New Features
- All common models can be converted to TorchScript format by tracing or scripting (tutorial). Requires pytorch≥1.8.
- Support fvcore parameter schedulers (originally from ClassyVision) that are composable, scale-invariant, and can be used on parameters other than learning rate.
- Refactor PointRend as a mask head (instead of an ROIHead).
- New export and C++ deployment examples.
- Release d2go which provides end-to-end production pipeline.
New Features in DensePose:
Release DensePose CSE (a framework to extend DensePose to various categories using 3D models) and DensePose Evolution (a framework to bootstrap DensePose on unlabeled data). See here for more details.
Deprecations:
- Deprecate cfg argument from COCO/LVIS evaluator; Deprecate num_classes and ignore_label argument from
SemSegEvaluator
- Deprecate
WarmupMultiStepLR
,WarmupCosineLR
in favor of fvcore schedulers - Deprecated features will be removed in future releases
Pre-built Linux binaries are available for the following environment:
CUDA | torch 1.8 | torch 1.7 | torch 1.6 |
---|---|---|---|
11.1 | install | ||
11.0 | install | ||
10.2 | install | install | install |
10.1 | install | install | install |
9.2 | install | install | |
cpu | install | install | install |