Struct rust_bert::models::marian::MarianForConditionalGeneration

pub struct MarianForConditionalGeneration { /* private fields */ }

Marian Model for conditional generation

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

• base_model: BartModel Base BART model
• linear: Linear layer with bias tied to the weights of the token id embeddings

Implementations

impl MarianForConditionalGeneration

pub fn new<'p, P>(p: P, config: &MarianConfig) -> MarianForConditionalGeneration

where P: Borrow<Path<'p>>,

Build a new MarianForConditionalGeneration

Arguments

• p - Variable store path for the root of the BART model
• config - MarianConfig 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::marian::{MarianConfig, MarianForConditionalGeneration};
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 = MarianConfig::from_file(config_path);
let model = MarianForConditionalGeneration::new(&p.root(), &config);

pub fn forward_t( &self, input_ids: Option<&Tensor>, attention_mask: Option<&Tensor>, encoder_outputs: Option<&Tensor>, 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. initialized 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
  • 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_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;
use rust_bert::marian::MarianForConditionalGeneration;
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(|| {
    marian_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