Struct OpenAiGptModel

pub struct OpenAiGptModel { /* private fields */ }

GPT Base model

Base architecture for GPT model. Usually complemented with a task-specific head, such as a language model head. As opposed to GPT2, GPT does not give the possibility to re-use past activations as an input. It is made of the following blocks:

• tokens_embed: token embeddings
• positions_embed: position embeddings
• h: Encoder (transformer) made of a vector of layers. Each layer is made of a multi-head attention layer, layer-normalization layers and a MLP made of linear layers.
• 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 OpenAiGptModel

pub fn new<'p, P>(p: P, config: &Gpt2Config) -> OpenAiGptModel

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

Build a new OpenAiGptModel

Arguments

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

Example

use rust_bert::openai_gpt::{OpenAiGptConfig, OpenAiGptModel};
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 = OpenAiGptConfig::from_file(config_path);
let gpt2: OpenAiGptModel = OpenAiGptModel::new(&p.root() / "gpt", &config);

pub fn forward_t( &self, input_ids: Option<&Tensor>, attention_mask: Option<&Tensor>, token_type_ids: Option<&Tensor>, position_ids: Option<&Tensor>, input_embeds: Option<&Tensor>, train: bool, ) -> Result<OpenAiGptModelOutput, 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)
• 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

• OpenAiGptModelOutput containing:
  • output - Tensor of shape (batch size, sequence_length, hidden_size) representing the activations of the last hidden state
  • 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::gpt2::Gpt2Config;
use rust_bert::openai_gpt::OpenAiGptModel;
let (batch_size, sequence_length, past_sequence_length) = (64, 128, 56);
let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, 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 position_ids = Tensor::arange(sequence_length, (Int64, device))
    .expand(&[batch_size, sequence_length], true);

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