Struct rust_bert::prophetnet::ProphetNetForCausalGeneration

pub struct ProphetNetForCausalGeneration { /* fields omitted */ }

ProphetNet Model for causal generation

ProphetNet decoder with a vocabulary decoding head It is made of the following blocks:

• base_model: ProphetNetDecoder Base ProphetNet decoder
• word_embeddings: word embeddings used by the decoder
• lm_head: Linear layer without bias to project the hidden states to the vocabulary

Implementations

impl ProphetNetForCausalGeneration

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

Build a new ProphetNetForCausalGeneration

Arguments

• p - Variable store path for the root of the ProphetNet model
• config - ProphetNetConfig object defining the model architecture

Example

use rust_bert::prophetnet::{ProphetNetConfig, ProphetNetForCausalGeneration};
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 = ProphetNetConfig::from_file(config_path);
let prophetnet_model = ProphetNetForCausalGeneration::new(&p.root(), &config);

pub fn forward_t(
    &self,
    input_ids: Option<&Tensor>,
    attention_mask: Option<&Tensor>,
    input_embeds: Option<&Tensor>,
    encoder_hidden_states: Option<&Tensor>,
    encoder_attention_mask: Option<&Tensor>,
    old_layer_states: Option<Vec<(Option<LayerState>, Option<LayerState>)>>,
    train: bool
) -> Result<ProphetNetGenerationOutput, 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.
• input_embeds - Optional input tensor of shape (batch size, sequence_length, embeddings dimension). This or input_ids must be provided.
• 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)
• old_layer_states - Optional Vector Option<Vec<Option<&LayerState>, Option<&LayerState>>> of length n_layer containing tuples with the past keys and values for both the self attention and the encoder cross attention of each layer of the decoder.
• decoder_input_embeds - Optional input tensor of shape (batch size, target_sequence_length, embeddings dimension). This or decoder_input_ids must be provided.
• train - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference.

Returns

• ProphetNetGenerationOutput containing:
  • logits - Tensor of shape (batch size, target_sequence_length, vocabulary_size) representing the activations of the last hidden state for the decoder
  • ngram_logits - Tensor of shape (ngram, batch size, target_sequence_length, vocabulary_size) representing the activations of the last hidden state for the decoder ngram stream
  • next_decoder_cache - Option<Vec<Option<LayerState>>> of length n_layer containing the past content for the the attention layers with shape (past_sequence_length, batch size, hidden_size)
  • all_decoder_hidden_states - Option<Vec<Tensor>> of length n_layer with shape (batch size, target_sequence_length, hidden_size)
  • all_ngram_decoder_hidden_states - Option<Vec<Tensor>> of length n_layer with shape (ngram, batch size, target_sequence_length, hidden_size)
  • all_attentions - Option<Vec<Tensor>> of length n_layer with shape (batch size, target_sequence_length, hidden_size)
  • all_ngram_attentions - Option<Vec<Tensor>> of length n_layer with shape (ngram, batch size, target_sequence_length, hidden_size)
  • all_cross_attentions - Option<Vec<Tensor>> of length n_layer with shape (batch size, target_sequence_length, hidden_size)

Example

use rust_bert::prophetnet::{ProphetNetModel, ProphetNetConfig, ProphetNetForCausalGeneration};
let (batch_size, sequence_length, target_sequence_length) = (64, 128, 32);
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 decoder_input_ids = Tensor::ones(&[batch_size, target_sequence_length], (Kind::Float, device));

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