tensorflow: 激活函数(Activation_Functions) 探究
tensorflow: 激活函数(Activation_Functions) 探究
JNingWei
发布于 2018-09-28 15:57:07
发布于 2018-09-28 15:57:07
文章被收录于专栏：JNing的专栏
激活函数概念
From TensorFlow - Activation_Functions：
在神经网络中，我们有很多的 非线性函数 来作为 激活函数
连续 、平滑
tf.sigmoid(x, name = None) == 1 / (1 + exp(-x))
import numpy as np
import tensorflow as tf
sess = tf.Session()
bn = np.random.normal(0, 5, [3, 5])
print bn.shape, type(bn), ':'
print bn
print
output = tf.nn.sigmoid(bn)
print output.shape, type(output), ':'
print sess.run(output)(3, 5) <type 'numpy.ndarray'> :
[[ 2.42429203 -1.89521415 4.52536321 2.02200042 -0.46109594]
[-5.37984794 3.82258344 3.05039891 5.35911657 4.04462726]
[-3.79266918 -7.12570837 1.74167827 -0.85649631 -3.77669239]]
(3, 5) <class 'tensorflow.python.framework.ops.Tensor'> :
[[ 9.18661034e-01 1.30651098e-01 9.89285269e-01 8.83087698e-01
3.86725869e-01]
[ 4.58738158e-03 9.78596887e-01 9.54799746e-01 9.95316973e-01
9.82785304e-01]
[ 2.20387197e-02 8.03517003e-04 8.50900112e-01 2.98071887e-01
2.23857102e-02]]tf.tanh(x, name = None) == ( exp(x) - exp(-x) ) / ( exp(x) + exp(-x) )
import numpy as np
import tensorflow as tf
sess = tf.Session()
bn = np.random.normal(0, 5, [3, 5])
print bn.shape, type(bn), ':'
print bn
print
output = tf.nn.tanh(bn)
print output.shape, type(output), ':'
print sess.run(output)(3, 5) <type 'numpy.ndarray'> :
[[-1.43756487 -0.82183219 2.83650212 -0.86855883 -2.54894335]
[ 2.3639829 -5.23813843 6.94823124 -6.59737671 3.62198313]
[ 9.15073151 2.82883771 -4.40860502 -5.96409016 -2.74915937]]
(3, 5) <class 'tensorflow.python.framework.ops.Tensor'> :
[[-0.89320646 -0.67606587 0.99314851 -0.70064117 -0.98785491]
[ 0.98246619 -0.99994361 0.99999816 -0.99999628 0.99857208]
[ 0.99999998 0.99304304 -0.99970372 -0.9999868 -0.99184608]]tf.nn.softplus(features, name = None) == log ( exp( features ) + 1)
import numpy as np
import tensorflow as tf
sess = tf.Session()
bn = np.random.normal(0, 5, [3, 5])
print bn.shape, type(bn), ':'
print bn
print
output = tf.nn.softplus(bn)
print output.shape, type(output), ':'
print sess.run(output)(3, 5) <type 'numpy.ndarray'> :
[[ 2.3897838 -9.86605463 -7.58004249 -4.38702367 -1.44367065]
[ 7.52588384 6.49497224 -4.37733996 -0.68677868 -2.12110005]
[-6.35464811 -1.70150615 6.51252343 -0.12833586 4.36898049]]
(3, 5) <class 'tensorflow.python.framework.ops.Tensor'> :
[[ 2.47747365e+00 5.19057707e-05 5.10409248e-04 1.23609802e-02
2.11928637e-01]
[ 7.52642265e+00 6.49648212e+00 1.24805143e-02 4.07592450e-01
1.13239092e-01]
[ 1.73713721e-03 1.67553530e-01 6.51400705e+00 6.31036601e-01
4.38156511e+00]]连续、不平滑
tf.nn.relu(features, name = None) == max (features, 0)
import numpy as np
import tensorflow as tf
sess = tf.Session()
bn = np.random.normal(0, 5, [3, 5])
print bn.shape, type(bn), ':'
print bn
print
output = tf.nn.relu(bn)
print output.shape, type(output), ':'
print sess.run(output)(3, 5) <type 'numpy.ndarray'> :
[[ 4.34288636 -3.14906286 -5.21796011 -2.77006242 -4.92871322]
[ 9.07049557 -9.64290379 -5.91523423 1.59385546 -2.04672855]
[-4.10765782 1.51740207 -0.5572445 8.21818142 -4.67065521]]
(3, 5) <class 'tensorflow.python.framework.ops.Tensor'> :
[[ 4.34288636 0. 0. 0. 0. ]
[ 9.07049557 0. 0. 1.59385546 0. ]
[ 0. 1.51740207 0. 8.21818142 0. ]]tf.nn.relu6(features, name = None) == min ( max(features, 0), 6 )
import numpy as np
import tensorflow as tf
sess = tf.Session()
bn = np.random.normal(0, 5, [3, 5])
print bn.shape, type(bn), ':'
print bn
print
output = tf.nn.relu6(bn)
print output.shape, type(output), ':'
print sess.run(output)(3, 5) <type 'numpy.ndarray'> :
[[ 6.08205437 7.72360999 -1.62220085 5.41621866 5.8087728 ]
[-5.07454654 3.85471614 1.44742944 2.77378759 3.61971044]
[ 5.43383943 1.9598894 -2.5352505 -1.38550512 3.64028622]]
(3, 5) <class 'tensorflow.python.framework.ops.Tensor'> :
[[ 6. 6. 0. 5.41621866 5.8087728 ]
[ 0. 3.85471614 1.44742944 2.77378759 3.61971044]
[ 5.43383943 1.9598894 0. 0. 3.64028622]]tf.nn.bias_add(value, bias, name = None) == value + bias (bias是一维的)
import numpy as np
import tensorflow as tf
sess = tf.Session()
bn = np.random.normal(0, 5, [3, 5])
print bn.shape, type(bn), ':'
print bn
print
output = tf.nn.bias_add(value=bn, bias=np.ones_like(bn[0]))
print output.shape, type(output), ':'
print sess.run(output)(3, 5) <type 'numpy.ndarray'> :
[[-7.24470546 1.40561024 2.27976912 -6.22879516 4.98934916]
[-9.75160657 6.78796922 0.60843038 -4.94145474 -0.98402315]
[-7.02590057 1.98236592 0.85727947 0.08917467 -5.54994355]]
(3, 5) <class 'tensorflow.python.framework.ops.Tensor'> :
[[-6.24470546 2.40561024 3.27976912 -5.22879516 5.98934916]
[-8.75160657 7.78796922 1.60843038 -3.94145474 0.01597685]
[-6.02590057 2.98236592 1.85727947 1.08917467 -4.54994355]]随机正则化
tf.nn.dropout(x, keep_prob, noise_shape = None, seed = None, name = None) == keep_prob概率 的神经元输出值将被放大到原来的 1/keep_prob 倍，其余神经元的输出置 0
import numpy as np
import tensorflow as tf
sess = tf.Session()
bn = np.random.normal(0, 5, [3, 5])
print bn.shape, type(bn), ':'
print bn
print
output = tf.nn.dropout(x=bn, keep_prob=0.5)
print output.shape, type(output), ':'
print sess.run(output)(3, 5) <type 'numpy.ndarray'> :
[[ -6.63260663 5.18248388 -2.64777118 -0.98104194 -4.21568201]
[ 0.94315835 5.73277238 -0.27942206 0.93593509 10.41087634]
[ 0.18322279 5.72198372 5.00533604 -1.80672579 -2.32201658]]
(3, 5) <class 'tensorflow.python.framework.ops.Tensor'> :
[[ -0. 0. -5.29554235 -1.96208389 -0. ]
[ 1.88631669 0. -0.55884411 1.87187017 20.82175267]
[ 0. 11.44396745 0. -0. -0. ]]本文参与 腾讯云自媒体同步曝光计划，分享自作者个人站点/博客。
原始发表：2017年09月05日，如有侵权请联系 cloudcommunity@tencent.com 删除
评论
登录后参与评论
推荐阅读
目录
腾讯云开发者
