autogluon/autogluon

Features

• End-to-End Training Pipelines - Automates the end-to-end machine learning pipeline to build high-accuracy models for diverse data types.
• Automated ML (AutoML) - Provides an end-to-end automated pipeline for selecting, tuning, and deploying high-accuracy models across multiple data types.
• Time Series Forecasting - Provides a complete framework for predicting future values and probabilistic trends for multiple univariate sequences.
• Automated Machine Learning - Automates the end-to-end pipeline from data preprocessing to high-accuracy model training and validation.
• Pretrained Model Integrations - Provides utilities to execute pretrained foundation models for immediate zero-shot predictions without custom training.
• Ensemble Methods - Implements multi-layer stacking to combine predictions from base models as features for higher-level meta-models.
• Feature Extraction - Converts raw image or text data into high-dimensional feature vectors using zero-shot prediction for downstream tasks.
• Regression Models - Predicts continuous numerical values from text, images, and tabular inputs and evaluates them using error metrics.
• Computer Vision Pipelines - Provides automated workflows for object detection, semantic segmentation, and image classification.
• Data Preprocessing - Internally handles preprocessing for text and image data to prepare raw inputs for model training.
• Model Fine-Tuning - Supports adapting pre-trained models to specific downstream tasks and datasets to improve domain-specific accuracy.
• Model Architecture Selection - Automatically specifies the most appropriate pretrained backbones or MLP types for various data modalities.
• Model Ensembling - Combines user-defined models with automated models to build multi-layer stack ensembles for improved performance.
• Finetuning Workflows - Adapts large pretrained foundation models to specific tasks using parameter-efficient finetuning.
• Model Performance Selection - Automatically selects the best local, global, and ensemble models based on quality presets or time limits.
• Multimodal Embedding Models - Converts raw multimodal samples into shared numerical embedding vectors for downstream predictive tasks.
• Multimodal Integration Libraries - Provides a library for training and deploying models that integrate tabular, image, text, and time series data.
• Multimodal Machine Learning - Predicts discrete labels or class probabilities for a set of combined multimodal input features.
• Multimodal Models - Integrates diverse data types by embedding images and text into a shared vector space for joint predictive modeling.
• Multimodal Training - Implements training workflows for neural networks that embed and fuse different data modalities into a single model.
• Parameter Efficient Fine-Tuning - Implements memory-efficient adaptation techniques such as LoRA, IA3, and BitFit by updating only a small subset of model parameters.
• Cross-Modal Similarity Scoring - Determines semantic similarity between different data modalities, including text-to-text, image-to-image, and image-to-text.
• Tabular Feature Engineering - Converts raw data into model-ready formats by encoding categories and extracting datetime components.
• Text Model Training - Trains high-quality models for text data using various backbones and supports cross-lingual transfer.
• Automated Model Selection - Fits, tunes, and selects the best forecasting models for a panel of time series based on a prediction horizon.
• Probabilistic Forecasting - Produces multi-step ahead probabilistic predictions providing value ranges and uncertainty intervals for univariate time series.
• Zero-Shot Forecasting - Uses pretrained large-scale time series models to generate forecasts on new data without requiring training.
• Vector Embeddings - Generates high-dimensional vector representations of text and images to compute semantic similarity.
• Zero-Shot Inference - Executes predictions across various modalities using pretrained foundation models without requiring task-specific training data.
• Foundation Model Adaptation - Provides methods to adapt large pre-trained foundation models using parameter-efficient techniques like IA^3 and BitFit.
• Tabular Predictive Models - Implements automated predictive modeling for structured tabular data to forecast values or categories.
• Batch Inference Engines - Uses a dependency graph to coordinate predictions across multiple models, eliminating redundant computations and increasing throughput.
• Batch Prediction Processing - Produces prediction probabilities from multiple models simultaneously using a dependency graph.
• Class Probability Estimation - Calculates the likelihood of each possible class for a given input for granular analysis.
• Image Augmentation - Applies geometric and color transformations to image data to improve model generalization.
• Image Classification Models - Develops image classification and regression models, including support for zero-shot classification using pretrained models.
