Tensorflow embedding lookup using onehot encoding

Question

I currently have onehot encoddings that I want to use embeddings for. However when I call

embed=tf.nn.embedding_lookup(embeddings, train_data) 
print(embed.get_shape())

embed data shape (11, 32, 729, 128)

This shape should be (11, 32, 128) but it gives me the wrong dimensions because train_data is onehot encoded.

train_data2=tf.matmul(train_data,tf.range(729))

give me error:

ValueError: Shape must be rank 2 but is rank 3

Help me out please! Thanks.

Answer 1

A small fix to your example:

encoding_size = 4
one_hot_batch = tf.constant([[0, 0, 0, 1], [0, 1, 0, 0], [1, 0, 0, 0]])
one_hot_indexes = tf.matmul(one_hot_batch, np.array([range(encoding_size)], 
    dtype=np.int32).T)

with tf.Session() as session:
  print one_hot_indexes.eval()

Another way:

batch_size = 3
one_hot_batch = tf.constant([[0, 0, 0, 1], [0, 1, 0, 0], [1, 0, 0, 0]])
one_hot_indexes = tf.where(tf.not_equal(one_hot_batch, 0))
one_hot_indexes = one_hot_indexes[:, 1]
one_hot_indexes = tf.reshape(one_hot_indexes, [batch_size, 1])
with tf.Session() as session:
  print one_hot_indexes.eval()

Result in both cases:

[[3]
 [1]
 [0]]