bilm-tf

Tensorflow implementation of the pretrained biLM used to compute ELMo representations from "Deep contextualized word representations".

This repository supports both training biLMs and using pre-trained models for prediction.

We also have a pytorch implementation available in AllenNLP.

You may also find it easier to use the version provided in Tensorflow Hub if you just like to make predictions.

Citation:

@inproceedings{Peters:2018,
  author={Peters, Matthew E. and  Neumann, Mark and Iyyer, Mohit and Gardner, Matt and Clark, Christopher and Lee, Kenton and Zettlemoyer, Luke},
  title={Deep contextualized word representations},
  booktitle={Proc. of NAACL},
  year={2018}
}

Installing

Install python version 3.5 or later, tensorflow version 1.2 and h5py:

pip install tensorflow-gpu==1.2 h5py
python setup.py install

Ensure the tests pass in your environment by running:

python -m unittest discover tests/

Installing with Docker

To run the image, you must use nvidia-docker, because this repository requires GPUs.

sudo nvidia-docker run -t allennlp/bilm-tf:training-gpu

Using pre-trained models

We have several different English language pre-trained biLMs available for use. Each model is specified with two separate files, a JSON formatted "options" file with hyperparameters and a hdf5 formatted file with the model weights. Links to the pre-trained models are available here.

There are three ways to integrate ELMo representations into a downstream task, depending on your use case.

1. Compute representations on the fly from raw text using character input. This is the most general method and will handle any input text. It is also the most computationally expensive.
2. Precompute and cache the context independent token representations, then compute context dependent representations using the biLSTMs for input data. This method is less computationally expensive then #1, but is only applicable with a fixed, prescribed vocabulary.
3. Precompute the representations for your entire dataset and save to a file.

We have used all of these methods in the past for various use cases. #1 is necessary for evaluating at test time on unseen data (e.g. public SQuAD leaderboard). #2 is a good compromise for large datasets where the size of the file in #3 is unfeasible (SNLI, SQuAD). #3 is a good choice for smaller datasets or in cases where you'd like to use ELMo in other frameworks.

In all cases, the process roughly follows the same steps. First, create a Batcher (or TokenBatcher for #2) to translate tokenized strings to numpy arrays of character (or token) ids. Then, load the pretrained ELMo model (class BidirectionalLanguageModel). Finally, for steps #1 and #2 use weight_layers to compute the final ELMo representations. For #3, use BidirectionalLanguageModel to write all the intermediate layers to a file.

Shape conventions

Each tokenized sentence is a list of str, with a batch of sentences a list of tokenized sentences (List[List[str]]).

The Batcher packs these into a shape (n_sentences, max_sentence_length + 2, 50) numpy array of character ids, padding on the right with 0 ids for sentences less then the maximum length. The first and last tokens for each sentence are special begin and end of sentence ids added by the Batcher.

The input character id placeholder can be dimensioned (None, None, 50), with both the batch dimension (axis=0) and time dimension (axis=1) determined for each batch, up the the maximum batch size specified in the BidirectionalLanguageModel constructor.

After running inference with the batch, the return biLM embeddings are a numpy array with shape (n_sentences, 3, max_sentence_length, 1024), after removing the special begin/end tokens.

Vocabulary file

The Batcher takes a vocabulary file as input for efficency. This is a text file, with one token per line, separated by newlines (\n). Each token in the vocabulary is cached as the appropriate 50 character id sequence once. Since the model is completely character based, tokens not in the vocabulary file are handled appropriately at run time, with a slight decrease in run time. It is recommended to always include the special <S> and </S> tokens (case sensitive) in the vocabulary file.

ELMo with character input

See usage_character.py for a detailed usage example.

ELMo with pre-computed and cached context independent token representations

To speed up model inference with a fixed, specified vocabulary, it is possible to pre-compute the context independent token representations, write them to a file, and re-use them for inference. Note that we don't support falling back to character inputs for out-of-vocabulary words, so this should only be used when the biLM is used to compute embeddings for input with a fixed, defined vocabulary.

To use this option:

1. First create a vocabulary file with all of the unique tokens in your dataset and add the special <S> and </S> tokens.
2. Run dump_token_embeddings with the full model to write the token embeddings to a hdf5 file.
3. Use TokenBatcher (instead of Batcher) with your vocabulary file, and pass use_token_inputs=False and the name of the output file from step 2 to the BidirectonalLanguageModel constructor.

See usage_token.py for a detailed usage example.

Dumping biLM embeddings for an entire dataset to a single file.

To take this option, create a text file with your tokenized dataset. Each line is one tokenized sentence (whitespace separated). Then use dump_bilm_embeddings.

The output file is hdf5 format. Each sentence in the input data is stored as a dataset with key str(sentence_id) where sentence_id is the line number in the dataset file (indexed from 0). The embeddings for each sentence are a shape (3, n_tokens, 1024) array.

See usage_cached.py for a detailed example.

Training a biLM on a new corpus

Broadly speaking, the process to train and use a new biLM is:

1. Prepare input data and a vocabulary file.
2. Train the biLM.
3. Test (compute the perplexity of) the biLM on heldout data.
4. Write out the weights from the trained biLM to a hdf5 file.
5. See the instructions above for using the output from Step #4 in downstream models.

