Struct rust_bert::marian::MarianGenerator

pub struct MarianGenerator { /* private fields */ }

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::marian::MarianGenerator;
use rust_bert::pipelines::generation_utils::GenerateConfig;
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)?;

pub fn new_with_tokenizer(
    generate_config: GenerateConfig,
    tokenizer: TokenizerOption
) -> Result<MarianGenerator, RustBertError>

Trait Implementations

impl LanguageGenerator<MarianForConditionalGeneration, MarianVocab, MarianTokenizer> for MarianGenerator

fn generate<S>(
    &self,
    prompt_texts: Option<&[S]>,
    generate_options: Option<GenerateOptions<'_>>
) -> Vec<GeneratedTextOutput> where
    S: AsRef<str> + Sync, 

Generate text based on a vector of promp texts.

fn generate_indices<S>(
    &self,
    prompt_texts: Option<&[S]>,
    generate_options: Option<GenerateOptions<'_>>
) -> Vec<GeneratedIndicesOutput> where
    S: AsRef<str> + Sync, 

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>,
    generate_options: Option<GenerateOptions<'_>>
) -> Vec<GeneratedIndicesOutput>

Generate token indices given a list of indices (useful when the input has been pre-tokenized). Returns a list of output tokens that need to be decoded using a tokenizer.

fn get_tokenizer(&self) -> &TokenizerOption

Returns a reference to the text generator’s tokenizer

fn half(&mut self)

fn float(&mut self)

fn set_device(&mut self, device: Device)