Install¶

The core needs numpy and OpenCV only. Everything else is an optional extra, so you install the one your models actually need. Releases are on PyPI at pypi.org/project/vizor.

pip install vizor

That gives you Vizor, Refiner, Tracks, Vote and the video helpers. It does not give you any model, and it does not give you a detector.

Extras¶

Each extra pulls in the dependency for one group of model wrappers.

Combine them with a comma.

pip install "vizor[api,hf]"

Pkl needs no extra. It replays predictions you saved earlier, so it works on the core install.

From source¶

Clone and install in editable mode if you want to change the package or run the tests.

git clone https://github.com/Y-T-G/vizor
cd vizor
pip install -e ".[dev]"
pytest

The suite needs no GPU, no API key and no model weights. It stubs the secondary instead of calling one.

107 passed in 3.82s

The four replay tests in tests/test_replay.py skip themselves when the cached predictions are not on disk, and those files are not in the repo, so a fresh clone reports 103 passed and 4 skipped.

Lazy imports¶

import vizor never imports torch, transformers or openai. The model wrappers are resolved through a module-level __getattr__ and imported the first time you name one.

import vizor as vz   # numpy and OpenCV only
model = vz.Florence()  # transformers and torch are imported here

The practical consequence is that a missing extra shows up as an ImportError when you construct the model, not when you import the package.