深度学习 – 在keras和千层面的公路网 – 显着的性能差异

我用keras和烤宽面条实现了高速公路网络,而keras版本一直低于烤宽面条版本.我在它们中使用相同的数据集和元参数.这是keras版本的代码：

X_train, y_train, X_test, y_test, X_all = hacking_script.load_all_data()
data_dim = 144
layer_count = 32
dropout = 0.04
hidden_units = 32
nb_epoch = 10

model = Sequential()
model.add(Dense(hidden_units, input_dim=data_dim))
model.add(Dropout(dropout))
for index in range(layer_count):
    model.add(Highway(activation = 'relu'))
    model.add(Dropout(dropout))
model.add(Dropout(dropout))
model.add(Dense(2, activation='softmax'))


print 'compiling...'
model.compile(loss='binary_crossentropy', optimizer='adagrad')
model.fit(X_train, y_train, batch_size=100, nb_epoch=nb_epoch,
    show_accuracy=True, validation_data=(X_test, y_test), shuffle=True, verbose=0)

predictions = model.predict_proba(X_test)

这是烤宽面条版的代码：

class MultiplicativeGatingLayer(MergeLayer):
    def __init__(self, gate, input1, input2, **kwargs):
        incomings = [gate, input1, input2]
        super(MultiplicativeGatingLayer, self).__init__(incomings, **kwargs)
        assert gate.output_shape == input1.output_shape == input2.output_shape

    def get_output_shape_for(self, input_shapes):
        return input_shapes[0]

    def get_output_for(self, inputs, **kwargs):
        return inputs[0] * inputs[1] + (1 - inputs[0]) * inputs[2]


def highway_dense(incoming, Wh=Orthogonal(), bh=Constant(0.0),
                  Wt=Orthogonal(), bt=Constant(-4.0),
                  nonlinearity=rectify, **kwargs):
    num_inputs = int(np.prod(incoming.output_shape[1:]))

    l_h = DenseLayer(incoming, num_units=num_inputs, W=Wh, b=bh, nonlinearity=nonlinearity)
    l_t = DenseLayer(incoming, num_units=num_inputs, W=Wt, b=bt, nonlinearity=sigmoid)

    return MultiplicativeGatingLayer(gate=l_t, input1=l_h, input2=incoming)

# ==== Parameters ====

num_features = X_train.shape[1]
epochs = 10

hidden_layers = 32
hidden_units = 32
dropout_p = 0.04

# ==== Defining the neural network shape ====

l_in = InputLayer(shape=(None, num_features))
l_hidden1 = DenseLayer(l_in, num_units=hidden_units)
l_hidden2 = DropoutLayer(l_hidden1, p=dropout_p)
l_current = l_hidden2
for k in range(hidden_layers - 1):
    l_current = highway_dense(l_current)
    l_current = DropoutLayer(l_current, p=dropout_p)
l_dropout = DropoutLayer(l_current, p=dropout_p)
l_out = DenseLayer(l_dropout, num_units=2, nonlinearity=softmax)

# ==== Neural network definition ====

net1 = NeuralNet(layers=l_out,
                 update=adadelta, update_rho=0.95, update_learning_rate=1.0,
                 objective_loss_function=categorical_crossentropy,
                 train_split=TrainSplit(eval_size=0), verbose=0, max_epochs=1)

net1.fit(X_train, y_train)
predictions = net1.predict_proba(X_test)[:, 1]

现在keras版本几乎没有比logistic回归更好,而烤宽面条版本是目前为止最好的评分算法.任何想法为什么？

以下是一些建议(我不确定它们是否会真正缩小您所观察到的性能差距)：

根据Keras documentation,使用Glorot Uniform权重初始化Highway层,而在您的Lasagne代码中,您使用正交权重初始化.除非你的代码的另一部分用于为Keras Highway层设置权重初始化为Orthogonal,否则这可能是性能差距的来源.

您似乎也在使用Adagrad作为Keras型号,但您正在使用Adadelta作为您的烤宽面条型号.

此外,我不是100％确定这一点,但您可能还想验证您的变换偏差项是否以相同的方式初始化.