Skip to main content
Help the Python Software Foundation raise $60,000 USD by December 31st!  Building the PSF Q4 Fundraiser

Handy library to work with pre-trained ELMo embeddings in TensorFlow

Project description

Simple_elmo is a Python library to work with pre-trained ELMo embeddings in TensorFlow.

This is a significantly updated wrapper to the original ELMo implementation. The main changes are:

  • more convenient and transparent data loading (including from compressed files)
  • code adapted to modern TensorFlow versions (including TensorFlow 2).


pip install --upgrade simple_elmo

Make sure to update the package regularly, we are actively developing.


from simple_elmo import ElmoModel

model = ElmoModel()


First, let's load a pretrained model from disk:


Required arguments

PATH_TO_ELMO is either a ZIP archive downloaded from the NLPL vector repository, OR a directory containing 2 files:

  • *.hdf5, pre-trained ELMo weights in HDF5 format (simple_elmo assumes the file is named model.hdf5; if it is not found, the first existing file with the .hdf5 extension will be used);
  • options.json, description of the model architecture in JSON;

One can also provide a vocab.txt/vocab.txt.gz file in the same directory: a one-word-per-line vocabulary of words to be cached (as character id representations) before inference. Even if it is not present at all, ELMo will still process all words normally. However, providing the vocabulary file can slightly increase inference speed when working with very large corpora (by reducing the amount of word to char ids conversions).

Optional arguments

  • max_batch_size: integer, default 32 the maximum number of sentences/documents in a batch during inference; your input will be automatically split into chunks of the respective size; if your computational resources allow, you might want to increase this value.
  • limit: integer, default 100 the number of words from the vocabulary file to actually cache (counted from the first line). Increase the default value if you are sure these words occur in your training data much more often than 1 or 2 times.

Working with models

Currently, we provide two methods for loaded models (will be expanded in the future):

  • model.get_elmo_vectors(SENTENCES)

  • model.get_elmo_vector_average(SENTENCES)

SENTENCES is a list of input sentences (lists of words).

The get_elmo_vectors() method produces a tensor of contextualized word embeddings. Its shape is (number of sentences, the length of the longest sentence, ELMo dimensionality).

The get_elmo_vector_average() method produces a tensor with one vector per each input sentence, constructed by averaging individual contextualized word embeddings. Its shape is (number of sentences, ELMo dimensionality).

Both methods can be used with the layers argument, which takes one of the three values:

  • average (default): return the average of all ELMo layers for each word;
  • top: return only the top (last) layer for each word;
  • all: return all ELMo layers for each word (an additional dimension appears in the produced tensor, with the shape equal to the number of layers in the model, 3 as a rule)

Use these tensors for your downstream tasks.

Example scripts

We provide three example scripts to make it easier to start using simple_elmo right away:

Inferring token embeddings

python3 -i test.txt -e ~/PATH_TO_ELMO/

This script simply returns contextualized ELMo embeddings for the words in your input sentences.

Text pairs classification

python3 -i paraphrases_lemm.tsv.gz -e ~/PATH_TO_ELMO/

This script can be used to perform document pair classification (like in text entailment or paraphrase detection). Simple average of ELMo embeddings for all words in a document is used; then, the cosine similarity between two documents is calculated and used as a single classifier feature. Evaluated with macro F1 score and 10-fold cross-validation.

Example paraphrase dataset for English (adapted from MRPC):

Example paraphrase datasets for Russian (adapted from

Word sense disambiguation

python3 -i senseval3.tsv -e ~/PATH_TO_ELMO/

This script takes as an input a word sense disambiguation (WSD) dataset and a pre-trained ELMo model. It extracts token embeddings for ambiguous words and trains a simple Logistic Regression classifier to predict word senses. Averaged macro F1 score across all words in the test set is used as the evaluation measure (with 5-fold cross-validation).

Example WSD datasets for English (adapted from Senseval 3):

Example WSD datasets for Russian (adapted from RUSSE'18):

Frequently Asked Questions

Where can I find pre-trained ELMo models?

Several repositories are available where one can download ELMo models compatible with simple_elmo:

Can I load ELMoForManyLangs models?

Unfortunately not. These models are trained using a slightly different architecture. Therefore, they are not compatible neither with AllenNLP nor with simple_elmo. You should use the original ELMoForManyLangs code to work with these models.

I see a lot of warnings about deprecated methods

This is normal. The simple_elmo library is based on the original ELMo implementation which was aimed at the versions of TensorFlow which are very outdated today. We significantly updated the code and fixed many warnings - but not all of them yet. The work continues (and will eventually lead to a complete switch to TensorFlow 2).

Meaniwhile, these warnings can be ignored: they do not harm the resulting embeddings in any way.

Can I train my own ELMo with this library?

Currently we provide ELMo training code (updated and improved in the same way compared to the original implementation) in a separate repository. It will be integrated into the simple_elmo package in the nearest future.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for simple-elmo, version 0.6.0
Filename, size File type Python version Upload date Hashes
Filename, size simple_elmo-0.6.0-py3-none-any.whl (33.0 kB) File type Wheel Python version py3 Upload date Hashes View
Filename, size simple_elmo-0.6.0.tar.gz (22.1 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page