目录

ModelSVMWrapper包：

应用实现：

构建输入模型（model)

开始训练模型：

ModelSVMWrapper包：

ModelSVMWrapper函数提供一个包，该包内用SVM有效替换Keras模型的softmax。 model_svm_wrapper.py代码实现：

from keras.models import Model
from sklearn.svm import SVC
from keras.utils import to_categorical

class ModelSVMWrapper:
    """
    Linear stack of layers with the option to replace the end of the stack with a Support Vector Machine
    # Arguments
        layers: list of layers to add to the model.
        svm: The Support Vector Machine to use.
    """
    def __init__(self, model, svm=None):
        super().__init__()

        self.model = model
        self.intermediate_model = None  # type: Model
        self.svm = svm

        if svm is None:
            self.svm = SVC(kernel='linear')

    def add(self, layer):
        return self.model.add(layer)

    def fit(self, x=None, y=None, batch_size=None, epochs=1, verbose=1, callbacks=None, validation_split=0.,
            validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0,
            steps_per_epoch=None, validation_steps=None, **kwargs):
        
        fit = self.model.fit(x, to_categorical(y), batch_size, epochs, verbose, callbacks, validation_split,
                             validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch,
                             validation_steps, **kwargs)

        self.fit_svm(x, y, self.__get_split_layer())

        return fit

    def fit_svm(self, x, y, split_layer):
      
        # Store intermediate model
        self.intermediate_model = Model(inputs=self.model.input,
                                        outputs=split_layer.output)
        # Use output of intermediate model to train SVM
        intermediate_output = self.intermediate_model.predict(x)
        self.svm.fit(intermediate_output, y)

    def evaluate(self, x=None, y=None, batch_size=None, verbose=1, steps=None):
       
        if self.intermediate_model is None:
            raise Exception("A model must be fit before running evaluate")
        output = self.predict(x, batch_size, verbose, steps)
        correct = [output[i] == y[i]
                   for i in range(len(output))]

        accuracy = sum(correct) / len(correct)

        return accuracy

    def predict(self, x, batch_size=None, verbose=0, steps=None):
       
        intermediate_prediction = self.intermediate_model.predict(x, batch_size, verbose, steps)
        output = self.svm.predict(intermediate_prediction)

        return output

    def __get_split_layer(self):
     
        if len(self.model.layers) < 3:
            raise ValueError('self.layers to small for a relevant split')

        for layer in self.model.layers:
            if layer.name == "split_layer":
                return layer

        # if no specific cut of point is specified we can assume we need to remove only the last (softmax) layer
        return self.model.layers[-3]

其中：

x：大量需要训练的数据输入。

y：目标（标签）数据的数字数组。

model:是已经定义好的模型；

应用实现：

构建输入模型（model)

# Build a classical model
def build_model():
  model = models.Sequential()
  model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)))
  model.add(layers.MaxPooling2D((2, 2)))
  model.add(layers.Conv2D(64, (3, 3), activation='relu'))
  model.add(layers.MaxPooling2D((2, 2)))
  model.add(layers.Conv2D(64, (3, 3), activation='relu'))
  model.add(layers.Flatten(name="intermediate_output"))
  model.add(layers.Dense(64, activation='relu'))
  model.add(layers.Dense(10, activation='softmax'))
  
  # The extra metric is important for the evaluate function
  model.compile(optimizer='rmsprop',
              loss='categorical_crossentropy',
              metrics=['accuracy'])
  return model

# Wrap it in the ModelSVMWrapper
wrapper = ModelSVMWrapper(build_model())

开始训练模型：

accuracy = {
    "with_svm": [],
    "without_svm": []
}

epochs = 10
for i in range(epochs):
  print('Starting run: {}'.format(i))
  wrapper.fit(train_images, train_labels, epochs=1, batch_size=64)
  accuracy["with_svm"].append(wrapper.evaluate(test_images, test_labels))