How sktime/pytorch-forecasting works
PyTorch Forecasting is a high-level deep-learning toolkit for time-series forecasting. The documented workflow starts with tabular observations, prepares them as a TimeSeriesDataSet with transformations and history/forecast horizons, then trains a selected forecasting model with Lightning and evaluates multi-horizon predictions using a chosen loss or metric. Users can also tune hyperparameters. The graph includes the documented model and metric capabilities; unsampled implementations are represented by their exact tree paths, and edges reflect README statements or source-index references rather than assumptions about internal calls.
Main entry points: _base_model.py Forecasting model API · _timeseries.py Time-series dataset · _base_pkg.py Model configuration and run · _encoder_decoder_data_module.py Encoder-decoder data module
Components 24
Data preparation
- Time-series dataset
pytorch_forecasting/data/timeseries/_timeseries.py - Encoders and normalizers
pytorch_forecasting/data/encoders.py - Batch samplers
pytorch_forecasting/data/samplers.py - Encoder-decoder data module
pytorch_forecasting/data/data_module/_encoder_decoder_data_module.py - Time-series data module
pytorch_forecasting/data/data_module/_tslib_data_module.py
Forecasting models
- Forecasting model API
pytorch_forecasting/models/base/_base_model.py - Temporal Fusion Transformer
pytorch_forecasting/models/temporal_fusion_transformer/_tft.py - TFT submodules
pytorch_forecasting/models/temporal_fusion_transformer/sub_modules.py - Model architectures
pytorch_forecasting/models - Baseline forecast
pytorch_forecasting/models/baseline.py
Training and evaluation
- Lightning Trainer
- Multi-horizon metrics
pytorch_forecasting/metrics/base_metrics/_base_metrics.py - Distribution losses
pytorch_forecasting/metrics/distributions.py - Neural loss adapter
pytorch_forecasting/metrics/nn_loss_adapter.py - MQF2 utilities
pytorch_forecasting/metrics/_mqf2_utils.py - Hyperparameter tuning
pytorch_forecasting/tuning/tuner.py
Shared model components
- Model base objects
pytorch_forecasting/base/_base_object.py - Model configuration and run
pytorch_forecasting/base/_base_pkg.py - Categorical embeddings
pytorch_forecasting/models/nn/embeddings.py - Positional embeddings
pytorch_forecasting/layers/_embeddings/_positional_embedding.py - Scaler adapter
pytorch_forecasting/adapters/scaler_adapters.py
Other components
- Forecasting user
- Tabular observations
- Forecast predictions
Connections 28
- Forecasting user → Tabular observations: provides
- Forecasting user → Time-series dataset: configures
- Tabular observations → Time-series dataset: converts to
- Time-series dataset → Encoders and normalizers: transforms with
pytorch_forecasting/data/timeseries/_timeseries.py - Time-series dataset → Batch samplers: uses
pytorch_forecasting/data/timeseries/_timeseries.py - Forecasting user → Forecasting model API: selects
- Forecasting model API → Temporal Fusion Transformer: supports
pytorch_forecasting/models/base/_base_model.py - Forecasting model API → Model architectures: supports
pytorch_forecasting/models/base/_base_model.py - Forecasting model API → Baseline forecast: supports
- Temporal Fusion Transformer → TFT submodules: uses
pytorch_forecasting/models/temporal_fusion_transformer/_tft.py - Forecasting user → Lightning Trainer: trains with
- Lightning Trainer → Forecasting model API: trains
- Lightning Trainer → Time-series dataset: uses
- Forecasting model API → Forecast predictions: produces
pytorch_forecasting/models/base/_base_model.py - Forecast predictions → Multi-horizon metrics: evaluated by
- Forecasting model API → Distribution losses: supports
pytorch_forecasting/models/base/_base_model.py - Distribution losses → MQF2 utilities: uses
pytorch_forecasting/metrics/distributions.py - Multi-horizon metrics → Neural loss adapter: adapts through
pytorch_forecasting/metrics/base_metrics/_base_metrics.py - Forecasting user → Hyperparameter tuning: tunes
- Hyperparameter tuning → Lightning Trainer: configures
- Encoder-decoder data module → Encoders and normalizers: uses
pytorch_forecasting/data/data_module/_encoder_decoder_data_module.py - Encoder-decoder data module → Time-series dataset: uses
pytorch_forecasting/data/data_module/_encoder_decoder_data_module.py - Scaler adapter → Encoders and normalizers: adapts
pytorch_forecasting/adapters/scaler_adapters.py - Categorical embeddings → Positional embeddings: uses
pytorch_forecasting/models/nn/embeddings.py - Model configuration and run → Lightning Trainer: creates
pytorch_forecasting/base/_base_pkg.py - Model configuration and run → Encoder-decoder data module: builds
- Model configuration and run → Forecasting model API: builds
pytorch_forecasting/base/_base_pkg.py - Forecasting model API → Categorical embeddings: uses
pytorch_forecasting/models/base/_base_model.py