TensorFlow中 tf.nn.embedding_lookup

Posted 2022-11-21 yqtaowhu

import tensorflow as tf  
src_vocab_size = 10
src_embed_size = 5
source = [1,3]
with tf.variable_scope("encoder"):
    embedding_encoder = tf.get_variable(
        "embedding_encoder", [src_vocab_size, src_embed_size], tf.float32)
encoder_emb_inp = tf.nn.embedding_lookup(
          embedding_encoder, source)
init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    emb_mat = sess.run(embedding_encoder)
    for line in emb_mat:
        print line
    en_input = sess.run(encoder_emb_inp)
    print
    for line in en_input:
        print line

输出结果：

[ 0.56113797  0.04369807  0.18308383 -0.48125005 -0.43450889]
[-0.6047132  -0.21060479  0.40796143 -0.40531671  0.55036896]
[ 0.31311834 -0.4060598   0.36560428  0.2722581  -0.02451819]
[ 0.18635517 -0.12266624 -0.39344144 -0.1277926  -0.45468265]
[ 0.30129766  0.56903845 -0.03529584 -0.33247966  0.45404953]
[-0.58887643  0.50933784 -0.19886917 -0.03041148 -0.44376266]
[ 0.35494697  0.25374722  0.41377074  0.06932443 -0.21179438]
[ 0.10084659 -0.60172981  0.49977249 -0.28413546 -0.33590576]
[-0.01577765  0.41795093  0.43442172  0.59790486  0.58752233]
[ 0.42998117 -0.0969131  -0.34563044  0.16796118  0.62855309]

[-0.6047132  -0.21060479  0.40796143 -0.40531671  0.55036896]
[ 0.18635517 -0.12266624 -0.39344144 -0.1277926  -0.45468265]

1、对于one-hot的编码embedding操作
2、embedding_lookup即去矩阵中的某一行，同时其不是简单的查表，id对应的向量是可以训练，即其实一个全连接
3、在分类模型中用id类的特征，注意希望模型能够记住信息，但是id的维度太高，同一个商品数量也不大，因此可以用iterm embedding来代替id