我有10000张图像，其中5000张是病变的医学图像，另外5000张是健康图像。我使用了vgg16并修改了最后几层，如下所示：
Layer (type) Output Shape Param # =================================================================input_1 (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 (Flatten) (None, 25088) 0 _________________________________________________________________fc1 (Dense) (None, 256) 6422784 _________________________________________________________________fc2 (Dense) (None, 128) 32896 _________________________________________________________________output (Dense) (None, 2) 258 =================================================================Total params: 21,170,626Trainable params: 6,455,938Non-trainable params: 14,714,688我的代码如下：
import numpy as npimport osimport timefrom vgg16 import VGG16from keras.preprocessing import imagefrom imagenet_utils import preprocess_input, decode_predictionsfrom keras.layers import Dense, Activation, Flattenfrom keras.layers import merge, Inputfrom keras.models import Modelfrom keras.utils import np_utilsfrom sklearn.utils import shufflefrom sklearn.cross_validation import train_test_split# 加载训练数据PATH = '/mount'# 定义数据路径data_path = PATH data_dir_list = os.listdir(data_path)img_data_list=[]y=0;for dataset in data_dir_list: img_list=os.listdir(data_path+'/'+ dataset) print ('已加载数据集的图像-'+'{}\n'.format(dataset)) for img in img_list: img_path = data_path + '/'+ dataset + '/'+ img img = image.load_img(img_path, target_size=(224, 224)) x = image.img_to_array(img) x = np.expand_dims(x, axis=0) x = preprocess_input(x) x = x/255 y=y+1 print('输入图像形状:', x.shape) print(y) img_data_list.append(x)from keras.optimizers import SGDsgd = SGD(lr=1e-3, decay=1e-6, momentum=0.9, nesterov=True)img_data = np.array(img_data_list)#img_data = img_data.astype('float32')print (img_data.shape)img_data=np.rollaxis(img_data,1,0)print (img_data.shape)img_data=img_data[0]print (img_data.shape)# 定义类别数量num_classes = 2num_of_samples = img_data.shape[0]labels = np.ones((num_of_samples,),dtype='int64')labels[0:5001]=0labels[5001:]=1names = ['YES','NO']# 将类别标签转换为独热编码Y = np_utils.to_categorical(labels, num_classes)# 打乱数据集x,y = shuffle(img_data,Y, random_state=2)# 拆分数据集X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=2)image_input = Input(shape=(224, 224, 3))model = VGG16(input_tensor=image_input, include_top=True,weights='imagenet')model.summary()last_layer = model.get_layer('block5_pool').outputx= Flatten(name='flatten')(last_layer)x = Dense(256, activation='relu', name='fc1')(x)x = Dense(128, activation='relu', name='fc2')(x)out = Dense(num_classes, activation='softmax', name='output')(x)custom_vgg_model2 = Model(image_input, out)custom_vgg_model2.summary()# 冻结除了全连接层之外的所有层for layer in custom_vgg_model2.layers[:-3]: layer.trainable = Falsecustom_vgg_model2.summary()custom_vgg_model2.compile(loss='categorical_crossentropy',optimizer=sgd,metrics=['accuracy'])t=time.time()# t = now()hist = custom_vgg_model2.fit(X_train, y_train, batch_size=128, epochs=50, verbose=1, validation_data=(X_test, y_test))print('训练时间: %s' % (t - time.time()))(loss, accuracy) = custom_vgg_model2.evaluate(X_test, y_test, batch_size=10, verbose=1)print("[INFO] 损失={:.4f}, 准确率: {:.4f}%"。format(loss,accuracy * 100))model.save("vgg_10000.h5")我发布的结果是前5个和最后5个纪元
Epoch 1/508000/8000 [==============================] - 154s - loss: 0.6960 - acc: 0.5354 - val_loss: 0.6777 - val_acc: 0.5745Epoch 2/508000/8000 [==============================] - 134s - loss: 0.6684 - acc: 0.5899 - val_loss: 0.6866 - val_acc: 0.5490Epoch 3/508000/8000 [==============================] - 134s - loss: 0.6608 - acc: 0.6040 - val_loss: 0.6625 - val_acc: 0.5925Epoch 4/508000/8000 [==============================] - 134s - loss: 0.6518 - acc: 0.6115 - val_loss: 0.6668 - val_acc: 0.5810Epoch 5/508000/8000 [==============================] - 134s - loss: 0.6440 - acc: 0.6280 - val_loss: 0.6990 - val_acc: 0.5580最后5个
Epoch 25/508000/8000 [==============================] - 134s - loss: 0.5944 - acc: 0.6720 - val_loss: 0.6271 - val_acc: 0.6485Epoch 26/508000/8000 [==============================] - 134s - loss: 0.5989 - acc: 0.6699 - val_loss: 0.6483 - val_acc: 0.6135Epoch 27/508000/8000 [==============================] - 134s - loss: 0.5950 - acc: 0.6789 - val_loss: 0.7130 - val_acc: 0.5785Epoch 28/508000/8000 [==============================] - 134s - loss: 0.5853 - acc: 0.6838 - val_loss: 0.6263 - val_acc: 0.6395结果并不是很好。我尝试调整了最后两层的节点数为128和128，使用了adam优化器等，但结果仍然不令人满意。任何帮助都将非常感激。
回答：
你可以尝试以下方法：
-
进行分层train_test_split
train_test_split(x, y, stratify=y, test_size=0.2, random_state=2) -
检查你的数据，看看图像中是否有异常值。
- 使用adam优化器：
from keras.optimizers import Adam替代SGD -
尝试在适用的地方使用不同的种子，即替代
random_state=2，使用其他值：X_train, X_test, y_train, y_test = train_test_split( x, y, test_size=0.2, random_state=382938) -
尝试使用
include_top=False：model = VGG16(input_tensor=image_input, include_top=False,weights='imagenet') -
使用
(训练, 验证, 测试)集或(交叉验证, 保留集)来获得更可靠的性能指标。
