@rokikon/pipelines

Static, type-safe wrappers for machine learning pipelines with lazy loading.

This TypeScript library provides a simple and efficient way to interact with machine learning models, inspired by @xenova/transformers pipelines.

Benefits:

• Static Typing: Benefit from strong typing for enhanced code maintainability and error detection.
• Lazy Loading: Optimize performance by loading models only when needed.
• Simplified Interface: Enjoy an easy-to-use API for common machine learning tasks.

Installation

npm install @rokikon/pipelines

Usage

This library provides static classes for each pipeline task.

Example:

import { TextClassificationPipeline } from '@rokikon/pipelines';

async function main() {
  const results = await TextClassificationPipeline.run("This is a test sentence.");
  console.log(results);
}

main();

Customize pipeline behavior with different options:

import { TextGenerationPipeline } from '@rokikon/pipelines';

async function main() {
  const results = await TextGenerationPipeline.run("This is a test sentence", {
    model: "onnx-community/Llama-3.2-1B-Instruct-q4f16",
    options: {
      dtype: "q4f16",
      device: "webgpu",
    },
  });
  console.log(results);
}

main();

Warm up the model:

You can use the warmup method to pre-load a model and avoid latency on the first run. This is especially useful for large models that take time to initialize. Don't forget to run same model later.

await TextGenerationPipeline.warmup({
  model: "onnx-community/Llama-3.2-1B-Instruct-q4f16",
  options: {
    dtype: "q4f16",
    device: "webgpu",
    max_new_tokens: 2048,
  },
});

Available Pipelines

Audio

• AudioClassificationPipeline - Classify audio into predefined categories.
• AutomaticSpeechRecognitionPipeline - Transcribe spoken language into text (alternative to SpeechRecognitionPipeline).
• ZeroShotAudioClassificationPipeline - Classify audio into arbitrary categories without prior training.

Computer Vision

• DepthEstimationPipeline - Predict the depth of an image.
• ImageCaptioningPipeline - Generate descriptive captions for images.
• ImageFeatureExtractionPipeline - Extract meaningful features from images.
• ImageSegmentationPipeline - Partition an image into multiple segments.
• ImageToImagePipeline - Transform images from one style to another.
• ImageToTextPipeline - Generate textual descriptions from images (similar to Image Captioning but may have different underlying implementations).
• ObjectDetectionPipeline - Locate and identify objects within an image.
• ZeroShotImageClassificationPipeline - Classify images into arbitrary categories without prior training.
• ZeroShotObjectDetectionPipeline - Detect objects in an image without specific training for those objects.

Natural Language Processing

• DocumentQuestionAnsweringPipeline - Extract answers to questions from a given document.
• FeatureExtractionPipeline - Extract meaningful features from text.
• FillMaskPipeline - Predict missing words in a text.
• QuestionAnsweringPipeline - Answer questions based on a given context.
• SummarizationPipeline - Condense lengthy text into concise summaries.
• TextClassificationPipeline - Classify text into predefined categories.
• TextGenerationPipeline - Generate new text, such as creative writing or code.
• TextToAudioPipeline - Convert text into spoken audio.
• TextToTextGenerationPipeline - Transform or generate new text from given text, including tasks like translation and paraphrasing.
• TokenClassificationPipeline - Classify individual tokens (words or sub-words) in a text, useful for tasks like Named Entity Recognition (NER).
• TranslationPipeline - Translate text from one language to another.
• ZeroShotClassifierPipeline - Classify text into arbitrary categories without any prior training.

Contributing

Contributions are welcome! Feel free to open issues and pull requests. https://github.com/rokikon/pipelines-ts

License

This project is licensed under the MIT License.