Struct rust_bert::bart::BartForConditionalGeneration

pub struct BartForConditionalGeneration { /* fields omitted */ }

BART Model for conditional generation

BART model with a vocabulary decoding head It is made of the following blocks:

• base_model: BartModel Base BART model
• linear: Linear layer without bias tied to the weights of the token id embeddings

Implementations

impl BartForConditionalGeneration

pub fn new<'p, P>(
    p: P,
    config: &BartConfig,
    generation_mode: bool
) -> BartForConditionalGeneration where
    P: Borrow<Path<'p>>, 

Build a new BartForConditionalGeneration

Arguments

• p - Variable store path for the root of the BART model
• config - BartConfig object defining the model architecture
• generation_mode - flag indicating if the model should run in generation mode (a decoder start token must then be provided)

Example

use rust_bert::bart::{BartConfig, BartForConditionalGeneration};
use rust_bert::Config;
use std::path::Path;
use tch::{nn, Device};

let config_path = Path::new("path/to/config.json");
let device = Device::Cpu;
let p = nn::VarStore::new(device);
let config = BartConfig::from_file(config_path);
let generation_mode = true;
let bart: BartForConditionalGeneration =
    BartForConditionalGeneration::new(&p.root() / "bart", &config, generation_mode);

pub fn forward_t(
    &self,
    input_ids: Option<&Tensor>,
    attention_mask: Option<&Tensor>,
    encoder_output: Option<BartEncoderOutput>,
    decoder_input_ids: Option<&Tensor>,
    decoder_attention_mask: Option<&Tensor>,
    old_layer_states: Option<Vec<(Option<LayerState>, Option<LayerState>)>>,
    train: bool
) -> BartModelOutput

Forward pass through the model

Arguments

• input_ids - Optional input tensor of shape (batch size, source_sequence_length). Must be provided when not running in generation mode
• attention_mask - Optional attention mask of shape (batch size, source_sequence_length) for the encoder positions. Positions with a mask with value 0 will be masked.
• encoder_outputs - Optional tuple made of a tensor of shape (batch size, source_sequence_length, encoder_hidden_dim) and optional vectors of tensors of length num_encoder_layers with shape (batch size, source_sequence_length, hidden_size). These correspond to the encoder last hidden state and optional hidden states/attention weights for encoder layers. When provided, the encoder hidden state will not be recalculated. Useful for generation tasks.
• decoder_input_ids - Optional input tensor of shape (batch size, target_sequence_length). Must be provided when running in generation mode (e.g. initialiazed with a BOS token)
• decoder_attention_mask - Optional attention mask of shape (batch size, target_sequence_length) for the decoder positions. Positions with a mask with value 0 will be masked.
• train - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference.

Returns

• BartModelOutput containing:
  • decoder_output - Tensor of shape (batch size, target_sequence_length, vocab_size) representing the logits for each vocabulary item and position
  • encoder_hidden_states - Tensor of shape (batch size, source_sequence_length, hidden_size) representing the activations of the last encoder hidden state
  • cache - (Option<Tensor>, Option<Vec<&LayerState, &LayerState>>) of length n_layer containing the encoder padding mask and past keys and values for both the self attention and the encoder cross attention of each layer of the decoder.
  • all_encoder_hidden_states - Option<Vec<Tensor>> of length num_encoder_layers with shape (batch size, source_sequence_length, hidden_size)
  • all_encoder_attentions - Option<Vec<Tensor>> of length num_encoder_layers with shape (batch size, source_sequence_length, hidden_size)
  • all_decoder_hidden_states - Option<Vec<Tensor>> of length num_decoder_layers with shape (batch size, target_sequence_length, hidden_size)
  • all_decoder_attentions - Option<Vec<Tensor>> of length num_decoder_layers with shape (batch size, target_sequence_length, hidden_size)