• Object Detection - Identifies and localizes objects within images using bounding boxes via both zero-shot and trained models.
• Automated Training - Automates the end-to-end training of high-quality object detection models using COCO-formatted data.
• Segmentation Model Training - Trains models to perform pixel-level semantic image segmentation for detailed visual understanding.
• Continuous Model Training - Updates models with new data over time to prevent performance decay and adapt to changing patterns.
• Custom Model Integrations - Wraps external time series models into a compatible interface for use within the automated forecasting pipeline.
• Forecasting Input Preparation - Formats dataset splits into historical data and future covariates to prepare for forecasting.
• Decision Threshold Calibration - Implements techniques for converting continuous probability estimates into discrete class labels using optimized thresholds for binary classification.
• Detection Accuracy Metrics - Calculates object detection accuracy using standard COCO and VOC mean Average Precision (mAP) metrics.
• Distributed Training - Executes training across multiple GPUs or nodes using data parallel and distributed strategies.
• Bagging Ensembles - Fits base models across multiple data splits using bagging to improve model stability and reduce variance.
• Stacking Ensembles - Implements multi-layer stacking by using base model predictions as features for subsequent model layers.
• Weighted Ensembles - Combines multiple models using optimized weights via ensemble selection for final predictions.
• Feature Contribution Analysis - Identifies key model drivers by quantifying feature impact through permutation shuffling.
• Feature Engineering - Implements manual data transformations and custom feature generators that execute during both training and inference.
• Custom Feature Pipelines - Defines custom sequences of feature generators to control how specific data types are handled and transformed.
• Text Embedding Extraction - Converts groups of sentences into vector representations for similarity and retrieval tasks.
• Feature Importance Attribution - Quantifies the impact of input features on predictive accuracy using value shuffling.
• Few-Shot Learning Frameworks - Provides a framework for training classification models on multimodal data using minimal examples.
• Few-Shot Learning Mechanisms - Combines foundation models with support vector machines to implement few-shot learning.
• Tabular Foundation Model Application - Leverages pre-trained foundation models to improve predictive performance on small tabular datasets.
• Image-Text Ranking - Measures and ranks visual-semantic similarity between images and text to identify the most relevant matches.
• Knowledge Distillation - Trains small student models to mimic complex teacher ensembles for more efficient production deployment.
• Covariate Integration - Improves forecast accuracy by integrating static features and known future covariates into the models.
• Image Encoder Embedding Extractions - Converts images into feature vectors to enable the calculation of semantic similarity scores.
• GPU Training Accelerators - Accelerates model training by utilizing parallelization strategies across GPUs.
• Model Inference Accelerators - Integrates TensorRT as a model inference accelerator to increase prediction speed and reduce deployment latency.
• ONNX Model Exporters - Provides the ability to convert machine learning models into the standardized ONNX format to increase prediction throughput.
• Model Comparison Interfaces - Provides summary tables comparing trained models across validation scores, training times, and inference speeds.
• Model Evaluation Metrics - Allows the definition of specific evaluation criteria to guide model selection and optimization.
• Inference Optimization - Improves prediction speed by persisting models in memory or refitting ensembles into single models.
• Text Classification - Trains models to perform binary classification, multiclass classification, or regression on datasets with text fields.
• Model Interpretability - Constructs interpretable glass-box generalized additive models with automatic interaction detection.
• Foundation Model Ensembles - Combines multiple foundation models in a single predictor through stacking and ensembling.
• Deployment Optimizations - Refines predictor artifacts to improve inference speed and resource efficiency for production environments.
• Model Prediction Evaluation - Compares predicted probabilities and labels against ground truth data to generate comprehensive performance reports.
• Image Labeling Engines - Generates class predictions or probability distributions for images provided as paths or byte arrays.
• Model Quality Presets - Provides predefined quality presets to balance prediction accuracy against training speed and inference latency.
• Forecast Accuracy Validation - Evaluates time series forecast precision by comparing predictions against held-out historical data using error metrics.
• Model Training Optimizers - Tunes hyperparameters, early-stopping, and ensemble weights to maximize evaluation metrics.
