tensorflow - 如何实现卷积神经网络的 L2 正则化成本函数

Question

我已经实现了一个CNN数字分类模型。我的模型过度拟合了很多，为了克服过度拟合，我试图L2 Regularization在我的成本函数中使用。我有一个小小的困惑，我该如何选择<weights>放入成本方程（代码的最后一行）。

...

x = tf.placeholder(tf.float32, shape=[None, img_size, img_size, num_channels], name='x') # Input
y_true = tf.placeholder(tf.float32, shape=[None, num_classes], name='y_true') # Labels

<Convolution Layer 1>

<Convolution Layer 2>

<Convolution Layer 3>

<Fully Coonected 1>

<Fully Coonected 2> O/P = layer_fc2 

# Loss Function
lambda = 0.01
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=layer_fc2, labels=y_true)
# cost        = tf.reduce_mean(cross_entropy) # Nornmal Loss
cost          = tf.reduce_mean(cross_entropy + lambda * tf.nn.l2_loss(<weights>)) # Regularized Loss

...

Answer 1

您应该定义关于权重的 L2 损失 -trainable_variables用于：

C = tf.nn.softmax_cross_entropy_with_logits(logits=layer_fc2, labels=y_true)
l2_loss = tf.add_n([tf.nn.l2_loss(v) for v in tf.trainable_variables()])
C = C + lambda * l2_loss