models

Auto-generated from tensorflow/models by Mutable.ai Auto WikiRevise

The TensorFlow Models repository is a comprehensive collection of well-maintained, tested, and optimized machine learning models implemented using TensorFlow's high-level APIs. This repository serves as a central hub for various state-of-the-art models across different domains, including natural language processing (NLP), computer vision, and recommendation systems.

The most important parts of the repository are the official and research directories. The official directory contains a diverse set of computer vision models, such as image classification, object detection, instance segmentation, and semantic segmentation, as well as NLP models, including text classification, pretraining, and question answering. The research directory provides a wide range of specialized machine learning models and research projects, covering tasks like adversarial text models, attention-based optical character recognition, audio processing, data augmentation, hierarchical reinforcement learning, and more.

The repository also includes the orbit directory, which provides a flexible and lightweight framework for writing custom training loops in TensorFlow 2. This framework simplifies the process of building and training machine learning models by offering a modular and extensible design, with abstract base classes for trainers and evaluators, as well as a Controller class for managing the outer training and evaluation loop.

The key algorithms and technologies used in this repository include:

• Transformer-based models for NLP tasks, leveraging attention mechanisms and pre-training techniques
• Convolutional neural networks (CNNs) and region-based models (e.g., Faster R-CNN, Mask R-CNN) for computer vision tasks
• Recommendation models, such as Deep Learning Recommendation Model (DLRM) and Deep Cross Network (DCN) v2, for click-through rate prediction and other recommendation tasks
• Hierarchical reinforcement learning algorithms, including the HIRO (Hierarchical Inverse Reinforcement Learning) method
• Adversarial training techniques for text classification models
• Attention-based models for optical character recognition

The key design choices of the repository include:

• Modular and extensible architecture, allowing for easy integration and customization of the various components
• Extensive use of Keras-based primitives for building and training models, promoting flexibility and interoperability
• Centralized configuration management, with separate directories for model configurations, data pipelines, and task-specific implementations
• Comprehensive documentation and unit tests to ensure the correctness and robustness of the implementations

Official Models Research Models Orbit Training Framework

References: official

The official directory is a comprehensive collection of well-maintained, tested, and optimized machine learning models implemented using TensorFlow's high-level APIs. This directory serves as a central repository for various state-of-the-art models across different domains, including natural language processing (NLP), computer vision, and recommendation systems.

References: official/vision

The …/modeling directory contains the core functionality for building various computer vision models, including image classification, object detection, instance segmentation, and video classification.

References: official/nlp

The …/modeling directory provides a comprehensive library of Keras primitives for building transformer-based natural language processing (NLP) models, including pretraining, fine-tuning, and serving functionality.

References: official/recommendation

The …/recommendation directory contains the implementation of various recommendation models and associated data preprocessing and training pipelines. The key components in this directory include:

References: official/projects/edgetpu

The …/nlp directory contains the implementation of the MobileBERT-EdgeTPU model, which is a version of the MobileBERT language model optimized for deployment on the Edge TPU hardware. The key components in this directory include:

References: official/projects/movinet, official/projects/perceiver, official/projects/videoglue

The …/movinet directory contains the implementation of the MoViNet (Mobile Video Transformer) video classification model, which is part of the TensorFlow official models repository. The directory includes the following key components:

References: official/projects/pointpillars

The …/pointpillars directory contains the implementation of the PointPillars object detection model, which is a 3D object detection model for autonomous driving applications.

References: official/projects/simclr, official/projects/const_cl

The …/simclr directory contains the core functionality and implementation of the Simple Contrastive Learning of Visual Representations (SimCLR) model. The key components in this directory include:

References: official/projects/panoptic

The …/modeling directory contains the core implementation of the Panoptic Deeplab and Panoptic Mask R-CNN models for panoptic segmentation.

References: official/projects/yolo

The …/modeling directory contains the core implementation of the YOLO (You Only Look Once) and YOLOv7 object detection models. The key components in this directory include:

References: official/projects/deepmac_maskrcnn

The …/modeling directory contains the core implementation of the DeepMAC-MaskRCNN model, a deep learning model for object detection and instance segmentation.

References: official/projects/mosaic

The …/mosaic directory contains the implementation of the MOSAIC semantic segmentation model, a deep learning model designed for efficient image segmentation, particularly on mobile devices.

References: research

The research directory contains a diverse collection of machine learning models and research projects, each with its own specialized functionality and implementation details.