1. Prepare input data and a vocabulary file.

To train and evaluate a biLM, you need to provide:

• a vocabulary file
• a set of training files
• a set of heldout files

The vocabulary file is a a text file with one token per line. It must also include the special tokens <S>, </S> and <UNK> (case sensitive) in the file.

IMPORTANT: the vocabulary file should be sorted in descending order by token count in your training data. The first three lines should be the special tokens (<S>, </S> and <UNK>), then the most common token in the training data, ending with the least common token.

NOTE: the vocabulary file used in training may differ from the one use for prediction.

The training data should be randomly split into many training files, each containing one slice of the data. Each file contains pre-tokenized and white space separated text, one sentence per line. Don't include the <S> or </S> tokens in your training data.

All tokenization/normalization is done before training a model, so both the vocabulary file and training files should include normalized tokens. As the default settings use a fully character based token representation, in general we do not recommend any normalization other then tokenization.

Finally, reserve a small amount of the training data as heldout data for evaluating the trained biLM.

2. Train the biLM.

The hyperparameters used to train the ELMo model can be found in bin/train_elmo.py.

The ELMo model was trained on 3 GPUs. To train a new model with the same hyperparameters, first download the training data from the 1 Billion Word Benchmark. Then download the vocabulary file. Finally, run:

export CUDA_VISIBLE_DEVICES=0,1,2
python bin/train_elmo.py \
    --train_prefix='/path/to/1-billion-word-language-modeling-benchmark-r13output/training-monolingual.tokenized.shuffled/*' \
    --vocab_file /path/to/vocab-2016-09-10.txt \
    --save_dir /output_path/to/checkpoint

3. Evaluate the trained model.

Use bin/run_test.py to evaluate a trained model, e.g.

export CUDA_VISIBLE_DEVICES=0
python bin/run_test.py \
    --test_prefix='/path/to/1-billion-word-language-modeling-benchmark-r13output/heldout-monolingual.tokenized.shuffled/news.en.heldout-000*' \
    --vocab_file /path/to/vocab-2016-09-10.txt \
    --save_dir /output_path/to/checkpoint

4. Convert the tensorflow checkpoint to hdf5 for prediction with bilm or allennlp.

First, create an options.json file for the newly trained model. To do so, follow the template in an existing file (e.g. the original options.json and modify for your hyperpararameters.

Important: always set n_characters to 262 after training (see below).

Then Run:

python bin/dump_weights.py \
    --save_dir /output_path/to/checkpoint
    --outfile /output_path/to/weights.hdf5

Frequently asked questions and other warnings

Can you provide the tensorflow checkpoint from training?

The tensorflow checkpoint is available by downloading these files:

• vocabulary
• checkpoint
• options
• 1
• 2
• 3

How to do fine tune a model on additional unlabeled data?

First download the checkpoint files above. Then prepare the dataset as described in the section "Training a biLM on a new corpus", with the exception that we will use the existing vocabulary file instead of creating a new one. Finally, use the script bin/restart.py to restart training with the existing checkpoint on the new dataset. For small datasets (e.g. < 10 million tokens) we only recommend tuning for a small number of epochs and monitoring the perplexity on a heldout set, otherwise the model will overfit the small dataset.

Are the softmax weights available?

They are available in the training checkpoint above.

Can you provide some more details about how the model was trained?

The script bin/train_elmo.py has hyperparameters for training the model. The original model was trained on 3 GTX 1080 for 10 epochs, taking about two weeks.

For input processing, we used the raw 1 Billion Word Benchmark dataset here, and the existing vocabulary of 793471 tokens, including <S>, </S> and <UNK>. You can find our vocabulary file here. At the model input, all text used the full character based representation, including tokens outside the vocab. For the softmax output we replaced OOV tokens with <UNK>.

The model was trained with a fixed size window of 20 tokens. The batches were constructed by padding sentences with <S> and </S>, then packing tokens from one or more sentences into each row to fill completely fill each batch. Partial sentences and the LSTM states were carried over from batch to batch so that the language model could use information across batches for context, but backpropogation was broken at each batch boundary.

Why do I get slightly different embeddings if I run the same text through the pre-trained model twice?

As a result of the training method (see above), the LSTMs are stateful, and carry their state forward from batch to batch. Consequently, this introduces a small amount of non-determinism, expecially for the first two batches.

Why does training seem to take forever even with my small dataset?

The number of gradient updates during training is determined by:

• the number of tokens in the training data (n_train_tokens)
• the batch size (batch_size)
• the number of epochs (n_epochs)

Be sure to set these values for your particular dataset in bin/train_elmo.py.

What's the deal with n_characters and padding?

During training, we fill each batch to exactly 20 tokens by adding <S> and </S> to each sentence, then packing tokens from one or more sentences into each row to fill completely fill each batch. As a result, we do not allocate space for a special padding token. The UnicodeCharsVocabulary that converts token strings to lists of character ids always uses a fixed number of character embeddings of n_characters=261, so always set n_characters=261 during training.

However, for prediction, we ensure each sentence is fully contained in a single batch, and as a result pad sentences of different lengths with a special padding id. This occurs in the Batcher see here. As a result, set n_characters=262 during prediction in the options.json.

How can I use ELMo to compute sentence representations?

Simple methods like average and max pooling of the word level ELMo representations across sentences works well, often outperforming supervised methods on benchmark datasets. See "Evaluation of sentence embeddings in downstream and linguistic probing tasks", Perone et al, 2018 arxiv link.