jdb78/pytorch-forecasting
Pytorch Forecasting
هذا إطار عمل للتعلم العميق للتنبؤ بالقيم المستقبلية في البيانات المتسلسلة باستخدام معماريات PyTorch. يوفر مجموعة أدوات للتنبؤ بالسلاسل الزمنية طويلة المدى والاحتمالية، مع دمج خط أنابيب بيانات لتحويل إطارات البيانات الجدولية إلى تسلسلات لتدريب التعلم العميق الخاضع للإشراف.
تستخدم المكتبة غلاف تدريب لتوسيع نطاق تنفيذ النموذج عبر وحدات المعالجة المركزية (CPUs) ووحدات معالجة الرسومات (GPUs). تدعم توليد توزيعات احتمالية للنتائج المستقبلية بدلاً من تقديرات النقطة الواحدة لتحديد عدم اليقين في التنبؤ.
يتضمن إطار العمل قدرات لتنفيذ نماذج التنبؤ، وتحسين المعلمات الفائقة، وتقييم الدقة من خلال مقاييس متعددة الآفاق. كما يوفر طرقًا لقياس أداء المعماريات المعقدة مقابل نماذج أساسية بسيطة.
Features
- Deep Learning Forecasting - Offers a comprehensive toolkit of deep learning architectures for long-horizon and probabilistic time series forecasting.
- PyTorch Lightning Training Orchestration - Provides a high-level API wrapping PyTorch Lightning to scale model training across CPUs and GPUs.
- PyTorch Lightning Workflows - Uses PyTorch Lightning workflows to scale and organize the execution of deep learning training across compute devices.
- Time Series Model Training - Implements deep learning architectures specifically designed for fitting and predicting temporal sequences.
- Tabular-to-Tensor Sequence Mapping - Transforms tabular dataframes into multi-dimensional tensors and sequences required for deep learning model input.
- Time Series Data Engineering - Provides an end-to-end pipeline for preparing raw temporal data, scaling, and transforming it into tensors for ML.
- Time Series Deep Learning Libraries - Provides a deep learning library specifically designed for forecasting and prediction of sequential time series data.
- Time Series Forecasting - Provides models and architectures for predicting future values in sequential data to identify trends and patterns.
- Long-Horizon Forecasting - Provides specialized deep learning architectures and hierarchical interpolation for accurate long-term future value predictions.
- ML-Based Forecasting Models - Provides a toolkit of ML-based forecasting models, such as RNNs and Transformers, for time series prediction.
- Probabilistic Forecasting - Implements probabilistic forecasting by parameterizing output layers to generate probability distributions instead of point estimates.
- Time Series Windowing - Provides time series windowing to map historical data windows to target future windows for supervised learning.
- Time Series Tensor Pipelines - Implements pipelines that convert tabular time series data into normalized tensors for neural network training.
- Deep Learning Training Orchestration - Executes model training on hardware accelerators with performance logging and prediction dependency visualization.
- Hierarchical Temporal Predictions - Implements hierarchical temporal predictions to refine coarse forecasts into finer time-step resolutions for long-term sequences.
- Hyperparameter Optimization - Provides automated methods for hyperparameter optimization to improve the accuracy of forecasting architectures.
- Multi-Horizon Evaluation Metrics - Calculates multi-horizon evaluation metrics to measure forecast accuracy across different future time windows.
- Quantile Regression Models - Utilizes quantile regression models to optimize pinball loss and predict a distribution of potential future outcomes.
- Training Sequence Integration - Provides utilities to convert tabular data into specialized sequences for supervised neural network training.
- Time Series Management - Includes tools for managing time series datasets, including data transformations and missing value handling.
- Analysis Toolkits - Deep learning forecasting implementations using PyTorch.
