rust_bert/models/t5/
mod.rs

1//! # T5 (Text-To-Text Transfer Transformer)
2//!
3//! Implementation of the T5 language model ([Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li, Liu, 2019).
4//! The base model is implemented in the `t5_model::T5Model` struct. This model includes a language model head: `t5_model::T5ForConditionalGeneration`
5//! implementing the common `generation_utils::LanguageGenerator` trait shared between the models used for generation (see `pipelines` for more information).
6//!
7//! # Model set-up and pre-trained weights loading
8//!
9//! A full working example (summarization) is provided in `examples/summarization_t5`, run with `cargo run --example summarization_t5`.
10//! All models expect the following resources:
11//! - Configuration file expected to have a structure following the [Transformers library](https://github.com/huggingface/transformers)
12//! - Model weights are expected to have a structure and parameter names following the [Transformers library](https://github.com/huggingface/transformers). A conversion using the Python utility scripts is required to convert the `.bin` weights to the `.ot` format.
13//! - `T5Tokenizer` using a `spiece.model` sentence piece model
14//!
15//! Pretrained models for a number of language pairs are available and can be downloaded using RemoteResources.
16//!
17//! ```no_run
18//! # fn main() -> anyhow::Result<()> {
19//! #
20//! use tch::{nn, Device};
21//! # use std::path::PathBuf;
22//! use rust_bert::resources::{LocalResource, ResourceProvider};
23//! use rust_bert::t5::{T5Config, T5ForConditionalGeneration};
24//! use rust_bert::Config;
25//! use rust_tokenizers::tokenizer::T5Tokenizer;
26//!
27//! let config_resource = LocalResource {
28//!     local_path: PathBuf::from("path/to/config.json"),
29//! };
30//! let sentence_piece_resource = LocalResource {
31//!     local_path: PathBuf::from("path/to/spiece.model"),
32//! };
33//! let weights_resource = LocalResource {
34//!     local_path: PathBuf::from("path/to/model.ot"),
35//! };
36//! let config_path = config_resource.get_local_path()?;
37//! let spiece_path = sentence_piece_resource.get_local_path()?;
38//! let weights_path = weights_resource.get_local_path()?;
39//!
40//! let device = Device::cuda_if_available();
41//! let mut vs = nn::VarStore::new(device);
42//! let tokenizer = T5Tokenizer::from_file(spiece_path.to_str().unwrap(), true);
43//! let config = T5Config::from_file(config_path);
44//! let t5_model = T5ForConditionalGeneration::new(&vs.root(), &config);
45//! vs.load(weights_path)?;
46//!
47//! # Ok(())
48//! # }
49//! ```
50
51mod attention;
52mod encoder;
53mod layer_norm;
54mod t5_model;
55
56pub use attention::LayerState;
57pub(crate) use attention::{get_relative_position_bucket, T5Attention, T5LayerCrossAttention};
58pub(crate) use encoder::{T5Block, T5BlockOutput, T5LayerFF, T5StackOutput};
59pub(crate) use layer_norm::T5LayerNorm;
60pub(crate) use t5_model::{FeedForwardProj, TaskSpecificParams};
61pub use t5_model::{
62    T5Config, T5ConfigResources, T5ForConditionalGeneration, T5ForSentenceEmbeddings, T5Generator,
63    T5Model, T5ModelOutput, T5ModelResources, T5Prefix, T5SourceLanguages, T5TargetLanguages,
64    T5VocabResources,
65};
rust_bert/models/t5/mod.rs

rust_bert/models/t5/
mod.rs
