pytorch-fast-elmo

Introduction

A fast ELMo implementation with features:

• Lower execution overhead. The core components are reimplemented in Libtorch in order to reduce the Python execution overhead (45% speedup).
• A more flexible design. By redesigning the workflow, the user could extend or change the ELMo behavior easily.

Benchmark

Hardware:

• CPU: i7-7800X
• GPU: 1080Ti

Options:

• Batch size: 32
• Warm up iterations: 20
• Test iterations: 1000
• Word length: [1, 20]
• Sentence length: [1, 30]
• Random seed: 10000

Item	Mean Of Durations (ms)	cumtime(synchronize)%
Fast ELMo (CUDA, no synchronize)	31	N/A
AllenNLP ELMo (CUDA, no synchronize)	56	N/A
Fast ELMo (CUDA, synchronize)	47	26.13%
AllenNLP ELMo (CUDA, synchronize)	57	0.02%
Fast ELMo (CPU)	1277	N/A
AllenNLP ELMo (CPU)	1453	N/A

Usage

Please install torch==1.0.0 first. Then, simply run this command to install.

pip install pytorch-fast-elmo

FastElmo should have the same behavior as AllenNLP's ELMo.

from pytorch_fast_elmo import FastElmo, batch_to_char_ids

options_file = '/path/to/elmo_2x4096_512_2048cnn_2xhighway_options.json'
weight_file = '/path/to/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5'

elmo = FastElmo(options_file, weight_file)

sentences = [['First', 'sentence', '.'], ['Another', '.']]
character_ids = batch_to_ids(sentences)

embeddings = elmo(character_ids)

Use FastElmoWordEmbedding if you have disabled char_cnn in bilm-tf, or have exported the Char CNN representation to a weight file.

from pytorch_fast_elmo import FastElmoWordEmbedding, load_and_build_vocab2id, batch_to_word_ids

options_file = '/path/to/elmo_2x4096_512_2048cnn_2xhighway_options.json'
weight_file = '/path/to/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5'

vocab_file = '/path/to/vocab.txt'
embedding_file = '/path/to/cached_elmo_embedding.hdf5'

elmo = FastElmoWordEmbedding(
        options_file,
        weight_file,
        # Could be omitted if the embedding weight is in `weight_file`.
        word_embedding_weight_file=embedding_file,
)
vocab2id = load_and_build_vocab2id(vocab_file)

sentences = [['First', 'sentence', '.'], ['Another', '.']]
word_ids = batch_to_word_ids(sentences, vocab2id)

embeddings = elmo(word_ids)

CLI commands:

# Cache the Char CNN representation.
fast-elmo cache-char-cnn ./vocab.txt ./options.json ./lm_weights.hdf5 ./lm_embd.hdf5

# Export word embedding.
fast-elmo export-word-embd ./vocab.txt ./no-char-cnn.hdf5 ./embd.txt

Credits

This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.