Struct rust_bert::reformer::ReformerModel

pub struct ReformerModel { /* private fields */ }

Reformer Base model

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

• embeddings: ReformerEmbeddings Reformer embeddings, combining word and position embeddings
• encoder: ReformerEncoder (transformer) made of a vector of Reformer layer with local or LSH attention. caching is implemented for the decoder to avoid recalculating static states (encoder key/values and previously calculated decoder key/values)
• least_common_mult_chunk_length: least common chunk length for all attention layers
• min_chunk_length: minimum chunk length for all attention layers
• pad_token_id: padding token id used to pad to chunk length multiple if input is long enough to be chunked.

Implementations

impl ReformerModel

pub fn new<'p, P>(
    p: P,
    config: &ReformerConfig
) -> Result<ReformerModel, RustBertError>where
    P: Borrow<Path<'p>>,

Build a new ReformerModel

Arguments

• p - Variable store path for the root of the BART model
• config - ReformerConfig object defining the model architecture

Example

use rust_bert::reformer::{ReformerConfig, ReformerModel};
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 = ReformerConfig::from_file(config_path);
let reformer_model: ReformerModel =
    ReformerModel::new(&p.root() / "reformer", &config).unwrap();

pub fn forward_t(
    &self,
    input_ids: Option<&Tensor>,
    position_ids: Option<&Tensor>,
    input_embeds: Option<&Tensor>,
    attention_mask: Option<&Tensor>,
    num_hashes: Option<i64>,
    old_layer_states: Option<Vec<Option<LayerState>>>,
    train: bool
) -> Result<ReformerModelOutput, RustBertError>

Forward pass through the model

Arguments

• input_ids - Optional input tensor of shape (batch size, sequence_length). Must be provided when no pre-computed embeddings are given.
• position_ids - Optional input tensor of shape (batch size, sequence_length). If not provided will be calculated on the fly starting from position 0.
• input_embeds - Optional input tensor of shape (batch size, sequence_length, embeddings_dim). Must be provided when no input ids are given.
• attention_mask - Optional attention mask of shape (batch size, sequence_length). Positions with a mask with value 0 will be masked.
• num_hashes - Optional specification of the number of hashes to use. If not provided will use the value provided in the model configuration.
• old_layer_states - Optional cached input (Option<Vec<Option<LayerState>>>) containing previous values for the cached states and buckets.
• train - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference.

Returns

• ReformerModelOutput containing:
  • hidden_states - Tensor of shape (batch size, sequence_length, hidden_size) representing the activations of the last hidden state
  • all_hidden_states - Option<Vec<Tensor>> of length n_layers with shape (batch size, sequence_length, hidden_size)
  • all_attentions - Option<Vec<Tensor>> of length n_layers with shape (batch size, sequence_length, hidden_size)
  • cache - Option<Vec<Option<LayerState>>> of length n_layer containing values for the states and buckets for future use.

Example

use rust_bert::reformer::{ReformerConfig, ReformerModel};
let (batch_size, sequence_length) = (64, 128);
let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device));
let input_positions = Tensor::arange(sequence_length, (Kind::Int64, device))
    .unsqueeze(0)
    .expand(&[batch_size, sequence_length], true);
let attention_mask = Tensor::ones(&[batch_size, sequence_length], (Int64, device));

let model_output = no_grad(|| {
    reformer_model.forward_t(
        Some(&input_tensor),
        Some(&input_positions),
        None,
        Some(&attention_mask),
        Some(4),
        None,
        false,
    )
});