Struct rust_bert::bart::BartGenerator

pub struct BartGenerator { /* fields omitted */ }

Language generation model based on the Bart architecture

Implementations

impl BartGenerator

pub fn new(
    generate_config: GenerateConfig
) -> Result<BartGenerator, RustBertError>

Build a new BartGenerator

Arguments

• vocab_path - Path to the model vocabulary, expected to have a structure following the Transformers library convention
• merges_path - Path to the bpe merges, expected to have a structure following the Transformers library convention
• config_path - Path to the model configuration, expected to have a structure following the Transformers library convention
• weights_path - Path to the model weight files. These need to be converted form the .bin to .ot format using the utility script provided.
• device - Device to run the model on, e.g. Device::Cpu or Device::Cuda(0)

Example

use rust_bert::bart::BartGenerator;
use rust_bert::pipelines::generation_utils::GenerateConfig;
let device = Device::cuda_if_available();
let generate_config = GenerateConfig {
    max_length: 30,
    do_sample: true,
    num_beams: 5,
    temperature: 1.1,
    num_return_sequences: 3,
    ..Default::default()
};
let bart_generator = BartGenerator::new(generate_config)?;

Trait Implementations

impl LanguageGenerator<BartForConditionalGeneration, RobertaVocab, RobertaTokenizer> for BartGenerator

fn generate<'a, S>(
    &self,
    prompt_texts: Option<S>,
    attention_mask: Option<Tensor>,
    min_length: impl Into<Option<i64>>,
    max_length: impl Into<Option<i64>>,
    decoder_start_token_id: impl Into<Option<i64>>
) -> Vec<String> where
    S: AsRef<[&'a str]>, 

Generate text based on a vector of promp texts.

fn generate_indices<'a, S>(
    &self,
    prompt_texts: Option<S>,
    attention_mask: Option<Tensor>,
    min_length: impl Into<Option<i64>>,
    max_length: impl Into<Option<i64>>,
    decoder_start_token_id: impl Into<Option<i64>>
) -> Vec<Vec<i64>> where
    S: AsRef<[&'a str]>, 

Generate token indices without decoding (useful for token-level operations before returning final text or as validation step during training).

fn generate_from_ids_and_past(
    &self,
    input_ids: Tensor,
    attention_mask: Option<Tensor>,
    min_length: impl Into<Option<i64>>,
    max_length: impl Into<Option<i64>>,
    decoder_start_token_id: impl Into<Option<i64>>
) -> Vec<Vec<i64>>