tensorflow/tfjs-coreArchived
Tfjs Core
TensorFlow.js is a JavaScript machine learning library and browser-based runtime used to build, train, and execute models. It functions as a WebGL accelerated tensor engine, providing a foundation for high-performance linear algebra operations and an automatic differentiation framework for computing gradients.
The project distinguishes itself through its ability to run machine learning directly in web environments, supporting both client-side inference and browser-based training. It enables the deployment of Python-based models by converting Keras or TensorFlow models into compatible formats and provides native support for TFLite models via flatbuffers.
The library covers a broad surface of capabilities, including model construction and transfer learning, hardware-accelerated inference, and the management of model lifecycles. It also includes utilities for data preprocessing, media decoding from camera feeds and images, and a suite of visualization tools for monitoring training progress and model architecture.
Features
- In-Browser ML Runtimes - Provides a browser-based runtime for executing pre-trained machine learning models including TFLite and Keras conversions.
- Hardware-Accelerated WebGL Execution - Functions as a WebGL accelerated tensor engine that offloads linear algebra operations to the GPU.
- Automatic Differentiation - Implements a full automatic differentiation framework for computing gradients during model training.
- Automatic Differentiation Frameworks - Provides a comprehensive engine for computing exact gradients to support model training and optimization.
- Browser-Based Deep Learning - Enables running pre-trained models or training new ones entirely within the web browser.
- Client-Side Inference - Executes machine learning predictions directly on user devices to enable real-time responses.
- Hardware-Accelerated Inference - Executes loaded models using GPU-accelerated operations for real-time inference performance.
- In-Browser Model Execution - Enables the execution of machine learning models directly in the web browser using JavaScript.
- JavaScript Machine Learning Libraries - Provides a complete library for implementing and training neural networks and ML models using JavaScript.
- JavaScript Model Training - Provides a full environment for developing and optimizing neural networks using JS and automatic differentiation.
- Eager-Execution Frameworks - Creates and optimizes machine learning models using an eager API for definition and training.
- Backend-Agnostic Engines - Provides a backend-agnostic engine that decouples tensor operations from specific hardware like WebGL or WASM.
- Tensor Initialization - Initializes multidimensional tensors of varying ranks and shapes from arrays or hardware buffers.
- Linear Algebra - Executes high-performance mathematical routines for vector and matrix operations on tensors.
- Pretrained Model Integrations - Retrieves model architecture and weight shards from web servers to perform inference or continue training.
- Eager Execution Modes - Supports an eager execution mode where operations are evaluated immediately as they are called.
- FlatBuffer Model Interpreters - Includes native support for TFLite models via FlatBuffers for efficient inference without a full framework.
- Data Preprocessing - Provides tools for cleaning, normalizing, and transforming raw data before it is used in model training.
- Model Architecture Visualizations - TensorFlow.js displays tabular information about model architecture and detailed histograms of layer parameters.
- Model Persistence - Saves and retrieves trained model artifacts and topology to local storage or static resources.
- Weight Sharding - Retrieves large model parameters in fragmented binary shards to optimize network transfer and memory.
- Training Progress Monitors - Produces real-time plots of metrics like loss and accuracy during model training via callbacks.
- Transfer Learning - Implements techniques for adapting pre-trained models to new tasks by training them on new datasets.
- Model Lifecycle Management - Manages the loading of pre-trained models from URLs and the release of associated hardware resources.
- TFLite Interpreters - Imports TFLite flatbuffers from URLs or memory buffers into an interpreter for execution.
- Web-Compatible Format Converters - Transforms models from frameworks like Keras into compatible JSON and binary formats for web execution.
- Deep Learning Frameworks - WebGL-accelerated linear algebra and ML for JavaScript.
