Scenic
A JAX and Flax codebase for research on attention-based computer vision models, providing shared libraries for training large-scale, multi-device vision models plus complete example projects.
Scenic is a codebase focused on research around attention-based models for computer vision, developed in JAX and using Flax. It has been used to build classification, segmentation and detection models across multiple modalities, including images, video, audio and multimodal combinations.
It provides a set of shared, lightweight libraries for tasks commonly needed when training large-scale, multi-device and multi-host vision models, such as boilerplate code for launching experiments, summary writing, logging and profiling, optimized training and evaluation loops, losses, metrics and bipartite matchers, input pipelines for popular vision datasets, and baseline models including strong non-attentional baselines. It also contains a number of complete projects that were developed in Scenic or reimplemented in it, such as ViViT, OmniNet and TokenLearner.
It is licensed under Apache-2.0.