open-edge-platform/anomalib

Features

• Benchmarking Suites - Provides a benchmarking suite that evaluates multiple anomaly detection models across standard datasets.
• Anomaly Detection Metrics - Computes a suite of standard metrics (AUROC, AUPR, F1, etc.) to measure anomaly detection performance.
• Backbone Integrations - Uses pretrained neural network backbones to compute deep feature representations from input data.
• Benchmarking Suites - Automates hyperparameter and model grid searches across accelerators and datasets, collecting metrics in parallel across worker processes.
• Model Performance Benchmarks - Runs grid searches over models, accelerators, and datasets to collect performance metrics in parallel or serial.
• Data Preprocessing - Anomalib applies transformations to raw images before passing them to the anomaly detection model.
• Experiment Logging - Records training metrics, hyperparameters, and inference images to multiple experiment tracking backends.
• Experiment Tracking - Logs training metrics, hyperparameters, and model artifacts to Weights & Biases, Comet, or TensorBoard.
• Experiment Visualization - Logs training parameters and outputs to loggers and generates visual plots for experiment analysis.
• Feature Extraction - Extracts intermediate feature maps from images using pretrained or custom backbone models with layer selection.
• Anomaly Detection Pipeline Assemblers - Assembles custom preprocessing, feature extraction, and classification steps into reusable anomaly detection workflows.
• Model Performance Benchmarking - Automatically collects accuracy, F1-score, and timing metrics when training and evaluating anomaly detection models.
• Anomaly Detection Training Workflows - Trains deep neural networks on image datasets to detect visual anomalies using configurable architectures and pipelines.
• Training and Evaluation Pipelines - Trains deep learning models on normal images and evaluates performance with specialized metrics.
• Model Benchmarking Tools - Computes metrics across multiple models via grid-search on defined configurations.
• Model Performance Evaluators - Computes performance metrics for anomaly detection models with stage-specific evaluation during training and inference.
• Anomaly Segmentation - Produces pixel-level anomaly maps that localize defects with automatic thresholding.
• Quality Defect Detection - Detects defects and irregularities in manufacturing images using deep learning models for quality control.
• Platforms - Ships a no-code web app and CLI for training and deploying anomaly detection models on manufacturing visual data.
• Anomaly Detection Algorithms - Provides a comprehensive model zoo with ready-to-use implementations of anomaly detection algorithms for multiple settings.
• Anomaly Detection Dataset Loaders - Loads and prepares visual anomaly detection data from standard benchmarks and custom folders.
• Benchmarking Tools - Evaluates state-of-the-art anomaly detection methods on public and private datasets using built-in benchmarking tools.
• Custom Model Frameworks - Provides core building blocks and utilities for assembling custom anomaly detection models from scratch.
• Custom Dataset Loading - Anomalib loads image datasets from custom folders and standard formats, structuring them for model training and evaluation.
• Inference Data Pipelines - Applies trained models to new data to generate anomaly scores and segmentation maps.
• Benchmark Dataset Loaders - Provides ready-to-use PyTorch Dataset implementations for standard anomaly detection benchmarks.
• Vision Dataset Loading - Anomalib loads standard and custom image datasets through dedicated datamodules, handling dataset-specific formats for anomaly detection.
• CLI Automation Tools - Provides a CLI to run training, evaluation, and inference without writing scripts.
• ML Experiment Loggers - Logs training metrics and configurations to TensorBoard, Weights & Biases, or Comet.ml through Lightning loggers.
• Modular Frameworks - Provides a modular framework for building custom anomaly detection pipelines with interchangeable components.
• Model Inference Deployment - Exports models to OpenVINO IR and provides inference tools for production anomaly detection.
• Visual Anomaly Detection Toolkits - Provides a library of deep learning models and tools for detecting visual defects in images and video using PyTorch.
• Model Export Pipelines - Exports trained anomaly detection models to ONNX, OpenVINO, and PyTorch for edge deployment.
• Anomaly Score Post-Processors - Provides configurable thresholding to convert raw anomaly scores into interpretable heatmaps and binary masks.
• Anomaly Detection Algorithm Plugins - Implements anomaly detection algorithms as pluggable LightningModules with a standardized interface.
• Model Evaluation - Evaluates trained anomaly detection models on test datasets and reports computed performance metrics.
• Score Post-Processors - Converts raw anomaly scores into binary labels and heatmaps using F1-based or adaptive thresholding.
• Anomaly Detection Data Module Extensions - Provides base classes and data formats for integrating new anomaly detection datasets.
• Custom Post-Processing Pipelines - Anomalib allows defining custom post-processing logic by subclassing the base post-processor and implementing a forward method.
• Dimensionality Reduction - Compresses high-dimensional feature vectors into lower-dimensional space using linear decomposition or random projection.
• Decision Threshold Calibration - Determines the classification threshold that maximizes the F1 score, removing the need for manual tuning.
• Feature Extractor Configurations - Anomalib selects a neural network backbone from a model zoo or a custom PyTorch module to extract features for anomaly detection.
• Feature-Based Clusterers - Groups data points into clusters based on learned feature representations using GMM or K-Means algorithms.
• Anomaly Visualization Tools - Generates visual representations of model predictions to highlight detected anomalies.
