Question

我在colab中拥有模型resnet 34，可从8000个样本中进行训练并从2000个样本中进行验证（7个类）它具有输出层：

x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dropout(0.7)(x)
predictions = Dense(num_classes, activation= 'softmax')(x)
model = Model(inputs = base_model.input, outputs = predictions)

数据拆分：

from google.colab import drive
drive.mount('/content/drive')
img_width, img_height = 160, 160
train_data_dir =( "/content/drive/My Drive/Skin_cancer_all/skin_cancer")


train_datagen = ImageDataGenerator(rescale=1./255,
    validation_split=0.2) # set validation split

train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=(img_width, img_height),
    batch_size=8,  
    subset='training') # set as training data

validation_generator = train_datagen.flow_from_directory(
    train_data_dir, # same directory as training data
    target_size=(img_width, img_height),
    batch_size=8,
    subset='validation') # set as validation data

我用过fit_Generator：

nb_epochs = 4
nb_train_samples = 8015
nb_validation_samples = 2000

model.compile(loss='categorical_crossentropy',
     optimizer='rmsprop',      
              metrics=['accuracy'])
    # but it used to start model.


history=model.fit_generator(train_generator,
                    steps_per_epoch=nb_train_samples,
                    epochs=nb_epochs
                        , validation_data=validation_generator
                        , validation_steps=nb_validation_samples)

我从4个时代开始

Epoch 1/4
8014/8015 [============================>.] - ETA: 0s - loss: 1.2813 - acc: 0.6681Epoch 1/4
8015/8015 [==============================] - 4406s 550ms/step - loss: 1.2812 - acc: 0.6681 - val_loss: 2.0039 - val_acc: 0.6750
Epoch 2/4
8014/8015 [============================>.] - ETA: 0s - loss: 1.1651 - acc: 0.6749Epoch 1/4
8015/8015 [==============================] - 1872s 234ms/step - loss: 1.1651 - acc: 0.6749 - val_loss: 10.8641 - val_acc: 0.6720
Epoch 3/4
8014/8015 [============================>.] - ETA: 0s - loss: 1.0921 - acc: 0.6818Epoch 1/4
8015/8015 [==============================] - 1881s 235ms/step - loss: 1.0921 - acc: 0.6818 - val_loss: 87.3402 - val_acc: 0.6920
Epoch 4/4
8014/8015 [============================>.] - ETA: 0s - loss: 1.0539 - acc: 0.6947Epoch 1/4
8015/8015 [==============================] - 1898s 237ms/step - loss: 1.0538 - acc: 0.6947 - val_loss: 126.0995 - val_acc: 0.7140

看起来不是很好，但是我的val_loss太高了，我认为这不仅仅是过度拟合，如果是我的数据集被分为2000 val和8000 train的话，那将是另一个问题，因为我之前从未见过此val_loss，而我之前和另一个模型（resnet-50）仍在训练相同的数据集，并且做了val_loss值不高（某些时期后，我看到的resnet 50的最大值是5）

我之前看到过合适的效果val_acc，但它仍然得到了改善，火车损耗也继续减少

我只想对此行为做出解释！如果一切都很好，但val_loss和val_acc是固定的抛出时期怎么办？

为什么val_loss急剧增加？但是train和val的精度足够好，并且train_loss也降低了吗？

0 个答案: