tensorflow prelu的實現細節

tensorflow prelu的實現細節

 

output = tf.nn.leaky_relu(input, alpha=tf_gamma_data,name=name)

#tf.nn.leaky_relu 限制了tf_gamma_data在[0 1]的范圍內 

內部實現方法是 output = tf.maxmum(alpha * input, input)

alpha > 1 時，會出現，正值*alpha， 負值不變

import numpy as np
import tensorflow as tf

#bn = np.loadtxt('tfbn.txt')
bn = np.array([[-0.9, -0.9 ,-0.9],[1.1,1.1,1.1]])
print("srcdata ", bn)
gamma_data = np.array([1.205321])
print("gamma_data ", gamma_data)
tf_gamma_data = tf.Variable(gamma_data, dtype=np.float32)
input_data = tf.Variable(bn, dtype=np.float32)
tf_prelu_test = tf.nn.leaky_relu(input_data, alpha=tf_gamma_data,name=None)
#tf_prelu_test = tf.nn.relu(input_data) + tf.multiply(tf_gamma_data, -tf.nn.relu(-input_data))
#tf_prelu_test = tf.nn.relu(input_data,name=None)
init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    tf_prelu_test = sess.run(tf_prelu_test)
    print("tf_prelu_test: \n", tf_prelu_test)

 

srcdata [[-0.9 -0.9 -0.9]
[ 1.1 1.1 1.1]]
gamma_data [1.205321]
tf_prelu_test:
[[-0.9 -0.9 -0.9 ]
[ 1.3258531 1.3258531 1.3258531]]
[Finished in 2.5s]

 

使用relu來代替
output = tf.nn.relu(data) + tf.multiply(alpha, -tf.nn.relu(-data))

 

import numpy as np
import tensorflow as tf

#bn = np.loadtxt('tfbn.txt')
bn = np.array([[-0.9, -0.9 ,-0.9],[1.1,1.1,1.1]])
print("srcdata ", bn)
gamma_data = np.array([1.205321])
print("gamma_data ", gamma_data)
tf_gamma_data = tf.Variable(gamma_data, dtype=np.float32)
input_data = tf.Variable(bn, dtype=np.float32)
#tf_prelu_test = tf.nn.leaky_relu(input_data, alpha=tf_gamma_data,name=None)
tf_prelu_test = tf.nn.relu(input_data) + tf.multiply(tf_gamma_data, -tf.nn.relu(-input_data))
#tf_prelu_test = tf.nn.relu(input_data,name=None)
init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    tf_prelu_test = sess.run(tf_prelu_test)
    print("tf_prelu_test: \n", tf_prelu_test)

 

srcdata [[-0.9 -0.9 -0.9]
[ 1.1 1.1 1.1]]
gamma_data [1.205321]
tf_prelu_test:
[[-1.0847888 -1.0847888 -1.0847888]
[ 1.1 1.1 1.1 ]]
[Finished in 2.7s]