TensorFlow中每个维度的元素数

你可以使用 tf.unique_with_counts :

y = tf.constant([[0,0,10,10,24,24],[99,0,0,12,12,12]], tf.int32)

out_g = []
out_c = []
#for each row
for w in tf.unstack(y,axis=0):
    # out gives unique elements in w 
    # idx gives index to the input w
    # count gives the count of each element of out in w
    out,idx, count = tf.unique_with_counts(w)

    #inverse of total non zero elements in w
    non_zero_count = 1/tf.count_nonzero(out)

    # gather the inverse of non zero unique counts
    g = tf.cast(tf.gather(1/count,idx), tf.float32) * tf.cast(tf.sign(w), tf.float32)
    out_g.append(g)
    out_c.append(non_zero_count)
out_g = tf.stack(out_g)
out_c = tf.stack(out_c)

with tf.Session() as sess:
   print(sess.run(out_g))
   print(sess.run(out_c))

#Output:

#[[0.   0.    0.5   0.5        0.5        0.5       ]
#[1.    0.    0.    0.33333334 0.33333334 0.33333334]]

# [0.5 0.5]

我不知道TensorFlow中有哪一个函数会产生这种情况,但使用列表理解来实现这一点相对简单:

import tensorflow as tf
import numpy as np

y = np.array([[ 0,  0, 10, 10, 24, 24],  [99,  0,  0, 12, 12, 12]])
y_ = [list(a) for a in y]

w1 = [[b.count(x)**(-1.0) if x != 0 else 0 for x in b ] for b in y_]
w2 = [len(set(b) - set([0]))**(-1.0) for b in y_]

w1
>>>[[0, 0, 0.5, 0.5, 0.5, 0.5], [1.0, 0, 0, 0.3333333333333333, 0.3333333333333333, 0.3333333333333333]]
w2
>>>[0.5, 0.5]

data_w1 = np.asarray(w1, np.float32)
data_w2 = np.asarray(w2, np.float32)

data_w1 = tf.convert_to_tensor(data_w1, np.float32)
data_w2 = tf.convert_to_tensor(data_w2, np.float32)