References: research/object_detection

The …/object_detection directory contains the core functionality of the TensorFlow Object Detection API, which is an open-source framework for building, training, and deploying object detection models.

References: research/deeplab

The …/deeplab directory contains the core functionality for the DeepLab semantic segmentation model, including custom layers, network architectures, preprocessing utilities, and various other supporting components.

References: research/delf

The …/python directory contains the core functionality of the DELF (DEep Local Features) project, including the implementation of the DELG (Discriminative Embedding Learning with Geometry) image retrieval system.

References: research/seq_flow_lite

The …/seq_flow_lite directory contains the implementation of various neural network models and custom TensorFlow operations that are part of the Sequence Flow Lite (SeqFlow Lite) framework. The key components in this directory include:

References: research/lstm_object_detection

The …/lstm_object_detection directory contains the core functionality for an LSTM-based object detection model. The main components include:

References: research/adversarial_text

The …/adversarial_text directory contains code for training text classification models using adversarial training methods. The main functionality includes:

References: research/efficient-hrl

The …/ directory contains the implementation of a Hierarchical Reinforcement Learning (HRL) framework, including the HIRO (Hierarchical Inverse Reinforcement Learning) algorithm and its variants.

References: research/cvt_text

The …/ directory contains the core functionality for the Cross-View Training (CVT) text model. The key components in this directory are:

References: research/attention_ocr

The …/python directory contains the core implementation of the Attention OCR (Optical Character Recognition) model. The key components are:

References: research/audioset

The …/audioset directory contains implementations of two pre-trained deep learning models for audio processing: VGGish and YAMNet. These models are designed to work with the AudioSet dataset, a large-scale dataset of over 2 million human-labeled 10-second YouTube video soundtracks with labels from an ontology of more than 600 audio event classes.

References: research/lfads

The …/lfads directory contains the implementation of the LFADS (Latent Factor Analysis via Dynamical Systems) model, a sequential variational autoencoder designed for analyzing time-series data, particularly in neuroscience applications.

References: research/deep_speech

The …/ directory contains the implementation of the DeepSpeech2 model, a deep learning-based automatic speech recognition (ASR) system. The key components in this directory are:

References: research/autoaugment

The …/autoaugment directory contains a comprehensive implementation of various deep learning models, data augmentation techniques, and training utilities for image classification tasks on the CIFAR-10 and CIFAR-100 datasets, including the AutoAugment data augmentation technique.

References: orbit

The Orbit Training Framework provides a flexible and lightweight library for writing custom training loops in TensorFlow 2. The key components of this framework include:

References: orbit/actions

The …/actions directory provides a set of abstractions and utilities for defining and applying "actions" within the Orbit framework. These actions can be used to perform various tasks during the training or evaluation of a machine learning model, such as reporting metrics, exporting models, or saving checkpoints.

References: orbit/controller.py

The Controller class in the …/controller.py file is responsible for managing the outer training and evaluation loop for machine learning models. It coordinates the execution of actions, such as saving checkpoints, running evaluations, and writing summaries, in addition to handling checkpoint restoration and saving.

References: orbit/runner.py

The models/orbit/runner.py file provides two abstract base classes, AbstractTrainer and AbstractEvaluator, that define the APIs for training and evaluating machine learning models.

References: orbit/standard_runner.py

The standard_runner.py file in the orbit directory provides two main classes, StandardTrainer and StandardEvaluator, which extend the AbstractTrainer and AbstractEvaluator APIs, respectively. These classes add additional structure and functionality on top of the bare APIs, making it easier to implement common training and evaluation workflows.

References: orbit/utils

The …/ directory provides a rich set of utility functions and classes that are essential for building and training machine learning models using the Orbit framework. The key functionalities include:

References: community

The community directory contains a curated collection of pre-trained machine learning models and implementations powered by TensorFlow 2. The models are organized into three main categories: Computer Vision, Natural Language Processing, and Recommendation Systems.

References: tensorflow_models

The TensorFlow Models repository serves as a central hub for various machine learning models and utilities within the TensorFlow ecosystem. It is divided into two main sub-directories: …/nlp and …/vision, each focusing on natural language processing (NLP) and computer vision tasks, respectively.

References: tensorflow_models/nlp

The …/nlp directory provides a comprehensive collection of natural language processing (NLP) related components, models, and functionalities within the TensorFlow Models repository.

References: tensorflow_models/vision

The …/vision directory contains the core components and functionality for computer vision tasks within the TensorFlow Models repository.