Detectron2 is a platform for object detection, segmentation and other visual recognition tasks.
Pre-built Linux binaries are available for the following environment:
CUDA | torch 1.10 | torch 1.9 | torch 1.8 |
---|---|---|---|
11.3 | install | ||
11.1 | install | install | install |
10.2 | install | install | install |
10.1 | install | ||
cpu | install | install | install |
Pre-built Linux binaries are available for the following environment:
CUDA | torch 1.9 | torch 1.8 | torch 1.7 |
---|---|---|---|
11.1 | install | install | |
11.0 | install | ||
10.2 | install | install | install |
10.1 | install | install | |
9.2 | install | ||
cpu | install | install | install |
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.
SemSegEvaluator
WarmupMultiStepLR
, WarmupCosineLR
in favor of fvcore schedulersPre-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 |
projects/
.point_rend
, deeplab
, panoptic_deeplab
) directly with import detectron2.projects.xxx
.cfg.SOLVER.AMP.ENABLED
) and inference.ade20k_sem_seg_train
, ade20k_sem_seg_val
).Pre-built Linux binaries are provided for the following environment:
CUDA | torch 1.7 | torch 1.6 | torch 1.5 |
---|---|---|---|
11.0 | install | ||
10.2 | install | install | install |
10.1 | install | install | install |
9.2 | install | install | install |
cpu | install | install | install |
CUDA | torch 1.6 | torch 1.5 | torch 1.4 |
---|---|---|---|
10.2 | install | install | |
10.1 | install | install | install |
10.0 | install | ||
9.2 | install | install | install |
cpu | install | install | install |
TransformGen
to Augmentation
and keep TransformGen
as an alias. Design the interface of Augmentation
so that it can access arbitrary custom data types. See augmentation tutorial for details.COCOEvaluator
by about 3xDefaultTrainer
. See cfg.SOLVER.REFERENCE_WORLD_SIZE
Pre-built Linux binaries are provided for the following environment:
CUDA | torch 1.5 | torch 1.4 |
---|---|---|
10.2 | install | |
10.1 | install | install |
10.0 | install | |
9.2 | install | install |
cpu | install | install |
Bugfix version.
We started to release pre-built wheels for multiple PyTorch versions:
CUDA | torch 1.5 | torch 1.4 |
---|---|---|
10.2 | install | |
10.1 | install | install |
10.0 | install | |
9.2 | install | install |
cpu | install | install |
_init_{box,mask,keypoint}_head
of StandardROIHeads
was changed from instance method to class method.The pre-built wheels for this version have to be used with an official binary release of PyTorch 1.5.
resume_or_load()
, training states like optimizer
, start_iter
will only be loaded when resume
is True and the last checkpoint is found. This matches users’ expectations better.output_size
in custom box head is renamed to .output_shape
feature_strides
and feature_channels
attributes are removed from ROIHeads
. Use the input argument input_shape
instead.layers()
method._forward_{box,mask,keypoint}
methods of StandardROIHeads
now accept dict of features.This release is made to be compatible with such changes in projects (Mesh R-CNN, PointRend, etc)
The pre-built wheels for this version have to be used with an official binary release of PyTorch 1.4.
Some major additional features since open source:
projects/
.detectron2.model_zoo
APIs.We start to provide pre-built binary wheels at https://dl.fbaipublicfiles.com/detectron2/wheels/index.html. The pre-built wheels for this version have to be used with an official binary release of PyTorch 1.4.