________________________________________ block1_pool (MaxPooling2D) (None, 24, 24, 64) 0 _________________________________________________________________ block2_conv1 (Conv2D) (None, 24, 24, 128) 73856 _________________________________________________________________ block2_conv2 (Conv2D) (None, 24, 24, 128) 147584 _________________________________________________________________ block2_pool (MaxPooling2D) (None, 12, 12, 128) 0 _________________________________________________________________ block3_conv1 (Conv2D) (None, 12, 12, 256) 295168 _________________________________________________________________ block3_conv2 (Conv2D) (None, 12, 12, 256) 590080 _________________________________________________________________ block3_conv3 (Conv2D) (None, 12, 12, 256) 590080 _________________________________________________________________ block3_pool (MaxPooling2D) (None, 6, 6, 256) 0 _________________________________________________________________ block4_conv1 (Conv2D) (None, 6, 6, 512) 1180160 _________________________________________________________________ block4_conv2 (Conv2D) (None, 6, 6, 512) 2359808 _________________________________________________________________ block4_conv3 (Conv2D) (None, 6, 6, 512) 2359808 _________________________________________________________________ block4_pool (MaxPooling2D) (None, 3, 3, 512) 0 _________________________________________________________________ block5_conv1 (Conv2D) (None, 3, 3, 512) 2359808 _________________________________________________________________ block5_conv2 (Conv2D) (None, 3, 3, 512) 2359808 _________________________________________________________________ block5_conv3 (Conv2D) (None, 3, 3, 512) 2359808 _________________________________________________________________ block5_pool (MaxPooling2D) (None, 1, 1, 512) 0 _________________________________________________________________ flatten (Flatten) (None, 512) 0 _________________________________________________________________ fc1 (Dense) (None, 4096) 2101248 _________________________________________________________________ fc2 (Dense) (None, 4096) 16781312 _________________________________________________________________ dropout_9 (Dropout) (None, 4096) 0 _________________________________________________________________ dense_5 (Dense) (None, 10) 40970 ================================================================= Total params: 33,638,218 Trainable params: 18,923,530 Non-trainable params: 14,714,688 _________________________________________________________________ _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_10 (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_5 (Flatten) (None, 25088) 0 _________________________________________________________________ fc1 (Dense) (None, 4096) 102764544 _________________________________________________________________ fc2 (Dense) (None, 4096) 16781312 _________________________________________________________________ dropout_10 (Dropout) (None, 4096) 0 _________________________________________________________________ prediction (Dense) (None, 10) 40970 ================================================================= Total params: 134,301,514 Trainable params: 119,586,826 Non-trainable params: 14,714,688 _________________________________________________________________ (1000, 48, 48, 3) (1000, 48, 48, 3) Train on 1000 samples, validate on 1000 samples Epoch 1/20 1000/1000 [==============================] - 175s 175ms/step - loss: 2.1289 - acc: 0.2350 - val_loss: 1.9100 - val_acc: 0.4230 Epoch 2/20 1000/1000 [==============================] - 190s 190ms/step - loss: 1.7685 - acc: 0.4420 - val_loss: 1.6503 - val_acc: 0.4930 Epoch 3/20 1000/1000 [==============================] - 265s 265ms/step - loss: 1.5582 - acc: 0.5140 - val_loss: 1.5005 - val_acc: 0.5440 Epoch 4/20 1000/1000 [==============================] - 373s 373ms/step - loss: 1.4210 - acc: 0.5710 - val_loss: 1.3019 - val_acc: 0.6160 Epoch 5/20 1000/1000 [==============================] - 295s 295ms/step - loss: 1.1946 - acc: 0.6490 - val_loss: 1.1182 - val_acc: 0.7280 Epoch 6/20 1000/1000 [==============================] - 277s 277ms/step - loss: 1.0291 - acc: 0.7330 - val_loss: 1.0279 - val_acc: 0.7430 Epoch 7/20 1000/1000 [==============================] - 177s 177ms/step - loss: 1.0065 - acc: 0.7060 - val_loss: 0.9229 - val_acc: 0.7690 Epoch 8/20 1000/1000 [==============================] - 169s 169ms/step - loss: 0.8438 - acc: 0.7810 - val_loss: 0.9716 - val_acc: 0.6670 Epoch 9/20 1000/1000 [==============================] - 169s 169ms/step - loss: 0.8898 - acc: 0.7230 - val_loss: 0.9710 - val_acc: 0.6660 Epoch 10/20 1000/1000 [==============================] - 166s 166ms/step - loss: 0.8258 - acc: 0.7460 - val_loss: 0.9026 - val_acc: 0.7130 Epoch 11/20 1000/1000 [==============================] - 169s 169ms/step - loss: 0.7592 - acc: 0.7640 - val_loss: 0.9691 - val_acc: 0.6730 Epoch 12/20 1000/1000 [==============================] - 165s 165ms/step - loss: 0.7793 - acc: 0.7520 - val_loss: 0.8350 - val_acc: 0.6800 Epoch 13/20 1000/1000 [==============================] - 164s 164ms/step - loss: 0.6677 - acc: 0.7780 - val_loss: 0.7203 - val_acc: 0.7730 Epoch 14/20 1000/1000 [==============================] - 164s 164ms/step - loss: 0.7018 - acc: 0.7630 - val_loss: 0.6947 - val_acc: 0.7760 Epoch 15/20 1000/1000 [==============================] - 163s 163ms/step - loss: 0.6129 - acc: 0.8100 - val_loss: 0.7025 - val_acc: 0.7610 Epoch 16/20 1000/1000 [==============================] - 163s 163ms/step - loss: 0.6104 - acc: 0.8190 - val_loss: 0.6385 - val_acc: 0.8220 Epoch 17/20 1000/1000 [==============================] - 163s 163ms/step - loss: 0.5507 - acc: 0.8320 - val_loss: 0.6273 - val_acc: 0.8290 Epoch 18/20 1000/1000 [==============================] - 164s 164ms/step - loss: 0.5205 - acc: 0.8360 - val_loss: 0.8740 - val_acc: 0.6750 Epoch 19/20 1000/1000 [==============================] - 163s 163ms/step - loss: 0.5852 - acc: 0.8150 - val_loss: 0.6614 - val_acc: 0.7890 Epoch 20/20 1000/1000 [==============================] - 166s 166ms/step - loss: 0.5310 - acc: 0.8340 - val_loss: 0.5718 - val_acc: 0.8250
三.解析
VGGNet是牛津大學計算機視覺組（Visual Geometry Group）和Google DeepMind公司的研究員一起研發的深度卷積神經網絡。VGG探索了卷積神經網絡的深度與其性能之間的關系，通過反復堆疊3*3的小型卷積核和2*2的最大池化層，VGG成功構築了16-19層深的卷積神經網絡。
VGG取得了2014年比賽分類項目第二名和定位項目第一名。同時，VGG拓展性很強，遷移到其他圖片數據上的泛化性非常好。VGG的結構簡潔，整個網絡都是使用了同樣大小的卷積核尺寸3*3和池化層2*2。VGG現在也還經常被用來提取圖像特征，可用來在圖像分類任務上進行再訓練，相當於提供了非常好的初始化權重。
VGG通過加深層次來提升性能，擁有5段卷積，每一段內有2-3個卷積層，同時每段尾部都會連接一個最大池化層來縮小圖片尺寸。每段內的卷積核數量一樣，越靠后段的卷積核數量越多，64-128-256-512-512。