Initializing the embedding tensor with a local numpy array

Question

I want to initialize the word embedding layer from a local numpy array with the same shape, which is a pre trained embedding from another model. It is OK if I did not add the partitioner param. def word_embedding(shape, dtype=tf.float32, name='word_embedding'):

  f = open('./cnn_embed_array', 'r')
  embedding_array = pickle.load(f)
  f.close()
  print 'embedding_array loaded......'
  with tf.device('/cpu:0'), tf.variable_scope(name):
    return tf.get_variable('embedding', shape, dtype=dtype, initializer=tf.constant_initializer(embedding_array), trainable = False)

But if I add partitioner=tf.fixed_size_partitioner(20) in the tf.get_variable function, it give me error saying that the param is redundant.

partitioner param tends to accelerate the training speed. Can I add the param in some other way ?

tomhosking · Accepted Answer · 2018-03-21 16:35:34Z

1

If trainable = False, then the variable will not be learned and so partitioning won't help you. On the other hand, if you want this variable to be updated then you'll need to set trainable = True.

answered Mar 21, 2018 at 16:35

tomhosking

612 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Initializing the embedding tensor with a local numpy array

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related