descriptron-vision

Mask and landmark prediction for Descriptron: torchvision detectors, SAM2-PAL propagation and DINOv3 landmark transfer.

pip install descriptron-vision
descriptron-train --task masks --coco-json annotations.json --img-dir images/ \
                  --output-dir out/ --total-iters 20000
descriptron-predict --checkpoint out/model_final_*.pth --coco-json annotations.json \
                    --img-dir images/ --annotations_out predictions.json

No compiler required. torch and torchvision are ordinary wheels on Linux, macOS and Windows. For CUDA, install torch from PyTorch's own index first:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install descriptron-vision

On a Mac the detectors can use the Apple GPU through PyTorch's MPS device, which is why this backend exists: Detectron2 has no macOS wheels and no Apple Silicon support.

Two optional dependencies that cannot be declared

PyPI rejects a git dependency in package metadata, so neither SAM2 nor Detectron2 is listed. Both are imported lazily, and the error names the command to install them. The Docker image carries both ready to run.

Choosing a backbone

--arch v1 (default) is maskrcnn_resnet50_fpn; --arch v2 is maskrcnn_resnet50_fpn_v2. v2 scores higher on COCO, but on a few hundred training images v1 won on 3 of 3 folds of a grouped species hold-out — its lighter two-FC box head converges further in the same budget. Measured on the data, not assumed from the model zoo.

Metadata

Release files for descriptron-vision 2.3.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for descriptron-vision 2.3.0
File	Size	Uploaded
descriptron_vision-2.3.0.tar.gz	178.8 kB	Sep 30, 2026	Details

Built distribution (wheel)

Table of built distributions (wheels) for descriptron-vision 2.3.0
File	Interpreter	ABI	Platform	Reset
descriptron_vision-2.3.0-py3-none-any.whl 196.2 kB Sep 30, 2026	Python 3	none	any	Details

Total release size: 375.0 kB

Release files / descriptron_vision-2.3.0.tar.gz

Download URL	descriptron_vision-2.3.0.tar.gz
Size	178.8 kB
Tags	Source
SHA-256 checksum How to use checksums	`f7c1a73b17ab6bcfd23eea2be3e6a7ebed74c8f7d6779c8b3dbb5d8db4cf7119`
BLAKE2b-256 checksum How to use checksums	`5622dd96c222376c402a26e32dcdd5b1a3cd37235b90da4982557d2a0222dd74`
Upload date	Sep 30, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.12.4`

Release files / descriptron_vision-2.3.0-py3-none-any.whl

Download URL	descriptron_vision-2.3.0-py3-none-any.whl
Size	196.2 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`d8b285b36141174e6b76b09ba9d0a7b6983685277e947ca5db3968b1157db45f`
BLAKE2b-256 checksum How to use checksums	`e9e7bee7ee11753c9b1123a67e3b435132b5d72d04cc8ccdf2d68e7ce4d68c5d`
Upload date	Sep 30, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.12.4`