rust-tokenizers

Rust-tokenizer offers high-performance tokenizers for modern language models, including WordPiece, Byte-Pair Encoding (BPE) and Unigram (SentencePiece) models. These tokenizers are used in the rust-bert crate. A broad range of tokenizers for state-of-the-art transformers architectures is included, including:

• Sentence Piece (unigram model)
• Sentence Piece (BPE model)
• BERT
• ALBERT
• DistilBERT
• RoBERTa
• GPT
• GPT2
• ProphetNet
• CTRL
• Pegasus
• MBart50
• M2M100
• DeBERTa
• DeBERTa (v2)

The wordpiece based tokenizers include both single-threaded and multi-threaded processing. The Byte-Pair-Encoding tokenizers favor the use of a shared cache and are only available as single-threaded tokenizers Using the tokenizers requires downloading manually the tokenizers required files (vocabulary or merge files). These can be found in the Transformers library.

The sentence piece model loads the same .model proto files as the C++ library

Usage example (Rust)

let vocab = Arc::new(rust_tokenizers::BertVocab::from_file(&vocab_path));

let test_sentence = Example::new_from_string("This is a sample sentence to be tokenized");
let bert_tokenizer: BertTokenizer = BertTokenizer::from_existing_vocab(vocab.clone());

println!("{:?}", bert_tokenizer.encode(&test_sentence.sentence_1,
                                       None,
                                       128,
                                       &TruncationStrategy::LongestFirst,
                                       0));

Python bindings set-up

Rust-tokenizer requires a rust nightly build in order to use the Python API. Building from source involves the following steps:

1. Install Rust and use the nightly tool chain
2. run python setup.py install in the /python-bindings repository. This will compile the Rust library and install the python API
3. Example use are available in the /tests folder, including benchmark and integration tests

The library is fully unit tested at the Rust level

Usage example (Python)

from rust_transformers import PyBertTokenizer
from transformers.modeling_bert import BertForSequenceClassification

rust_tokenizer = PyBertTokenizer('bert-base-uncased-vocab.txt')
model = BertForSequenceClassification.from_pretrained('bert-base-uncased', output_attentions=False).cuda()
model = model.eval()

sentence = '''For instance, on the planet Earth, man had always assumed that he was more intelligent than dolphins because 
              he had achieved so much—the wheel, New York, wars and so on—whilst all the dolphins had ever done was muck 
              about in the water having a good time. But conversely, the dolphins had always believed that they were far 
              more intelligent than man—for precisely the same reasons.'''

features = rust_tokenizer.encode(sentence, max_len=128, truncation_strategy='only_first', stride=0)
input_ids = torch.tensor([f.token_ids for f in features], dtype=torch.long).cuda()

with torch.no_grad():
    output = model(all_input_ids)[0].cpu().numpy()