在TensorFlow中手动初始化conv1d

Question

How can I set custom coefficients to tf.layers.conv1d . I found out how to read current coefficients, but how can I write them?

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

sess = tf.Session()
order = 5
x = np.zeros(30)
x[10] = 1
y = tf.layers.conv1d(inputs=tf.reshape(x,[1, len(x), 1]),
                     filters=1,
                     kernel_size=order,
                     padding='same')
sess.run(tf.global_variables_initializer())
y_out = sess.run(y)

# get coef
coef = sess.run(tf.all_variables()[-2].value())
print(coef.reshape(order))

Here is a link to notebook with code at google colab: https://colab.research.google.com/drive/1YNSzKmtC88b__LqYcfD-tFHFG3jOZIAz

In general, I'm interested in how to make a FIR-filter in TensorFlow.

Answer 1

I got it! There is kerner_initializer parameter.

And this is solution

init_coef = np.array([1,2,3,4,5])[::-1]
init_coef = tf.initializers.constant(init_coef)
y = tf.layers.conv1d(inputs=tf.reshape(x,[1, len(x), 1]),
                     filters=1,
                     kernel_size=order,
                     padding='same',
                     kernel_initializer=init_coef)