tensorflow 2 中是否有一个函数，我可以在其中创建一个一维张量和一个零之间距离相等的一维张量？

Question

我正在创建一个尖峰神经网络，并希望创建一个有 1000 个时间步长的尖峰序列，每 x 个时间步长出现尖峰。

inp_spikes = tf.reshape(tf.random.categorical(tf.math.log([[((1000-49)/1000), (49/1000)]]), 1*1000), [1, 1000])

上面的代码会给我一个张量，其中大约有 49 个，其余为零，但它是随机的。为了再现性，我想把随机数去掉，把它们平均分布。

Answer 1

我假设你想要一个长度为 1000 的张量，每 20 个索引有一个 1。

可能有更简单的方法，但一种方法是只使用 range、mod、与 0 的比较和 cast :

inp_spikes = tf.cast(tf.less_equal(tf.mod(tf.range(1000), 20), 0), tf.int32)

或者您可以使用 tf.scatter_nd:

list_of_inds = list(range(0, 1000, 20))
indices = tf.constant([[i] for i in list_of_inds])
updates = tf.constant([1] * len(list_of_inds))
shape = tf.constant([1000])
inp_spikes = tf.scatter_nd(indices, updates, shape)

Answer 2

由于你想要的张量中的非零值相对较少，所以最好将其定义为稀疏张量：

import tensorflow as tf

x = 20  # timesteps between spikes and index of first spike
n = 1000  # overall number of timesteps
s = tf.math.floordiv(n, x)  # number of spikes

indices = tf.cast(tf.linspace([x], [n-1], s), tf.int64)  # indices of spikes
values = tf.repeat(1, s)  # values of spikes
dense_shape = [n]  # the dense_shape of the sparse tensor

inp_spikes = tf.sparse.SparseTensor(indices=indices, values=values, dense_shape=dense_shape)

print(tf.sparse.to_dense(inp_spikes))  # just to illustrate the result

tf.Tensor(
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 1], shape=(1000,), dtype=int32)

tensorflow 2 中是否有一个函数，我可以在其中创建一个一维张量和一个零之间距离相等的一维张量？

Is there a function in tensorflow 2 where i can create a one dimensional tensor of ones and zeros with equal distance between ones?

python

tensorflow

tensorflow2.0