Feature Column Pre-trained Embedding

问题:

How to use pre-trained embedding with tf.feature_column.embedding_column.

I used pre_trained embedding in tf.feature_column.embedding_column. But it doesn't work. Error is

The error is :

ValueError: initializer must be callable if specified. Embedding of column_name: itemx

Here's my code:

weight, vocab_size, emb_size = _create_pretrained_emb_from_txt(FLAGS.vocab, 
FLAGS.pre_emb)

W = tf.Variable(tf.constant(0.0, shape=[vocab_size, emb_size]),
                trainable=False, name="W")
embedding_placeholder = tf.placeholder(tf.float32, [vocab_size, emb_size])
embedding_init = W.assign(embedding_placeholder)

sess = tf.Session()
sess.run(embedding_init, feed_dict={embedding_placeholder: weight})

itemx_vocab = tf.feature_column.categorical_column_with_vocabulary_file(
    key='itemx',
    vocabulary_file=FLAGS.vocabx)

itemx_emb = tf.feature_column.embedding_column(itemx_vocab,
                                               dimension=emb_size,
                                               initializer=W,
                                               trainable=False)

I have tried initializer = lambda w:W. like this:

itemx_emb = tf.feature_column.embedding_column(itemx_vocab,
                                               dimension=emb_size,
                                               initializer=lambda w:W,
                                               trainable=False)

it reports the error:

TypeError: () got an unexpected keyword argument 'dtype'

回答1:

I also take a issue here https://github.com/tensorflow/tensorflow/issues/20663

finally I got a right way with to solve it. although. i'm not clear why answer above is not effective!! if you know the question, Thanks to give some suggestion to me!!

ok~~~~here is current solvement. Actually from here Feature Columns Embedding lookup

code:

itemx_vocab = tf.feature_column.categorical_column_with_vocabulary_file(
    key='itemx',
    vocabulary_file=FLAGS.vocabx)

embedding_initializer_x = tf.contrib.framework.load_embedding_initializer(
    ckpt_path='model.ckpt',
    embedding_tensor_name='w_in',
    new_vocab_size=itemx_vocab.vocabulary_size,
    embedding_dim=emb_size,
    old_vocab_file='FLAGS.vocab_emb',
    new_vocab_file=FLAGS.vocabx
)
itemx_emb = tf.feature_column.embedding_column(itemx_vocab,
                                               dimension=128,
                                               initializer=embedding_initializer_x,
                                               trainable=False)

回答2:

You can also wrap your array into a function like this:

some_matrix = np.array([[0,1,2],[0,2,3],[5,6,7]])

def custom_init(shape, dtype):
    return some_matrix

embedding_feature = tf.feature_column.embedding_column(itemx_vocab, 
                                                       dimension=3, 
                                                       initializer=custom_init
                                                       )

It's an hacky way but does the job.