Module rust_bert::bert

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (Devlin et al.)

Implementation of the BERT language model (https://arxiv.org/abs/1810.04805 Devlin, Chang, Lee, Toutanova, 2018). The base model is implemented in the bert::BertModel struct. Several language model heads have also been implemented, including:

• Masked language model: bert::BertForMaskedLM
• Multiple choices: bert:BertForMultipleChoice
• Question answering: bert::BertForQuestionAnswering
• Sequence classification: bert::BertForSequenceClassification
• Token classification (e.g. NER, POS tagging): bert::BertForTokenClassification

Model set-up and pre-trained weights loading

A full working example is provided in examples/bert.rs, run with cargo run --example bert. The example below illustrate a Masked language model example, the structure is similar for other models. All models expect the following resources:

• Configuration file expected to have a structure following the Transformers library
• Model weights are expected to have a structure and parameter names following the Transformers library. A conversion using the Python utility scripts is required to convert the .bin weights to the .ot format.
• BertTokenizer using a vocab.txt vocabulary Pretrained models are available and can be downloaded using RemoteResources.

use rust_tokenizers::BertTokenizer;
use tch::{nn, Device};
use rust_bert::bert::{BertConfig, BertForMaskedLM};
use rust_bert::resources::{download_resource, LocalResource, Resource};
use rust_bert::Config;

let config_resource = Resource::Local(LocalResource {
    local_path: PathBuf::from("path/to/config.json"),
});
let vocab_resource = Resource::Local(LocalResource {
    local_path: PathBuf::from("path/to/vocab.txt"),
});
let weights_resource = Resource::Local(LocalResource {
    local_path: PathBuf::from("path/to/model.ot"),
});
let config_path = download_resource(&config_resource)?;
let vocab_path = download_resource(&vocab_resource)?;
let weights_path = download_resource(&weights_resource)?;
let device = Device::cuda_if_available();
let mut vs = nn::VarStore::new(device);
let tokenizer: BertTokenizer = BertTokenizer::from_file(vocab_path.to_str().unwrap(), true);
let config = BertConfig::from_file(config_path);
let bert_model = BertForMaskedLM::new(&vs.root(), &config);
vs.load(weights_path)?;

Structs

BertConfig	BERT model configuration
BertConfigResources	BERT Pretrained model config files
BertEmbeddings	BertEmbeddings implementation for BERT model
BertForMaskedLM	BERT for masked language model
BertForMultipleChoice	BERT for multiple choices
BertForQuestionAnswering	BERT for question answering
BertForSequenceClassification	BERT for sequence classification
BertForTokenClassification	BERT for token classification (e.g. NER, POS)
BertModel	BERT Base model
BertModelResources	BERT Pretrained model weight files
BertVocabResources	BERT Pretrained model vocab files

Enums

Activation

Activation function used in the attention layer and masked language model head

Traits

BertEmbedding

BertEmbedding trait (for use in BertModel or RoBERTaModel)