损失函数 对比 曲度 收敛 *滑

 

【1】

L2正则损失函数、欧拉损失函数：预测值与目标差值的*方和

 　　在目标值附*有更好的曲度，离目标越*收敛越慢

 

 

 

 

【2】

 L1正则损失函数，绝对值损失函数 

　　在目标值附*不*滑，导致算法不能很好地收敛

 

 

【1】【2】代码

 

import matplotlib.pyplot as plt
import tensorflow as tf

sess = tf.Session()
x_vals = tf.linspace(-1., 1., 500)
target = tf.constant(0.)
l2_y_vals = tf.square(target - x_vals)
l2_y_out = sess.run(l2_y_vals)

　　

 

 【3】

Pseudo-Huber损失函数是 Huber损失函数的连续、*滑估计，师徒利用L1和L2正则雪见极值处的陡峭，使得目标值附*连续。

它的表达式依赖delta。

 

 

 

 【4】分类损失函数

用来评估预测分类结果

 

 

 【5】

Hinge 损失函数 主要用来评估支持向量机算法，也用来评估神经网络算法

 

本例中，计算两个目标类(-1,1) 之间的损失，使用目标值1，故预测距离值1越*，

 

 

【6】两类交叉熵损失函数  逻辑损失函数

Cross-entropy loss

 当预测2类目标0或者1时，希望度量预测值到真实分类值(0或者1)的距离，这个距离经常是0到1之间的实数。为了度量这个距离，我们可以使用信息论汇总的交叉熵。

 

 【7】Sigmoid损失函数 Sigmoid cross entropy loss 与两类交叉熵损失函数类是，不同的是，它先把x_vals的值通过sigmoid函数转换，再计算交叉熵损失

 

 

import matplotlib.pyplot as plt
import tensorflow as tf

sess = tf.Session()
x_vals = tf.linspace(-1., 1., 500)
target = tf.constant(0.)

l2_y_vals = tf.square(target - x_vals)
l2_y_out = sess.run(l2_y_vals)

l1_y_vals = tf.abs(target - x_vals)
l1_y_out = sess.run(l1_y_vals)

delta1 = tf.constant(0.25)
phuber1_y_als = tf.multiply(tf.square(delta1), tf.sqrt(1. + tf.square((target - x_vals) / delta1)) - 1.)
phuber1_y_out = sess.run(phuber1_y_als)

delta2 = tf.constant(5.)
phuber2_y_als = tf.multiply(tf.square(delta2), tf.sqrt(1. + tf.square((target - x_vals) / delta2)) - 1.)
phuber2_y_out = sess.run(phuber2_y_als)

x_vals = tf.linspace(-3., 5., 500)
target = tf.constant(1.)
targets = tf.fill([500, ], 1.)

hinge_y_vals = tf.maximum(0., 1. - tf.multiply(target, x_vals))
hinge_y_out = sess.run(hinge_y_vals)

# [i for i  in xentropy_y_out if not sess.run(tf.is_nan(i))]
xentropy_y_vals = -tf.multiply(target, tf.log(x_vals)) - tf.multiply((1. - target), tf.log(1. - x_vals))
xentropy_y_out = sess.run(xentropy_y_vals)
not_nan = [i for i in xentropy_y_out if not sess.run(tf.is_nan(i))]

# logits and targets must have the same type and shape.
# ValueError: Only call `sigmoid_cross_entropy_with_logits` with named arguments (labels=..., logits=..., ...)
# xentropy_sigmoid_y_vals = tf.nn.sigmoid_cross_entropy_with_logits(x_vals, logits=targets)
# sess.run(tf.nn.sigmoid_cross_entropy_with_logits(labels=x_vals, logits=targets))
xentropy_sigmoid_y_vals = tf.nn.sigmoid_cross_entropy_with_logits(labels=x_vals, logits=targets)
xentropy_sigmoid_y_out = sess.run(xentropy_sigmoid_y_vals)

dd = 9

 