Example

use rust_bert::bart::{BartConfig, BartForConditionalGeneration};
 let (batch_size, source_sequence_length, target_sequence_length) = (64, 128, 56);
 let input_tensor = Tensor::rand(&[batch_size, source_sequence_length], (Int64, device));
 let target_tensor = Tensor::rand(&[batch_size, target_sequence_length], (Int64, device));
 let encoder_attention_mask = Tensor::ones(&[batch_size, source_sequence_length], (Int64, device));
 let decoder_attention_mask = Tensor::ones(&[batch_size, source_sequence_length], (Int64, device));

 let model_output = no_grad(|| {
   bart_model
        .forward_t(Some(&input_tensor),
                   Some(&encoder_attention_mask),
                   None,
                   Some(&target_tensor),
                   Some(&decoder_attention_mask),
                   None,
                   false)
   });

pub fn encode(
    &self,
    input_ids: &Tensor,
    attention_mask: Option<&Tensor>
) -> Tensor

Trait Implementations

impl LMHeadModel for BartForConditionalGeneration

pub fn forward_t(
    &self,
    input_ids: &Option<Tensor>,
    cache: Cache,
    attention_mask: &Option<Tensor>,
    _token_type_ids: &Option<Tensor>,
    _position_ids: &Option<Tensor>,
    _input_embeds: &Option<Tensor>,
    encoder_outputs: Option<&Tensor>,
    decoder_input_ids: &Option<Tensor>,
    train: bool
) -> Result<LMModelOutput, RustBertError>

Forward pass through the model

Arguments

• input_ids - Optional input tensor of shape (batch size, sequence_length). If None, pre-computed embeddings must be provided (see input_embeds)
• layer_past - Optional vector of length num_layers containing tuples of optional LayerStates containing th elast calculated key and value pairs for the decoder. This avoids recomputing attention weights at past positions and speeds up decoding.
• attention_mask - Optional mask of shape (batch size, sequence_length). Masked position have value 0, non-masked value 1. If None set to 1
• input_embeds - Unused for BART
• token_type_ids - Unused for BART
• position_ids - Unused for BART
• encoder_outputs - Optional tensor of shape (batch size, source_sequence_length, hidden_size). When provided, the encoder hidden state will not be recalculated. Useful for generation tasks.
• decoder_input_ids - Optional input tensor of shape (batch size, target_sequence_length). Must be provided when running in generation mode (e.g. initialized with a BOS token)
• train - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference.

Returns

• LMModelOutput containing:
  • lm_logits - Tensor of shape (batch size, sequence_length, vocab_size) representing the logits for each vocab item and position
  • cache - BartCache made of Option<Vec<(Option<Vec<&LayerState, &LayerState>>)>> of length n_layer containing the encoder past keys and values for both the self attention and the encoder cross attention of each layer of the decoder.
  • encoder_hidden_states - Option<Tensor> Hidden states for the encoder
  • all_hidden_states - None
  • all_attentions - None

Example

use rust_bert::pipelines::generation_utils::LMHeadModel;
use rust_bert::bart::{BartForConditionalGeneration, BartConfig};
 let (batch_size, source_sequence_length, target_sequence_length) = (64, 128, 56);
 let input_tensor = Tensor::rand(&[batch_size, source_sequence_length], (Int64, device));
 let target_tensor = Tensor::rand(&[batch_size, target_sequence_length], (Int64, device));
 let encoder_attention_mask = Tensor::ones(&[batch_size, source_sequence_length], (Int64, device));
 let decoder_attention_mask = Tensor::ones(&[batch_size, source_sequence_length], (Int64, device));

 let model_output = no_grad(|| {
   bart_model
        .forward_t(Some(&input_tensor),
                   Some(&encoder_attention_mask),
                   None,
                   Some(&target_tensor),
                   Some(&decoder_attention_mask),
                   None,
                   false)
   });

impl LanguageGenerator<BartForConditionalGeneration, RobertaVocab, RobertaTokenizer> for BartGenerator

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>>