Struct BertLayer

pub struct BertLayer { /* private fields */ }

BERT Layer

Layer used in BERT encoders. It is made of the following blocks:

• attention: self-attention BertAttention layer
• cross_attention: (optional) cross-attention BertAttention layer (if the model is used as a decoder)
• is_decoder: flag indicating if the model is used as a decoder
• intermediate: BertIntermediate intermediate layer
• output: BertOutput output layer

Implementations

impl BertLayer

pub fn new<'p, P>(p: P, config: &BertConfig) -> BertLayer

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

Build a new BertLayer

Arguments

• p - Variable store path for the root of the BERT model
• config - BertConfig object defining the model architecture

Example

use rust_bert::bert::{BertConfig, BertLayer};
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 = BertConfig::from_file(config_path);
let layer: BertLayer = BertLayer::new(&p.root(), &config);

pub fn forward_t( &self, hidden_states: &Tensor, mask: Option<&Tensor>, encoder_hidden_states: Option<&Tensor>, encoder_mask: Option<&Tensor>, train: bool, ) -> BertLayerOutput

Forward pass through the layer

Arguments

• hidden_states - input tensor of shape (batch size, sequence_length, hidden_size).
• mask - Optional mask of shape (batch size, sequence_length). Masked position have value 0, non-masked value 1. If None set to 1
• encoder_hidden_states - Optional encoder hidden state of shape (batch size, encoder_sequence_length, hidden_size). If the model is defined as a decoder and the encoder_hidden_states is not None, used in the cross-attention layer as keys and values (query from the decoder).
• encoder_mask - Optional encoder attention mask of shape (batch size, encoder_sequence_length). If the model is defined as a decoder and the encoder_hidden_states is not None, used to mask encoder values. Positions 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

• BertLayerOutput containing:
  • hidden_state - Tensor of shape (batch size, sequence_length, hidden_size)
  • attention_scores - Option<Tensor> of shape (batch size, sequence_length, hidden_size)
  • cross_attention_scores - Option<Tensor> of shape (batch size, sequence_length, hidden_size)

Example

let layer: BertLayer = BertLayer::new(&vs.root(), &config);
let (batch_size, sequence_length, hidden_size) = (64, 128, 512);
let input_tensor = Tensor::rand(
    &[batch_size, sequence_length, hidden_size],
    (Kind::Float, device),
);
let mask = Tensor::zeros(&[batch_size, sequence_length], (Kind::Int64, device));

let layer_output = no_grad(|| layer.forward_t(&input_tensor, Some(&mask), None, None, false));