Continuum: PyTorch Framework for Continual Learning Benchmarks

Continvvm released Continuum in 2023 as a PyTorch-based data loading framework for continual and incremental learning. It provides standardized access to common computer vision benchmarks partitioned into specific task-based scenarios.

Use Cases

• Evaluating class-incremental learning performance using the ClassIncremental scenario wrapper
• Benchmarking domain adaptation using the PermutedMNIST or RotatedMNIST scenarios
• Implementing online learning loops using the data stream iterators

Strengths

• Standardized PyTorch API for continual learning scenarios
• Supports multiple incremental learning types (class, domain, task)
• MIT licensed for open research

Limitations

• Does not contain original data; acts as a wrapper for existing benchmarks
• Requires internet connection to download underlying datasets
• Limited to datasets supported by the library's internal classes

Provenance

Source

Continvvm (GitHub)

Collection Method

Aggregated from existing public benchmarks

Freshness

Last updated June 2023; frequency is dependent on GitHub repository maintenance.