https://github.com/nicksenger/candle-token-classification

Token Classification for BERT-like models via Candle
https://github.com/nicksenger/candle-token-classification

candle deep-learning neural-networks nlp

Last synced: 5 months ago
JSON representation

Token Classification for BERT-like models via Candle

• Host: GitHub
• URL: https://github.com/nicksenger/candle-token-classification
• Owner: nicksenger
• Created: 2023-12-05T00:10:31.000Z (almost 2 years ago)
• Default Branch: main
• Last Pushed: 2024-08-16T06:17:24.000Z (about 1 year ago)
• Last Synced: 2025-05-12T12:13:46.488Z (5 months ago)
• Topics: candle, deep-learning, neural-networks, nlp
• Language: Rust
• Homepage:
• Size: 32.2 KB
• Stars: 2
• Watchers: 3
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Candle Token Classification

This project provides support for using pre-trained token classification models of the BERT lineage (BERT, RoBERTa, DeBERTa, Electra, etc) via [Candle](https://github.com/huggingface/candle)

## Usage

```rust

use candle_token_classification::BertLikeTokenClassificationHead; // Import the token classifier trait from this library

use candle_token_classification::BertTokenClassificationHead; // Import the concrete classifier (BERT & ELECTRA are provided)

fn main() -> Result<(), Box> {

    // ... retrieve Config, VarBuilder & Tokenizer from filesystem, HF hub, etc

    let classifier = BertTokenClassificationHead::load(vb, &config)?;

    // retrieve an ordered list of labels for the chosen classifier

    let labels = config.id2label.values().cloned().collect();

    let output = classifier.classify( // classify some text (or use `classifier.forward` to get the output tensor)

        "This is the text we'd like to classify.",

        &labels,

        &tokenizer,

        &classifier.device

    );

    println!("{:?}", output) // view human-readable output or send for downstream processing, etc

    Ok(())

}

```

## Contributing

This repo was made specifically for my use-case of UPOS tagging with pre-trained models. As such it is likely missing things required to support work outside that scope. If you need other functionality, contributions are welcome :)