Struct rust_bert::xlnet::XLNetLMHeadModel

pub struct XLNetLMHeadModel { /* fields omitted */ }

XLNetLMHeadModel

XLNet model with a language model head for language generation tasks It is made of the following blocks:

• base_model: XLNetModel
• lm_head: Linear language modeling head, projecting the hidden state logits to the vocabulary space

Implementations

impl XLNetLMHeadModel

pub fn new<'p, P>(p: P, config: &XLNetConfig) -> XLNetLMHeadModel where
    P: Borrow<Path<'p>>, 

Build a new XLNetLMHeadModel

Arguments

• p - Variable store path for the root of the XLNet model
• config - XLNetConfig object defining the model architecture

Example

use rust_bert::xlnet::{XLNetConfig, XLNetLMHeadModel};
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 = XLNetConfig::from_file(config_path);
let xlnet_model = XLNetLMHeadModel::new(&p.root(), &config);

pub fn forward_t(
    &self,
    input_ids: Option<&Tensor>,
    attention_mask: Option<&Tensor>,
    old_layer_states: Option<Vec<Option<LayerState>>>,
    perm_mask: Option<&Tensor>,
    target_mapping: Option<&Tensor>,
    token_type_ids: Option<&Tensor>,
    input_embeds: Option<Tensor>,
    train: bool
) -> Result<LMModelOutput, RustBertError>

Forward pass through the model

Arguments

• input_ids - Optional input tensor of shape (batch size, sequence_length). This or input_embeds must be provided.
• attention_mask - Optional attention mask of shape (batch size, sequence_length) for the encoder positions. Positions with a mask with value 0 will be masked.
• perm_mask - Optional tensor of shape (batch size, sequence_length, sequence_length). Mask to indicate the attention pattern for each input token (only used for pre-training over permutations, rather than simple token masking).
• target_mapping - Optional tensor of shape (batch size, num_tokens, sequence_length) indicating the position of the masked words to predict.
• token_type_ids - Optional tensor (batch size, sequence_length) indicating the sentence ID of the token (0: first sentence, 1: second sentence).
• input_embeds - Optional input tensor of shape (batch size, sequence_length, embeddings dimension). This or input_ids must be provided.
• old_layer_states - Optional vector of length num_layers containing optional LayerStates containing the last calculated content for the attention layers. This avoids recomputing attention weights at past positions and speeds up decoding.
• 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 - XLNetCache made of Option<Vec<Option<LayerState>>> of length n_layers and shape (past_sequence_length, batch size, hidden_size) containing the previous content
  • encoder_hidden_states - None
  • 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)

Example

use rust_bert::xlnet::{XLNetConfig, XLNetLMHeadModel};
let (batch_size, sequence_length) = (64, 128);
let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device));
let attention_mask = Tensor::ones(&[batch_size, sequence_length], (Int64, device));
let target_tensor = Tensor::ones(&[batch_size, sequence_length], (Int64, device));
let target_mapping = Tensor::zeros(&[64, 1, 128], (Kind::Float, device));
let _ = target_mapping.narrow(2, 3, 1).fill_(1.0);

let model_output = no_grad(|| {
    xlnet_model.forward_t(
        Some(&input_tensor),
        Some(&attention_mask),
        None,
        Some(&target_mapping),
        None,
        None,
        None,
        false,
    )
});

Trait Implementations

impl LMHeadModel for XLNetLMHeadModel

fn forward_t(
    &self,
    input_ids: &Option<Tensor>,
    layer_past: 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). This or input_embeds must be provided.
• attention_mask - Optional attention mask of shape (batch size, sequence_length) for the encoder positions. Positions with a mask with value 0 will be masked.
• perm_mask - Optional tensor of shape (batch size, sequence_length, sequence_length). Mask to indicate the attention pattern for each input token (only used for pre-training over permutations, rather than simple token masking).
• target_mapping - Optional tensor of shape (batch size, num_tokens, sequence_length) indicating the position of the masked words to predict.
• token_type_ids - Optional tensor (batch size, sequence_length) indicating the sentence ID of the token (0: first sentence, 1: second sentence).
• input_embeds - Optional input tensor of shape (batch size, sequence_length, embeddings dimension). This or input_ids must be provided.
• old_layer_states - Optional vector of length num_layers containing optional LayerStates containing the last calculated content for the attention layers. This avoids recomputing attention weights at past positions and speeds up decoding.
• 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 - XLNetCache made of Option<Vec<Option<LayerState>>> of length n_layers and shape (past_sequence_length, batch size, hidden_size) containing the previous content

Example

use rust_bert::xlnet::{XLNetConfig, XLNetLMHeadModel};
let (batch_size, sequence_length) = (64, 128);
let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device));
let attention_mask = Tensor::ones(&[batch_size, sequence_length], (Int64, device));
let target_tensor = Tensor::ones(&[batch_size, sequence_length], (Int64, device));
let target_mapping = Tensor::zeros(&[64, 1, 128], (Kind::Float, device));
let _ = target_mapping.narrow(2, 3, 1).fill_(1.0);

let model_output = no_grad(|| {
    xlnet_model.forward_t(
        Some(&input_tensor),
        Some(&attention_mask),
        None,
        Some(&target_mapping),
        None,
        None,
        None,
        false,
    )
});

impl LanguageGenerator<XLNetLMHeadModel, XLNetVocab, XLNetTokenizer> for XLNetGenerator

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.

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).

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>>