Apply L2 regularization to tensorflow variables

作者: ganhai 分类: 技术, 笔记发布时间: 2017-11-30 20:56 ė 1255 6 没有评论

Basic steps to apply L2 regularization variables

Firstly call tf.contrib.layers.l2_regularizer() to return a L2 regularization function, with regularization parameter lamda.
Next define a tensorflow variable, with keyword parameter “regularizer”, this variable will be added to tf.GraphKeys.REGULARIZATION_LOSSES collections.
Get the regularized variables from the collections, apply regularization and added to the loss.

regularizer = tf.contrib.layers.l2_regularizer(lamda)

W1 = tf.get_variable('W1', [3, 3, 3, 32], \

initializer = tf.contrib.layers.xavier_initializer(), \

regularizer = regularizer)

reg_variables = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)

reg_term = tf.contrib.layers.apply_regularization(regularizer, reg_variables)

loss += reg_term

Per Tensorflow 1.4 release notes, currently CUDA 8 and cuDNN 6 are supported.
CUDA 9 and cuDNN 7 will be supported in next release.
Download CUDA 8
Download cuDNN 6

本文出自扉启博客，转载时请注明出处及相应链接。

本文永久链接: https://www.feiqy.com/apply-l2-regularization-to-tensorflow-variables/