【tensorflow】tf.nn.embedding_lookup的使用

tf.nn.embedding_lookup即在给定的范围内做映射

下面直接看例子

样例1

t = np.asarray([1,2,3,0])

params = tf.constant([10,20,30,40])
embedded_inputs = tf.nn.embedding_lookup(params, t)

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

结果

[20 30 40 10]

可见结果即是按照t的顺序，将params中查找出来

样例2

t = np.asarray([1,2,3,0])

with tf.variable_scope('test', reuse=tf.AUTO_REUSE):
    t = tf.convert_to_tensor(t)
    embedding_table = tf.Variable([10,20,30,40],
        name='phone_embedding', shape=[4],
        dtype=tf.float32)
    embedded_inputs = tf.nn.embedding_lookup(embedding_table, t)

init_ops = [tf.global_variables_initializer(),
            tf.local_variables_initializer()]
with tf.Session() as sess:
    sess.run(init_ops)

    print(sess.run(t))
    print(sess.run(embedding_table))
    print(sess.run(embedded_inputs))

结果

[1 2 3 0]
[10. 20. 30. 40.]
[20. 30. 40. 10.]

样例3

索引值为矩阵

#Index t is a 2D array
t = np.asarray([1,2,3,0]).reshape([2,2])

with tf.variable_scope('test', reuse=tf.AUTO_REUSE):
    t = tf.convert_to_tensor(t)
    embedding_table = tf.get_variable(
        'phone_embedding', [4, 3],
        dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.5))
    embedded_inputs = tf.nn.embedding_lookup(embedding_table, t)

init_ops = [tf.global_variables_initializer(),
            tf.local_variables_initializer()]
with tf.Session() as sess:
    sess.run(init_ops)

    print('t = ', sess.run(t))
    print('embedding_table = ', sess.run(embedding_table))
    print('embedded_inputs = ', sess.run(embedded_inputs))

结果

t =  [[1 2]
 [3 0]]
embedding_table =  [[ 0.19044022 -0.9089187  -0.71793836]
 [-0.01651155  0.24362208 -0.02951453]
 [ 0.7932888   0.68590254  0.14450389]
 [ 0.4447717  -0.28524947 -0.79486984]]
embedded_inputs =  [[[-0.01651155  0.24362208 -0.02951453]
  [ 0.7932888   0.68590254  0.14450389]]

 [[ 0.4447717  -0.28524947 -0.79486984]
  [ 0.19044022 -0.9089187  -0.71793836]]]

t的序列即在embedding_table的第0维中查找。
[1,2]即为[ embedding_table[1] , embedding_table[2] ]
[3,0]即为[ embedding_table[3] , embedding_table[0] ]