Struct PegasusModel

pub struct PegasusModel { /* private fields */ }

Pegasus Base model

Base architecture for Pegasus model. Usually complemented with a task-specific head, such as a language model head. It is made of the following blocks:

• encoder: PegasusEncoder (transformer) made of a vector of encoding layers
• decoder: PegasusDecoder (transformer) made of a vector of decoding layers with self attention and encoder cross-attention. caching is implemented for the decoder to avoid recalculating static states (encoder key/values and previously calculated decoder key/values)

Implementations

impl PegasusModel

pub fn new<'p, P>(p: P, config: &PegasusConfig) -> PegasusModel

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

Build a new PegasusModel

Arguments

• p - Variable store path for the root of the Pegasus model
• config - PegasusConfig object defining the model architecture

Example

use rust_bert::pegasus::{PegasusConfig, PegasusModel};
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 = PegasusConfig::from_file(config_path);
let pegasus: PegasusModel = PegasusModel::new(&p.root() / "pegasus", &config);

pub fn forward_t( &self, input_ids: Option<&Tensor>, attention_mask: Option<&Tensor>, decoder_input_ids: &Tensor, encoder_output: Option<&Tensor>, decoder_attention_mask: Option<&Tensor>, 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.
• 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)
• 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_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

• PegasusModelOutput containing:
  • decoder_output - Tensor of shape (batch size, target_sequence_length, hidden_size) representing the activations of the last decoder hidden state
  • encoder_hidden_states - Option<Tensor> of shape (batch size, source_sequence_length, hidden_size) representing the activations of the last encoder hidden state if it was not provided, otherwise None
  • 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::pegasus::{PegasusConfig, PegasusModel};
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 decoder_input_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(|| {
    pegasus_model.forward_t(
        Some(&input_tensor),
        Some(&encoder_attention_mask),
        &decoder_input_tensor,
        None,
        Some(&decoder_attention_mask),
        None,
        false,
    )
});