一、基础正则化函数
tf.contrib.layers.l1_regularizer(scale, scope=None)
返回一个用来执行L1正则化的函数,函数的签名是func(weights).
- scale: 正则项的系数.
- scope: 可选的
scope name
tf.contrib.layers.l2_regularizer(scale, scope=None)
先看看tf.contrib.layers.l2_regularizer(weight_decay)都执行了什么：
import tensorflow as tfsess=tf.Session()weight_decay=0.1tmp=tf.constant([0,1,2,3],dtype=tf.float32)"""l2_reg=tf.contrib.layers.l2_regularizer(weight_decay)a=tf.get_variable("I_am_a",regularizer=l2_reg,initializer=tmp) """#**上面代码的等价代码a=tf.get_variable("I_am_a",initializer=tmp)a2=tf.reduce_sum(a*a)*weight_decay/2;a3=tf.get_variable(a.name.split(":")[0]+"/Regularizer/l2_regularizer",initializer=a2)tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES,a2)#**sess.run(tf.global_variables_initializer())keys = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)for key in keys: print("%s : %s" %(key.name,sess.run(key))) 我们很容易可以模拟出tf.contrib.layers.l2_regularizer都做了什么，不过会让代码变丑。
以下比较完整实现L2 正则化。
import tensorflow as tfsess=tf.Session()weight_decay=0.1 #(1)定义weight_decayl2_reg=tf.contrib.layers.l2_regularizer(weight_decay) #(2)定义l2_regularizer()tmp=tf.constant([0,1,2,3],dtype=tf.float32)a=tf.get_variable("I_am_a",regularizer=l2_reg,initializer=tmp) #(3)创建variable，l2_regularizer复制给regularizer参数。 #目测REXXX_LOSSES集合#regularizer定义会将a加入REGULARIZATION_LOSSES集合print("Global Set:")keys = tf.get_collection("variables")for key in keys: print(key.name)print("Regular Set:")keys = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)for key in keys: print(key.name)print("--------------------")sess.run(tf.global_variables_initializer())print(sess.run(a))reg_set=tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES) #(4)则REGULARIAZTION_LOSSES集合会包含所有被weight_decay后的参数和，将其相加l2_loss=tf.add_n(reg_set)print("loss=%s" %(sess.run(l2_loss)))"""此处输出0.7,即: weight_decay*sigmal(w*2)/2=0.1*(0*0+1*1+2*2+3*3)/2=0.7其实代码自己写也很方便，用API看着比较正规。在网络模型中，直接将l2_loss加入loss就好了。(loss变大，执行train自然会decay)""" 二、添加正则化方法
a、原始办法
正则化常用到集合，下面是最原始的添加正则办法（直接在变量声明后将之添加进'losses'集合或tf.GraphKeys.LOESSES也行）：
import tensorflow as tfimport numpy as npdef get_weights(shape, lambd): var = tf.Variable(tf.random_normal(shape), dtype=tf.float32) tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(lambd)(var)) return varx = tf.placeholder(tf.float32, shape=(None, 2))y_ = tf.placeholder(tf.float32, shape=(None, 1))batch_size = 8layer_dimension = [2, 10, 10, 10, 1]n_layers = len(layer_dimension)cur_lay = xin_dimension = layer_dimension[0]for i in range(1, n_layers): out_dimension = layer_dimension[i] weights = get_weights([in_dimension, out_dimension], 0.001) bias = tf.Variable(tf.constant(0.1, shape=[out_dimension])) cur_lay = tf.nn.relu(tf.matmul(cur_lay, weights)+bias) in_dimension = layer_dimension[i]mess_loss = tf.reduce_mean(tf.square(y_-cur_lay))tf.add_to_collection('losses', mess_loss)loss = tf.add_n(tf.get_collection('losses')) b、tf.contrib.layers.apply_regularization(regularizer, weights_list=None)
先看参数
- regularizer:就是我们上一步创建的正则化方法
- weights_list: 想要执行正则化方法的参数列表,如果为
None的话,就取GraphKeys.WEIGHTS中的weights.
函数返回一个标量Tensor,同时,这个标量Tensor也会保存到GraphKeys.REGULARIZATION_LOSSES中.这个Tensor保存了计算正则项损失的方法.
tensorflow中的Tensor是保存了计算这个值的路径(方法),当我们run的时候,tensorflow后端就通过路径计算出Tensor对应的值
现在,我们只需将这个正则项损失加到我们的损失函数上就可以了.
如果是自己手动定义
weight的话,需要手动将weight保存到GraphKeys.WEIGHTS中,但是如果使用layer的话,就不用这么麻烦了,别人已经帮你考虑好了.(最好自己验证一下tf.GraphKeys.WEIGHTS中是否包含了所有的weights,防止被坑)
c、使用slim
使用slim会简单很多：
with slim.arg_scope([slim.conv2d, slim.fully_connected], activation_fn=tf.nn.relu, weights_regularizer=slim.l2_regularizer(weight_decay)）： pass
此时添加集合为tf.GraphKeys.REGULARIZATION_LOSSES。