Struct rust_bert::t5::T5Model

pub struct T5Model { /* fields omitted */ }

T5 Base model

Base architecture for T5 model. Usually complemented with a task-specific head, such as a language model head. It is made of the following blocks:

• encoder: T5Stack (transformer) made of a vector of encoding layers
• decoder: T5Stack (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)
• embeddings: nn::Embedding Shared embeddings for the encoder and decoder.

Implementations

impl T5Model

pub fn new<'p, P>(
    p: P,
    config: &T5Config,
    output_attentions: bool,
    output_hidden_states: bool
) -> T5Model where
    P: Borrow<Path<'p>>, 

Build a new T5Model

Arguments

• p - Variable store path for the root of the BART model
• config - T5Config object defining the model architecture
• output_attention - flag indicating if the model should output the attention weights of intermediate layers
• output_hidden_states - flag indicating if the model should output the hidden states weights of intermediate layers

Example

use rust_bert::t5::{T5Config, T5Model};
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 = T5Config::from_file(config_path);
let output_attentions = true;
let output_hidden_states = true;
let t5: T5Model = T5Model::new(
    &p.root() / "t5",
    &config,
    output_attentions,
    output_hidden_states,
);

pub fn forward_t(
    &self,
    input_ids: Option<&Tensor>,
    attention_mask: Option<&Tensor>,
    encoder_outputs: Option<(Tensor, Option<Vec<Tensor>>, Option<Vec<Tensor>>)>,
    decoder_input_ids: Option<&Tensor>,
    decoder_attention_mask: Option<&Tensor>,
    input_embeds: Option<Tensor>,
    decoder_input_embeds: Option<Tensor>,
    old_layer_states: Option<Vec<(Option<LayerState>, Option<LayerState>)>>,
    train: bool
) -> (Tensor, Tensor, Option<Vec<(Option<LayerState>, Option<LayerState>)>>, Option<Vec<Tensor>>, Option<Vec<Tensor>>, Option<Vec<Tensor>>, Option<Vec<Tensor>>)

Forward pass through the model

Arguments

• input_ids - Optional input tensor of shape (batch size, source_sequence_length). This or input_embeds must be provided.
• 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). This or decoder_input_embeds must be provided.
• 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.
• input_embeds - Optional input tensor of shape (batch size, source_sequence_length, embeddings dimension). This or input_ids must be provided.
• decoder_input_embeds - Optional input tensor of shape (batch size, target_sequence_length, embeddings dimension). This or decoder_input_ids must be provided.
• old_layer_states - Optional vector of length num_layers containing tuples of optional LayerStates containing th elast calculated key and value pairs for the decoder. 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

• decoder_output - Tensor of shape (batch size, target_sequence_length, hidden_size) representing the activations of the last decoder hidden state
• encoder_hidden_states - Tensor of shape (batch size, source_sequence_length, hidden_size) representing the activations of the last encoder hidden state
• decoder_cache - Option<Vec<(Option<Vec<&LayerState, &LayerState>>)>> of length n_layer containing the encoder 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::t5::{T5Config, T5Model};
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 (
    decoder_output,
    encoder_hidden_states,
    decoder_cache,
    all_encoder_hidden_states,
    all_encoder_attentions,
    all_decoder_hidden_states,
    all_decoder_attentions,
) = no_grad(|| {
    t5_model.forward_t(
        Some(&input_tensor),
        Some(&encoder_attention_mask),
        None,
        Some(&target_tensor),
        Some(&decoder_attention_mask),
        None,
        None,
        None,
        false,
    )
});