Avalanche: an End-to-End Library for Continual Learning

Abstract

Learning continually from non-stationary data streams is a long-standing goal and a challenging problem in machine learning. Recently, we have witnessed a renewed and fast-growing interest in continual learning, especially within the deep learning community. However, algorithmic solutions are often difficult to re-implement, evaluate and port across different settings, where even results on standard benchmarks are hard to reproduce. In this work, we propose Avalanche, an open-source end-to-end library for continual learning research based on PyTorch. Avalanche is designed to provide a shared and collaborative codebase for fast prototyping, training, and reproducible evaluation of continual learning algorithms.

Avalanche RL: a Continual Reinforcement Learning Library

“…Avalanche RL is built as an extension of Avalanche [16], and it retains the same design principles and a similar API. The target users are practitioners and researchers, and therefore the library must be simple, allowing to setup an Fig.…”

“…While still in its early stages, CRL has seen a rising interest in publications in recent years (according to Dimensions [10] data). To support this growth, we focus on benchmarks and tools, introducing AvalancheRL: we extend Avalanche [16], the staple framework for Continual or Lifelong Learning, to support Reinforcement Learning in order to seamlessly train agents on a continuous stream tasks.…”

“…Avalanche RL is based on PyTorch [23] and supports any Ope-nAI Gym [4] environment. Its design is based on Avalanche [16], one of the more popular continual learning libraries, which allow us to reuse a large number of continual learning strategies and improve the interaction between reinforcement learning and continual learning researchers. Additionally, we propose Continual Habitat-Lab, a novel benchmark and a high-level library which enables the usage of the photorealistic simulator Habitat-Sim [28] for CRL research.…”

Continual Learning with Echo State Networks

“…We train the ESN readout with Adam optimizer and backpropagation, since CL requires to update the model continuously, possibly without storing its activations. We used the Avalanche [15] framework for all our CL experiments. We make publicly available the code together with configurations needed to reproduce all experiments 1 .…”

Continual Learning at the Edge: Real-Time Training on Smartphone Devices

“…other state-of-the-art strategies (Section 4); iv) propose a two phases consolidation to achieve both real-time fast update and off-line delayed optimization (Section 3). The native Android application source-code, along with the Avalanche-based [4] scripts used to reproduce the paper results are made available to further stimulate research in this area 1 .…”