tf.nn.fixed_unigram_candidate_sampler解释

最近看GraphSage的代码，发现了一个函数tf.nn.fixed_unigram_candidate_sampler

理解这个函数的难点主要在： num_true、range_max， 这两个参数上，直接上代码吧

import tensorflow as tf


def test1():
    vec = tf.constant([[1, 2, 3, 4, 6]], dtype=tf.int64)
    # vec = tf.reshape(vec, [-1, 1])
    """
        如果这里用了reshape，会报错：
        
tensorflow.python.framework.errors_impl.InvalidArgumentError: true_classes must have num_true columns, expected: 1 was: 5
        
    """
    ids, _, _ = tf.nn.fixed_unigram_candidate_sampler(
        true_classes=vec,
        num_true=5,
        num_sampled=2,
        unique=False,
        range_max=5,
        vocab_file='',
        distortion=1.0,
        num_reserved_ids=0,
        num_shards=1,
        shard=0,
        unigrams=(0.1, 0.2, 0.3, 0.1, 0.3),
        # unigrams=(1, 2, 3, 1, 3),
    )

    with tf.Session() as sess:
        print(sess.run(ids))


if __name__ == '__main__':
    test1()

import tensorflow as tf


def test1():
    vec = tf.constant([[1, 2, 3, 4, 6]], dtype=tf.int64)
    vec = tf.reshape(vec, [-1, 1])

    ids, _, _ = tf.nn.fixed_unigram_candidate_sampler(
        true_classes=vec,
        num_true=1,
        num_sampled=2,
        unique=False,
        range_max=5,
        vocab_file='',
        distortion=1.0,
        num_reserved_ids=0,
        num_shards=1,
        shard=0,
        unigrams=(0.1, 0.2, 0.3, 0.1, 0.3),
        # unigrams=(1, 2, 3, 1, 3),
    )
    with tf.Session() as sess:
        print(sess.run(ids))


if __name__ == '__main__':
    test1()

参考文章： 

如何通过自定义概率分布在Tensorflow中进行采样？

勘误：

unigrams 可为float

https://www.thinbug.com/q/49713210 

tf.nn.fixed_unigram_candidate_sampler解释

https://blog.csdn.net/u011026968/article/details/88537939

TensorFlow函数教程：tf.nn.fixed_unigram_candidate_sampler

https://www.w3cschool.cn/tensorflow_python/tf_nn_fixed_unigram_candidate_sampler.html