• Model Exporting - Saves model weights and configuration files to local directories for later reloading and deployment.
• Training Resumption - Allows resuming model training from an existing checkpoint to improve performance without restarting.
• Out-of-Fold Predictions - Implements out-of-fold prediction strategies to ensure unbiased performance estimation during training.
• Multi-Target Learning - Trains models to predict several target variables simultaneously instead of a single output.
• Multilingual NLP - Finetunes transformer models and uses multilingual presets to process text across different languages.
• Multimodal Entity Extraction - Identifies and classifies specific entities within datasets containing both text and image information.
• Named Entity Recognition - Identifies and classifies key entities such as people and organizations within text strings across multiple languages.
• NER Performance Evaluators - Measures the accuracy of named entity recognition using standard recall, precision, and F1 scores on test datasets.
• Model Finetuning - Adjusts pretrained object detection models using custom COCO-format datasets to improve task-specific accuracy.
• Predictor Snapshotting - Allows saving the predictor state to a directory to enable loading without retraining.
• Semantic Similarity Calculation - Determines conceptual similarity between pairs of text, images, or mixed-modal inputs to enhance search ranking.
• Semantic Search - Ranks documents by calculating cosine similarities between query and document embeddings.
• Semantic Segmentation - Identifies and outlines specific objects within images by predicting a class mask for every pixel.
• Temporal Train-Test Splitting - Divides datasets into training and testing sets based on a specified prediction length for evaluation.
• Time Series Backtesting - Estimates forecast quality by evaluating models across multiple historical cutoff points to simulate real-world deployment scenarios.
• Dynamic Covariate Integration - Uses external dynamic features, including known future events, to improve forecast accuracy.
• Forecast Aggregation - Aggregates predictions from multiple base models using weighted averages or gradient-based optimization.
• Forecast Evaluation - Measures time series prediction accuracy using built-in metrics like MAE and RMSE or custom scoring logic.
• Statistical Forecasting - Captures trends and seasonality using automated ARIMA and exponential smoothing models.
• Time Series Backtesting - Generates forecasts across multiple sliding or expanding validation windows to evaluate model performance over time.
• Training Curve Analysis - Visualizes performance metrics recorded during training to diagnose model behavior and accuracy improvement over time.
• Training Optimizations - Controls total training time via limits, presets, and data subsampling to optimize resource consumption.
• Vector Similarity Search - Determines semantic similarity between images by projecting them into high-dimensional vector embeddings.
• Zero and Few-Shot Learning - Classifies new categories based on a very small number of provided examples per class.
• Zero-Shot Classification Models - Categorizes images into previously unseen classes by leveraging pretrained vision-language models.
• Zero-Shot Classification Systems - Identifies image contents by comparing visual features against natural language text categories without training data.
• Model Serving & Deployment - Packages trained predictors into containers or serverless functions for real-time or batch inference.
• Text Matching - Trains models to determine if two sentences share the same meaning by projecting them into high-dimensional vectors.
• Multi-Label Classifiers - Predicts multiple target labels for a single input, supporting both mutually and non-mutually exclusive tags.
• Data Cleaning Procedures - Removes redundant information by dropping duplicate columns or columns with a single unique value.
• Foundation Model Training - Trains specialized pretrained foundation models like Mitra and TabPFNv2 for tabular classification and regression.
• Static Metadata Integration - Adds time-independent attributes like location or product IDs to help models group similar entities.
• Feature Generation - Transforms forecasting tasks into tabular regression problems by generating lag and time features.
• Cloud Deployment - Hosts trained models as persistent endpoints or batch transform jobs in cloud environments.
• Serverless Deployment - Packages trained models into container images for execution in serverless environments like AWS Lambda.
• Large Dataset Optimizations - Performs inference on files exceeding system memory by chunking data into smaller sizes.
• Tabular Data Type Inference - Automatically detects if columns are numerical, categorical, boolean, datetime, or text based on values.
• Machine Learning Libraries - Automated machine learning for tabular and multimodal data.
• Automated Machine Learning - Automated feature and model selection for tabular, image, and text data.

Star history