awslabs/gluon-ts

Features

• Probabilistic Forecasting - Provides a framework for predicting future values as probability distributions with confidence intervals using deep learning.
• Causal Convolutions - Provides causal convolutional encoding to process temporal signals without leaking future information into the past.
• Deep Learning Architectures - Constructs deep learning architectures using combinations of CNN and RNN blocks to model complex temporal patterns.
• Recurrent Model Definitions - Provides high-level abstractions for constructing recurrent neural network architectures including LSTM and GRU.
• Residual Networks - Implements residual networks with skip-connections to improve gradient flow and numerical stability in deep forecasting models.
• Probabilistic Models - Represents temporal data using probabilistic distributions, such as the Beta distribution, to model uncertainty.
• Sequence Decoders - Implements decoder components that transform encoder outputs into forecasted time series.
• Time Series Model Training - Provides procedures for fitting deep learning models, such as LSTMs and CNNs, to historical time series data.
• Time Series Data Engineering - Provides comprehensive tools for scaling, splitting, and transforming raw temporal data into tensors for machine learning workflows.
• Deep Learning Forecasting - Fits deep learning models using training and validation datasets to produce probabilistic predictors.
• Forecast Evaluation - Measures prediction quality using quantile loss, coverage, and other probabilistic metrics to assess uncertainty.
• Uncertainty Interval Measurements - Deno computes probabilistic metrics such as Mean Scaled Interval Score to assess uncertainty bounds.
• Zero-Shot Forecasting - Implements zero-shot forecasting to generate predictions for new time series using pretrained models without additional training.
• DataFrame Format Converters - Transforms Pandas DataFrames into a standardized dataset format compatible with probabilistic forecasting models.
• Time Series Data Loading - Reads time series datasets from Arrow, Feather, or Parquet files with automatic format detection.
• Time-Series Aggregations - Computes higher-level summaries from bottom-level leaf data using summation matrices for hierarchical time series.
• Time Series Data Normalization - Normalizes input tensors using mean absolute value or min-max scaling to ensure consistent numerical ranges.
• Time Series Tensor Pipelines - Converts variable-length temporal data into uniform tensors through padding and stacking for neural network processing.
• Time Series Transformations - Transforms features into latent representations using dilated convolutions or multilayer perceptrons.
• Forecasting Performance Metrics - Computes standard accuracy metrics, including absolute and squared errors, to quantify differences between predictions and actuals.
• Interval and Quantile Conversions - Standardizes uncertainty representation by transforming prediction interval levels into quantile levels and vice versa.
• Activation Functions - Integrates various non-linear activation functions like LipSwish and SoftRelu into neural network layers.
• Compute Context Management - Detects GPU support and assigns the appropriate compute context for executing forecasting models.
• Distribution Domain Bijections - Uses bijections to map neural network outputs into valid domains for probability distributions.
• Dataset Batch Loading - Loads and organizes data into fixed-size batches to optimize performance during training and inference.
• Dataset Splitting Utilities - Divides temporal datasets into training and testing subsets to evaluate forecasting model performance.
• External Forecasting Library Integrations - Integrates the Prophet library to generate probabilistic predictions within the framework.
• Temporal Feature Extractors - Creates lag, seasonality, and holiday features to incorporate temporal patterns into predictive models.
• Sparse and Dense Feature Importance - Calculates the influence of input features by aggregating gain or split metrics across forecast quantiles.
• Baseline Forecast Generators - Deno calculates the mean of training observations to return a constant value across the horizon.
• Incremental Model Training - Refines existing predictors by training them further on new data to improve forecast accuracy.
• Gated Convolutional Layers - Implements 1D convolutions with gating mechanisms to selectively control information flow within the network.
• Model Deserialization - Provides the ability to load saved models from file paths to resume forecasting immediately.
• Model Persistence - Persists model configurations and trained predictor lists to disk for future inference.
• Quantile Regression - Produces multiple quantile predictions and optimizes them using quantile loss to estimate uncertainty.
• Local Model Training - Supports training unique models for each individual time series to provide tailored predictions.
• Synthetic Time Series Generation - Generates artificial temporal datasets using mathematical and stochastic processes to simulate forecasting scenarios.
• Temporal Test Window Generators - Creates non-overlapping or overlapping test instances from a dataset to evaluate forecast performance across time windows.
• Third-Party Model Integration - Connects third-party time series libraries and statistical models to create a unified deep learning forecasting workflow.
• Baseline Forecasting - Produces simple baseline predictions using constant values or identity mapping to establish performance benchmarks.
• Uncertainty Visualizations - Provides plotting tools to visualize median forecasts and their associated prediction intervals for uncertainty assessment.
• Hierarchical Series Management - Manages and forecasts data organized in nested structures using aggregation matrices to link levels.
• Quantile Regression Models - Generates probabilistic forecasts by training individual quantile regression models for each prediction step.
• Statistical Forecasting - Wraps classical statistical models like ARIMA and ETS to provide probabilistic forecasts.
• Time Series Windowing - Loads raw data and splits it into training and testing windows for supervised forecasting.
• Seasonal Error Analysis - Deno computes the mean absolute difference of a series when shifted by its seasonality.
• Streaming Aggregators - Computes running totals and averages from data streams using a map-reduce approach for efficient aggregation.
• R Language Integrations - Calls external R libraries to perform univariate and hierarchical forecasting within a Python environment.
• Hierarchical Dataset Conversions - Transforms hierarchical structures into a standardized dataset format optimized for probabilistic forecasting.
• Time Series Management - Organizes collections of temporal data using sequential or interleaved iteration to prepare for training.
• Temporal Dataset Splitting - Divides datasets into training and testing subsets using fixed offsets or specific calendar dates.
• Future Covariate Integration - Combines temporal encoder outputs with dynamic future features to provide a comprehensive input for the forecast decoder.
• High-Performance and Parallel Computing - Distributes heavy forecast computation tasks across multiple CPU workers to increase throughput.
• Probability Distribution Transformations - Implements mathematical transformations to convert raw numerical vectors into probability distributions for forecasting.
• Time Series Analysis - Probabilistic time series modeling and forecasting.

Star history