Core Components Overview

Torch-RecHub uses a modular design that divides the core functionality of a recommender system into components for feature processing, data pipelines, model design, training, and evaluation. This modular design makes the framework easy to use and extend while preserving a well-organized codebase.

Core Component Architecture

Torch-RecHub's core component architecture is organized as follows:

1. Feature layer: Handles different types of features, including numerical, categorical, and sequence features
2. Data layer: Handles data loading, preprocessing, and DataLoader generation
3. Model layer: Implements recommendation models, including ranking, matching, multi-task, and generative recommendation models
4. Training layer: Provides unified training interfaces for model training, evaluation, prediction, and ONNX export
5. Tools layer: Provides utilities such as ONNX export, model visualization, callbacks, and loss functions

Component Relationships

The core components relate to one another as follows:

1. Feature layer -> Data layer: Feature definitions guide preprocessing and feature engineering in the data layer
2. Data layer -> Training layer: DataLoaders produced by data generators are used for model training and evaluation
3. Model layer -> Training layer: The training layer trains and evaluates models defined in the model layer
4. Training layer -> Tools layer: The training layer uses utilities for tasks such as ONNX export and model visualization

Component Details

Feature Processing

The feature-processing component defines and handles different types of features, including:

• DenseFeature: Handles numerical features
• SparseFeature: Handles categorical features
• SequenceFeature: Handles sequence or multi-hot features

See Feature Definitions for details.

Data Pipeline

The data-pipeline component handles data loading, preprocessing, and DataLoader generation, including:

• TorchDataset: A dataset for training and validation
• PredictDataset: A dataset for prediction
• DataGenerator: Generates DataLoaders for ranking and multi-task models
• MatchDataGenerator: Generates DataLoaders for matching models
• SequenceDataGenerator: Generates DataLoaders for HSTU/HLLM sequence tasks
• ParquetIterableDataset: Streams Parquet files batch by batch

See Data Pipeline for details.

Training and Evaluation

The training-and-evaluation component trains different types of recommendation models, including:

• CTRTrainer: Trains ranking models
• MatchTrainer: Trains matching models
• MTLTrainer: Trains multi-task models
• SeqTrainer: Trains HSTU/HLLM sequence-generation models

RQ-VAE and TIGER use dedicated data-processing and training workflows. See Generative Models and the corresponding reproduction guides.

See Training and Evaluation for details.