tensorflow离散采样tf.multinomial、tf.distributions.Categorical

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# &＃64;Time : 2021/10/21 21:51
# &＃64;Author : XXX
# &＃64;File : test_multinomial.py"""
记&＃xff1a;
logits: 为网络最后一层的输出&＃xff0c;无激活函数
p &＃61; softmax(logits): 概率
log_p &＃61; log_softmax(logits) &＃61; log(p): log概率
结论&＃xff1a;
"""
import tensorflow as tf
import numpy as nptf.random.set_random_seed(1111)
logits &＃61; tf.constant([[5., 3, 2]])
p &＃61; tf.nn.softmax(logits, axis&＃61;1)
log_p &＃61; tf.nn.log_softmax(logits, axis&＃61;1)# 对tf.multinomial&＃xff0c;输出ndarray,(batch_size, num_samples)
# sample &＃61; tf.multinomial(logits, 10000) # [8423. 1157. 420.]
# sample &＃61; tf.multinomial(log_p, 10000) # [8423. 1157. 420.]# 对tf.distributions.Categorical&＃xff0c;输出ndarray, (num_samples, batch_size)
# 参数1: logits&＃xff0c;得赋值为开头定义的logits或者log_p
# sample &＃61; tf.distributions.Categorical(logits&＃61;logits).sample((1, 10000)) # [8457. 1155. 388.]
# sample &＃61; tf.distributions.Categorical(logits&＃61;log_p).sample((1, 10000)) # [8457. 1155. 388.]
# 参数2&＃xff1a;probs
sample &＃61; tf.distributions.Categorical(probs&＃61;p).sample(sample_shape&＃61;(10000, )) # [8457. 1155. 388.]sess &＃61; tf.Session()
s &＃61; sess.run(sample)cnt &＃61; np.zeros((3, ))
for i in s:cnt[i] &＃43;&＃61; 1print(cnt)
print(sess.run(p)) # [[0.8437947 0.11419519 0.04201007]]