DOCSEARCHbyAlgolia
Submit Docs

TensorFlow Model Garden

repository·master·Indexed 12 days ago

https://github.com/tensorflow/models

A collection of state-of-the-art model implementations for TensorFlow 2, including official, research, and community-curated models. It covers domains such as Computer Vision (ResNet, MobileNet, Mask R-CNN), Natural Language Processing (BERT, T5, Transformer-XL), and Recommendation Systems (Wide & Deep, DLRM). The repository provides the `tensorflow_models` pip package, including `orbit`, and specialized NLP modeling layers for attention mechanisms, transformer architectures, and tokenization.

Tokens
149K
Snippets
382
Records
582
Agent score
99%

What's inside TensorFlow Model Garden

  1. Overview of the MARCO Crystallization Classification Model

    master

    The MARCO model is a pretrained Inception-v3 variant designed to automate the evaluation of crystallization experiments. It takes images of crystallization experiments as input and classifies them into one of four categories: crystals, precipitate, clear, or others.

    Important Compatibility Note: This model requires TensorFlow 1.x. TensorFlow 2.x is not supported.

  2. Overview of TF Model Garden Ranking Models

    master

    This repository provides implementations of DLRM and DCN v2 ranking models, designed for tasks like CTR (Click-Through Rate) prediction.

    Model Characteristics

    • Inputs: Numerical and categorical features.
    • Output: A scalar value (e.g., click probability).
    • Hardware Support: Optimized for GPU, TPU, and CPU.
      • CPUs: Best for large sparse embedding lookups.
      • GPUs: Best for fast compute.
      • TPUs: Designed for both; uses TPUEmbedding layer for categorical features to support massive embedding tables (up to 22.4 TB on TPU Pod v3-2048).

    Implementation Details

    • The core model code is located in the TensorFlow Recommenders library.
    • The input pipeline, configuration, and training loops are contained within this repository.

  3. Overview of TensorFlow Research Models

    master
    The research/ directory in the TensorFlow Models repository contains code implementations and pre-trained models derived from published research papers. These models are maintained by their respective original authors. The repository serves as a collection of state-of-the-art implementations across various domains including Computer Vision, NLP, Audio, and Reinforcement Learning.

  4. Overview of TF-NLP Model Garden

    master

    The TF-NLP Model Garden is a Natural Language Processing (NLP) modeling toolkit built for TensorFlow 2 (TF2). It is designed for researchers and developers to reproduce state-of-the-art (SOTA) model results and to train custom models for new research experiments.

    Key features include:

    • Modular Building Blocks: Reusable and modularized modeling components.
    • Reproducibility: Support for reproducing state-of-the-art results.
    • Extensibility: Easy to customize and extend for specific research needs.
    • End-to-End Training: Support for complete training pipelines.
    • Distributed Training: Capable of distributed training on both GPUs and TPUs.

  5. Overview of PCL RL Algorithms

    master

    The pcl_rl research repository provides implementations of several Reinforcement Learning (RL) algorithms used in various research papers regarding policy gradients, value-based learning, and trust region methods.

    Available Algorithms:

    • Actor Critic
    • TRPO
    • PCL
    • Unified PCL
    • Trust-PCL
    • PCL + Constraint Trust Region (un-published)
    • REINFORCE
    • UREX

  7. Overview of AudioSet Models

    master

    This repository contains models and supporting code for AudioSet, a large-scale dataset consisting of over 2 million human-labeled 10-second YouTube video soundtracks. The models are designed for audio event detection across an ontology of more than 600 audio event classes.

    Important Compatibility Note: This project requires TensorFlow 1.x. It is not compatible with TensorFlow 2.x.

  8. Overview of DETR (End-to-End Object Detection with Transformers)

    master
    DETR is a TensorFlow 2 implementation of the End-to-End Object Detection with Transformers architecture. It provides a framework for object detection that avoids the need for many hand-designed components like non-maximum suppression or anchor generation by using a transformer architecture to predict sets of objects directly.

  9. Overview of CircularNet for Waste Identification

    master

    CircularNet is a set of instance segmentation models designed to identify recyclables on conveyor belts. It is built using RF-DETR, a vision transformer model that performs both object detection and instance segmentation (assigning instance-level labels to every pixel in an input image).

    Model Capabilities

    The models categorize waste into two primary dimensions:

    • Material Type: Identifies the material (e.g., metal, paper). For plastics, it identifies specific resin types (e.g., HDPE, PET, LDPE).
    • Material Form: Categorizes the object's form factor (e.g., cup, bottle, bag).

    Example inference label: Plastics-PET_Bottle

  10. Overview of MaxViT: Multi-Axis Vision Transformer

    master

    MaxViT is a family of hybrid (CNN + ViT) vision backbone models designed for high parameter and FLOPs efficiency compared to standard ConvNets and Transformers. It uses a multi-axis approach that scales well to large datasets (like ImageNet-21K) and large image sizes (useful for object detection and segmentation) due to the linear complexity of its grid attention mechanism.

    Meta-architecture: The backbone is composed of homogeneously stacked blocks. Each MaxViT block contains:

    1. MBConv: Mobile Inverted Bottleneck Convolution.
    2. Block Attention: Window-based local attention.
    3. Grid Attention: Dilated global attention.

  11. Overview of the TensorFlow Object Detection API

    master

    The TensorFlow Object Detection API is an open-source framework built on top of TensorFlow designed to make it easy to construct, train, and deploy object detection models. It allows for localizing and identifying multiple objects within a single image.

    Important Note on Maintenance: The API is no longer being actively maintained for compatibility with new versions of external dependencies (e.g., pip, apt-get). While original installation instructions are preserved for legacy use, users seeking actively maintained detection or segmentation codebases are encouraged to use:

  12. Overview of Quantization Aware Training (QAT) for Vision

    master

    This project provides implementation examples for Quantization Aware Training (QAT) specifically for computer vision models. QAT is used to minimize quality loss during integer quantization compared to Post-Training Quantization (PTQ), making it ideal for models with strict latency and quality requirements.

    Supported tasks include:

    • Image Classification: ResNet, MobileNetV2, MobileNetV3 large, and Multi-hardware MobileNet.
    • Object Detection: RetinaNet (MobileNet v2).
    • Semantic Segmentation: DeepLabV3/V3+ (MobileNet v2).