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)
.
参数:
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 tf sess=tf.Session() weight_decay=0.1 tmp=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 tf sess=tf.Session() weight_decay=0.1 #(1)定义weight_decay l2_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 tf import numpy as np def 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 var x = tf.placeholder(tf.float32, shape=(None, 2)) y_ = tf.placeholder(tf.float32, shape=(None, 1)) batch_size = 8 layer_dimension = [2, 10, 10, 10, 1] n_layers = len(layer_dimension) cur_lay = x in_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。