如何向我的 python keras ANN 添加噪声（抖动）以避免过度拟合？

Question

我在 Python Keras 中实现了一个人工神经网络模型，我的训练准确率很高，但测试准确率很低。这意味着数据中存在一些过度拟合。

我想避免过度拟合，其中一种技术是抖动或噪声添加。但是，我的问题是：如何在 Python 中完成？

这是我的 ANN 代码：

def designANN(input_nodes, dropout, layer_nodes, output_nodes):

    classifier = Sequential()

    classifier.add(Dense(units = layer_nodes, kernel_initializer = "uniform",
                 activation = "relu", input_dim = input_nodes)) 

    classifier.add(Dropout(dropout))

    classifier.add(Dense(units = layer_nodes, kernel_initializer = "uniform",
                 activation = "relu"))
    classifier.add(Dropout(dropout))


    classifier.add(Dense(units = output_nodes, kernel_initializer = "uniform",
                 activation = "sigmoid"))


    classifier.compile(optimizer = "adam", loss = "binary_crossentropy", metrics = [npv])

    return classifier

Answer 1

您只需要 GaussianNoise 层。你可以把它放在你的网络中。我建议在激活函数之前使用它。这是 relu 的情况，如果我们添加随机噪声

，输出值可能超出范围 (<0)

def designANN(input_nodes, dropout, layer_nodes, output_nodes):

    classifier = Sequential()

    classifier.add(Dense(units = layer_nodes, kernel_initializer = "uniform",
                         input_dim = input_nodes))
    classifier.add(GaussianNoise(0.1))
    classifier.add(Activation('relu'))
    classifier.add(Dropout(dropout))

    classifier.add(Dense(units = layer_nodes, kernel_initializer = "uniform"))
    classifier.add(GaussianNoise(0.1))
    classifier.add(Activation('relu'))
    classifier.add(Dropout(dropout))

    classifier.add(Dense(units = output_nodes, kernel_initializer = "uniform",
                 activation = "sigmoid"))

    classifier.compile(optimizer = "adam", loss = "binary_crossentropy", metrics = [npv])

    return classifier

如何向我的 python keras ANN 添加噪声（抖动）以避免过度拟合？

How can I add noise (jittering) to my python keras ANN to avoid overfitting?

python

keras

neural-network

noise

jitter