 

【8】加权交叉熵损失函数 Weighted cross entropy loss 是Sigmoid交叉熵损失函数的加权，对正目标加权

本例中，对正目标加权权重0.5

 

 

【9】Softmax交叉熵损失函数 Softmax cross-entropy loss 作用于非归一化的输出结果，只针对单个目标分类的计算损失。通过softmax函数将输出结果

转化成概率分布，然后计算真值概率分布的损失。

 

 

【TOCHECK】 1.16012561

 

【10】稀疏Softmax交叉熵损失函数 Sparse softmax cross-entropy loss 和Softmax交叉熵损失函数类似，它是把目标分类为true的转化成index，

而Softmax交叉熵损失函数将目标转成概率分布。

 

 【TOCHECK】0.00012564

 

import matplotlib.pyplot as plt
import tensorflow as tf

sess = tf.Session()
x_vals = tf.linspace(-1., 1., 500)
target = tf.constant(0.)

l2_y_vals = tf.square(target - x_vals)
l2_y_out = sess.run(l2_y_vals)

l1_y_vals = tf.abs(target - x_vals)
l1_y_out = sess.run(l1_y_vals)

delta1 = tf.constant(0.25)
phuber1_y_als = tf.multiply(tf.square(delta1), tf.sqrt(1. + tf.square((target - x_vals) / delta1)) - 1.)
phuber1_y_out = sess.run(phuber1_y_als)

delta2 = tf.constant(5.)
phuber2_y_als = tf.multiply(tf.square(delta2), tf.sqrt(1. + tf.square((target - x_vals) / delta2)) - 1.)
phuber2_y_out = sess.run(phuber2_y_als)

x_vals = tf.linspace(-3., 5., 500)
target = tf.constant(1.)
targets = tf.fill([500, ], 1.)

hinge_y_vals = tf.maximum(0., 1. - tf.multiply(target, x_vals))
hinge_y_out = sess.run(hinge_y_vals)

# [i for i  in xentropy_y_out if not sess.run(tf.is_nan(i))]
xentropy_y_vals = -tf.multiply(target, tf.log(x_vals)) - tf.multiply((1. - target), tf.log(1. - x_vals))
xentropy_y_out = sess.run(xentropy_y_vals)
not_nan = [i for i in xentropy_y_out if not sess.run(tf.is_nan(i))]

# logits and targets must have the same type and shape.
# ValueError: Only call `sigmoid_cross_entropy_with_logits` with named arguments (labels=..., logits=..., ...)
# xentropy_sigmoid_y_vals = tf.nn.sigmoid_cross_entropy_with_logits(x_vals, logits=targets)
# sess.run(tf.nn.sigmoid_cross_entropy_with_logits(labels=x_vals, logits=targets))
xentropy_sigmoid_y_vals = tf.nn.sigmoid_cross_entropy_with_logits(labels=x_vals, logits=targets)
xentropy_sigmoid_y_out = sess.run(xentropy_sigmoid_y_vals)

weight = tf.constant(0.5)
xentropy_weigthed_y_vals = tf.nn.weighted_cross_entropy_with_logits(x_vals, targets, weight)
xentropy_weigthed_y_out = sess.run(xentropy_weigthed_y_vals)

unscaled_logits = tf.constant([1., -3., 10.])
target_dist = tf.constant([0.1, 0.02, 0.88])
softmax_xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=unscaled_logits, logits=target_dist)
print(sess.run(softmax_xentropy))
softmax_xentropy_out = sess.run(softmax_xentropy)

unscaled_logits = tf.constant([1., -3., 10.])
sparse_target_dist = tf.constant([2])
sparse_xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=unscaled_logits, logits=sparse_target_dist)
print(sess.run(sparse_xentropy))
sparse_xentropy_out = sess.run(sparse_xentropy)
dd = 9