Struct rust_bert::models::gpt_j::GptJModel

pub struct GptJModel { /* private fields */ }

GPT-J Base model

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

• wte: token embeddings
• h: Encoder (transformer) made of a vector of layers. Each layer is made of a multi-head attention layer, a layer-normalization layer, and a MLP made of linear layers.
• output_past: flag indicating if the model should return a past state. This can be fed back to the model to improve the quality of text generated.
• output_hidden_states: flag indicating if the model should return all hidden states (as opposed to only the last layer)
• output_attentions: flag indicating if the model should return activation weights

Implementations

impl GptJModel

pub fn new<'p, P>(p: P, config: &GptJConfig) -> GptJModel

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

Build a new GptJModel

Arguments

• p - Variable store path for the root of the GPT-J model
• config - GptJConfig object defining the model architecture

Example

use rust_bert::gpt_j::{GptJConfig, GptJModel};
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 = GptJConfig::from_file(config_path);
let gpt_j: GptJModel = GptJModel::new(&p.root() / "gpt_j", &config);

pub fn forward_t( &self, input_ids: Option<&Tensor>, layer_past: Option<Vec<Option<LayerState>>>, attention_mask: Option<&Tensor>, token_type_ids: Option<&Tensor>, _position_ids: Option<&Tensor>, input_embeds: Option<&Tensor>, train: bool ) -> Result<GptJModelOutput, RustBertError>

Forward pass through the model

Arguments

• input_ids - Optional input tensor of shape (batch size, sequence_length). If None, pre-computed embeddings must be provided (see input_embeds)
• layer_past - Optional vector of length n_layer containing the past keys and values of each layer of shape (2, batch size, number of heads, past_sequence_length, hidden size per head). When provided, these are concatenated with the current input keys and values.
• attention_mask - Optional mask of shape (batch size, sequence_length). Masked position have value 0, non-masked value 1. If None set to 1
• input_embeds - Optional pre-computed input embeddings of shape (batch size, sequence_length, hidden_size). If None, input ids must be provided (see input_ids)
• token_type_ids - Optional token type ids used to indicate the portion of the input the token belongs to. If not None, token type embeddings will be added to the token and position embeddings.
• _position_ids - Optional position ids of shape (batch size, sequence_length). If None, will be incremented starting from the length of the past input.
• train - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference.

Returns

• GptJModelOutput containing:
  • output - Tensor of shape (batch size, sequence_length, vocab_size) representing the activations of the last hidden state
  • cache - Option<Vec<Tensor>> of length n_layer containing the past keys and values of each layer of shape (2, batch size, number of heads, past_sequence_length, hidden size per head)
  • all_hidden_states - Option<Vec<Tensor>> of length num_hidden_layers with shape (batch size, sequence_length, hidden_size)
  • all_attentions - Option<Vec<Tensor>> of length num_hidden_layers with shape (batch size, sequence_length, hidden_size)

Example

use rust_bert::gpt_j::{GptJConfig, GptJModel, LayerState};
let (batch_size, sequence_length, past_sequence_length) = (64, 128, 56);
let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device));
let mut past: Vec<Option<LayerState>> = Vec::with_capacity(config.n_layer as usize);
for _ in 0..config.n_layer as usize {
    past.push(Some(LayerState {
        prev_key: Tensor::rand(
            &[
                batch_size,
                config.n_head,
                past_sequence_length,
                config.n_embd / config.n_head,
            ],
            (Double, device),
        ),
        prev_value: Tensor::rand(
            &[
                batch_size,
                config.n_head,
                past_sequence_length,
                config.n_embd / config.n_head,
            ],
            (Double, device),
        ),
    }))
}
let attention_mask = Tensor::zeros(&[batch_size, sequence_length], (Int64, device));
let token_type_ids = Tensor::ones(&[batch_size, sequence_length], (Int64, device));

let model_output = no_grad(|| {
    gpt_j_model
        .forward_t(
            Some(&input_tensor),
            Some(past),
            Some(&attention_mask),
            Some(&token_type_ids),
            None,
            None,
            false,
        )
        .unwrap()
});