如何在一系列输入中绘制张量流激活函数

Question

我真的很喜欢张量流中激活函数的二维表示。例如，绘制 :

(tf.nn.sigmoid(

对于一系列输入，类似于图像： plot

Answer 1

Tensorboard 不是为这样的情节设计的，所以很尴尬，但你可以这样做：

import numpy as np
import tensorflow as tf

# Create '/tmp/activations' directory or point to some other directory
summary_writer = tf.summary.FileWriter(logdir='/tmp/activations')

x_numpy = np.linspace(start=-6.0, stop=6.0, num=120)
x = tf.constant(x_numpy)
y = tf.nn.sigmoid(x)
with tf.Session() as sess:
    y_numpy = sess.run(y)

for x_i, y_i in zip(x_numpy, y_numpy):

    # Create a tensorboard summary with sigmoid output as the value
    # and sigmoid input as a step (which is normally
    # the `global_step` tensor). Tensorboards plots step in x-axis.
    # step must be an integer. So, we multiply x_i by 1M and convert 
    # it to int. This should be precise enough. In the plot,
    # just ignore the "M" suffix.
    value = tf.Summary.Value(tag='sigmoid', simple_value=y_i) 
    summary_writer.add_event(
        event=tf.summary.Event(summary=tf.Summary(value=[value]),
                               step=int(1000000 * x_i)))

summary_writer.close()

# Point tensorboad to '/tmp/activations'

如何在一系列输入中绘制张量流激活函数

How to plot Tensor flow activation functions in a range of inputs

python

plot

representation

tensorflow