python - 如何修复机器学习中恒定的验证准确性？

Question

我正在尝试使用预训练的 InceptionV3 模型对具有平衡类别的 dicom 图像进行图像分类。

def convertDCM(PathDCM) :
   data = []  
   for dirName, subdir, files in os.walk(PathDCM):
          for filename in sorted(files):
                     ds = pydicom.dcmread(PathDCM +'/' + filename)
                     im = fromarray(ds.pixel_array) 
                     im = keras.preprocessing.image.img_to_array(im)
                     im = cv2.resize(im,(299,299))
                     data.append(im) 
   return data

PathDCM = '/home/Desktop/FULL_BALANCED_COLOURED/'

data = convertDCM(PathDCM)

#scale the raw pixel intensities to the range [0,1]
data = np.array(data, dtype="float")/255.0
labels = np.array(labels,dtype ="int")


#splitting data into training and testing
#test_size is percentage to split into test/train data
(trainX, testX, trainY, testY) = train_test_split(
                            data,labels, 
                            test_size=0.2, 
                            random_state=42) 

img_width, img_height = 299, 299 #InceptionV3 size

train_samples =  300
validation_samples = 50
epochs = 25
batch_size = 15

base_model = keras.applications.InceptionV3(
        weights ='imagenet',
        include_top=False, 
        input_shape = (img_width,img_height,3))

model_top = keras.models.Sequential()
 model_top.add(keras.layers.GlobalAveragePooling2D(input_shape=base_model.output_shape[1:], data_format=None)),
model_top.add(keras.layers.Dense(300,activation='relu'))
model_top.add(keras.layers.Dropout(0.5))
model_top.add(keras.layers.Dense(1, activation = 'sigmoid'))
model = keras.models.Model(inputs = base_model.input, outputs = model_top(base_model.output))

#Compiling model 
model.compile(optimizer = keras.optimizers.Adam(
                    lr=0.0001),
                    loss='binary_crossentropy',
                    metrics=['accuracy'])

#Image Processing and Augmentation 
train_datagen = keras.preprocessing.image.ImageDataGenerator(
          rescale = 1./255,  
          zoom_range = 0.1,
          width_shift_range = 0.2, 
          height_shift_range = 0.2,
          horizontal_flip = True,
          fill_mode ='nearest') 

val_datagen = keras.preprocessing.image.ImageDataGenerator()


train_generator = train_datagen.flow(
        trainX, 
        trainY,
        batch_size=batch_size,
        shuffle=True)


validation_generator = train_datagen.flow(
                testX,
                testY,
                batch_size=batch_size,
                shuffle=True)

当我训练模型时，我总是获得 0.3889 的恒定验证精度，而验证损失会波动。

#Training the model
history = model.fit_generator(
    train_generator, 
    steps_per_epoch = train_samples//batch_size,
    epochs = epochs, 
    validation_data = validation_generator, 
    validation_steps = validation_samples//batch_size)

Epoch 1/25
20/20 [==============================]20/20 
[==============================] - 195s 49s/step - loss: 0.7677 - acc: 0.4020 - val_loss: 0.7784 - val_acc: 0.3889

Epoch 2/25
20/20 [==============================]20/20 
[==============================] - 187s 47s/step - loss: 0.7016 - acc: 0.4848 - val_loss: 0.7531 - val_acc: 0.3889

Epoch 3/25
20/20 [==============================]20/20 
[==============================] - 191s 48s/step - loss: 0.6566 - acc: 0.6304 - val_loss: 0.7492 - val_acc: 0.3889

Epoch 4/25
20/20 [==============================]20/20 
[==============================] - 175s 44s/step - loss: 0.6533 - acc: 0.5529 - val_loss: 0.7575 - val_acc: 0.3889


predictions= model.predict(testX)
print(predictions)

预测模型也只返回每个图像一个预测的数组:

[[0.457804  ]
 [0.45051473]
 [0.48343503]
 [0.49180537]...

为什么模型只预测两个类别之一？这是否与恒定的 val 精度或可能的过度拟合有关？

Answer 1

如果你有两个类别，每个图像都属于其中一个类别，因此一个类别的概率足以找到所有内容，因为每个图像的概率之和应该为 1。所以如果你有概率 p 1 类，另一类的概率为 1-p。

如果您希望能够对不属于这两个类别之一的图像进行分类，那么您应该创建第三个类别。

另外，这一行:

model_top.add(keras.layers.Dense(1, activation = 'sigmoid'))

表示输出是形状为(nb_sample,1)的向量，并且与训练标签具有相同的形状