Convert embeddings vectors (.txt format) into TensorFlowHub embedding lookup module.
- Save lookup keys and embeddings values in text file.
Keys should be in first column. Other columns treated as embedding values. Any space-like characters allowed as columns separator.
Non-existing keys will refer to "" key embeddings. You may provide embedding values for that, otherwise it will be initialized with zeros.
key1 1. 2. 3.
key2 4. 5. 6.
<UNQ> 0. -1. 0.
- Sonvert saved embeddings into TF Hub Module with "tfembedhub-convert" command.
tfembedhub-convert vectors.txt vectors-hub/- Use embedding hub via columns in your estimator.
from tfembedhub text_embedding_column, sequence_text_embedding_column
my_words_embedding = sequence_text_embedding_column(
key='sparse_key_from_features',
module_spec='path/to/my/hub'
)
# Then pass my_words_embedding to estimator "columns" list.