This folder contains a starter kit for new wisp projects. New projects can be quickly created by cloning this folder, and further customizing its files. The templates follow the popular example of Neural Radiance Fields. Other field types (i.e. Neural Signed Distance Functions) can also be derived from these templates.
template_neural_field is a one-stop for defining new neural fields.
In wisp, a neural field (nef) is as a combination of a feature grid structure, a decoder, and an optional input embedding.
New channels can be supported by implementing and registering additional forward functions here.
The input to forward functions is provided by BaseTracer implementation.
For example, a PackedRFTracer, which traces radiance fields, passes the input coordinates and ray direction.
Consequentially, Neural field implementations should be mindful of which tracers they're compatible with.
template_trainer orchestrates the optimization process, from data load to loss calculation and metrics logging.
BaseTrainer functions can be overrided to take control of lifecycle events such as pre_epoch(), a single training step(), and
invocation of validation(). The trainer typically holds a Pipeline object, which pairs the neural field currently being optimized,
and a compatible BaseTracer. Pipeline contains all components required to differentiably render a neural field.
template_app is an exemplary implementation of an interactive app.
The app controls which gui widgets are created and how visual properties of the canvas are customized.
It can also register event handlers due to changes within the shared WispState object.
The scene graph can also be manipulated here.
template_config contains configurations for customizing the templates defined above, as well as existing
Wisp components. The configurations are parsed by the config_parser, and passed directly to built
components.
template_main is a simple main script for tying the entire templates example together.
It registers newly added user files to ensure config_parser knows how to load them.
It then creates the shared WispState object, and the template trainer and app objects which
share this state. The app loop is finally run here to initiate the interactive app.
wisp.ops contains useful building blocks for further customizing the templates.
Advanced users may customize other building blocks such as Tracers, Datasets and BottomLevelRenderers, currently not provided in the form of templates.
Come up with useful components? We welcome you to submit a PR!