Struct rust_bert::pipelines::generation_utils::MarianGenerator

pub struct MarianGenerator { /* fields omitted */ }

Language generation model based on the Marian architecture for machine translation

Implementations

impl MarianGenerator

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

Build a new marianGenerator

Arguments

• vocab_path - Path to the model vocabulary, expected to have a structure following the Transformers library convention
• sentencepiece_model_path - Path to the sentencepiece model (native protobuf expected)
• 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::pipelines::generation_utils::{GenerateConfig, MarianGenerator};
let device = Device::cuda_if_available();
let generate_config = GenerateConfig {
    max_length: 512,
    do_sample: true,
    num_beams: 6,
    temperature: 1.0,
    num_return_sequences: 1,
    ..Default::default()
};
let marian_generator = MarianGenerator::new(generate_config)?;

Trait Implementations

impl LanguageGenerator<MarianForConditionalGeneration, MarianVocab, MarianTokenizer> for MarianGenerator

pub 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.

pub 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).

pub 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>>