• Model Checkpointing - Saves model weights to disk during validation or training, supporting zero-shot and few-shot scenarios.
• Anomaly-Specific Metrics - Measures model performance using anomaly-specific metrics like PRO, sPRO, and PIMO in addition to standard metrics.
• Multi-Metric Orchestrators - Coordinates computation of multiple evaluation metrics in a single pass for comprehensive performance reporting.
• Customizable Evaluation Pipelines - Uses an Evaluator class to automatically update, compute, and log multiple metrics when testing a model.
• Hyperparameter Optimization - Automates hyperparameter search for anomaly detection models using Weights & Biases or Comet.ml sweeps.
• Multi-Backend Inference Orchestration - Deploys anomaly detection models using Torch, Lightning, Gradio, or OpenVINO backends.
• Anomaly Detection Applications - Anomalib uses a pre-trained vision-language model and natural language prompts to compare images against reference normal images for detection.
• PyTorch Model Export - Converts trained models to PyTorch, ONNX, or OpenVINO format with optional compression and quantization.
• Synthetic Anomaly Generators - Anomalib creates synthetic anomaly masks and augmented images using Perlin noise patterns or real anomaly source images for training.
• Zero and Few-Shot Learning - Detects anomalies without training data by leveraging pretrained vision-language models or reference images.
• Video Anomaly Segmentations - Anomalib performs few-shot unsupervised segmentation of anomalous regions in video by factorizing spatio-temporal features with low-rank constraints.
• Video Anomaly Detection - Anomalib identifies anomalous events in video streams using attribute-based representations that provide accurate and interpretable detections.
• Attribute-Based Video Anomaly Detections - Anomalib extracts regions of interest from video frames and computes attribute-based representations to identify anomalous events.
• Anomaly Detection Data Structures - Defines typed dataclasses for image, video, and depth data used in anomaly detection.
• Anomaly Detection Data Containers - Defines typed dataclasses for anomaly detection samples with automatic field validation.
• Fast Anomaly Detectors - Anomalib uses a student-teacher architecture with contrastive learning to identify anomalous regions without requiring labeled anomalies.
• Few-Shot Anomaly Detectors - Supports few-shot and zero-shot anomaly detection by comparing image regions to reference images using CLIP embeddings.
• Zero-Shot Anomaly Detectors - Anomalib uses a pre-trained vision foundation model to identify image-level and pixel-level anomalies without requiring training data.
• Normalizing Flow Anomaly Detectors - Anomalib models normal data distributions with conditional normalizing flows to detect anomalies in images in real time.
• Normalizing Flow Detectors - Anomalib transforms input data through invertible flow blocks combining affine coupling, permutation, and global affine scaling to learn complex distributions.
• U-Normalizing Flow Detectors - Anomalib estimates a probability distribution over normal image features using a U-shaped normalizing flow and flags regions with low likelihood as anomalies.
• Patch Distribution Detectors - Anomalib fits a separate Gaussian distribution to patch embeddings at each location of a pretrained CNN to detect and localize anomalies.
• Patch k-Nearest Neighbors Detectors - Anomalib compares patch embeddings of test images to a memory bank of normal patches using k-nearest neighbor search to identify anomalous regions.
• Quantized Feature Anomaly Detectors - Anomalib learns a quantized feature codebook and generates synthetic defects internally to train decoders that identify anomalies without external defect datasets.
• Reconstruction-Based Anomaly Detectors - Anomalib compares an input image with its reconstruction to identify anomalous regions using a discriminatively trained embedding network.
• Reverse Distillation Anomaly Detectors - Anomalib trains a decoder to reconstruct multi-scale features from a compressed bottleneck and scores anomalies by the reconstruction error on normal images only.
• RGB-D Anomaly Detectors - Anomalib learns mappings between 2D appearance and 3D geometry to identify anomalies in visual and depth data jointly.
• Unsupervised Anomaly Detectors - Anomalib synthesizes global and local anomalies using gradient ascent on normal features to spot defects without labeled training samples.
• Sample Containers - Defines dataclasses holding input images, ground truth, masks, and predictions for each sample.
• Lightning DataModule Loaders - Implements Lightning DataModules for automated splitting, batching, and distributed data loading.
• Multi-Modal Loaders - Anomalib loads image, video, or depth datasets for anomaly detection using pre-built dataset classes.
• Stage-Specific Batch Transforms - Applies separate transform sets to data batches for each pipeline stage during training and inference.
• CLI Configuration Files - Configures experiments via command-line arguments and YAML/JSON configuration files.
• Pipeline Execution Interfaces - Orchestrates the complete anomaly detection workflow through a command-line interface.
• Anomaly Detection Pipelines - Executes multi-stage pipelines with sequential and parallel job stages for anomaly detection.
• Model Export Formats - Converts trained models to ONNX, OpenVINO IR, or PyTorch for production deployment.
• No-Code Training Interfaces - Provides a no-code web application for training and deploying anomaly detection models.
• Real-Time Model Inference on Frames - Deploys trained models for low-latency inference on edge devices using ONNX and OpenVINO.
• Edge Deployments - Exports models to OpenVINO and performs real-time anomaly detection on streaming data at the edge.
• Test Result Visualizers - Generates visual overlays of anomaly maps and segmentation masks during test and predict steps.

स्टार हिस्ट्री