UV is a drop-in replacement for pip and global python installation. I allows the management of python versions and packages. So far i only used it as a replacement for conda/venv so i only have a shallow understanding of basic feature. The build capabilities are not handled in this page yet.

What is uv

uv is a python package manager that can replace pip and the installation of python itself on any machine. Its coded in rust and is make the management of python project very fast and robust.

All the example provided in this article are assuming you run on a linux system.

it is installable with a simple line in the terminal

curl -LsSf https://astral.sh/uv/install.sh | sh

This should make the uv command available through your shell.

you can check the python versions already available by running

uv python list

if the version of python you want is not already installed you can run the next command (change the 3.14 to the required python version)

uv python install 3.14

Creation on the environment

Unlike conda, uv is based on project directories and not on global environments. So each project will have its own uv configuration. Here, you have two options.

Local quick environement

Only use a local environement that will be used only for this project. This option is good for prototyping.

In order to manage the part with GPU compatibilities, uv provides an help page to get the torch versions matching your locally installed cuda :

1uv venv --python 3.12
uv pip install torch torchvision --torch-backend=auto
2uv pip install -r requirements.txt

1: Creating the virtual environement in the current directory with the specified python version.
2: Install all packages listed in requirement.txt in the local env.

omegaconf

numpy
pandas
seaborn
matplotlib

typer

pycocotools
albumentationsX
opencv-python
torchmetrics
lightning
transformers

Complete and reproducible dev

If you want a complete reproducible management of dependencies you should use the project features of uv.

uv init --python 3.14 .
uv pip install torch torchvision --torch-backend=auto
uv add -r requirements.txt

Here, uv will create a full pyproject.toml where the python version and all the dependencies whith their versions are saved.

Simple Notebook setup

In the case you do not want to setup a big project with requirements for long term re use, but only a small env for fast and easy data analysis for example. You just need to run this command to open a notebook using the env created in previous steps, in this way, you do not need more setup than this.

uv run --with jupyter jupyter lab

The minimal setup requiring absolutly nothing is as follows, uv takes care of installing jupyter and notebooks utils and only the required packages are installed.

uv venv --python 3.12
uv pip install numpy seaborn matplotlib pandas seaborn scipy
uv run --with jupyter jupyter lab

--- title: "UV setup for computer vision using deep learning " author: "Julien Combes" date: "2025-04-13" categories: [DeepLearning, code, ComputerVision, python, linux] image: "image.jpg" --- UV is a drop-in replacement for pip and global python installation. I allows the management of python versions and packages. So far i only used it as a replacement for conda/venv so i only have a shallow understanding of basic feature. The **build capabilities** are not handled in this page yet. # What is uv uv is a python package manager that can replace pip and the installation of python itself on any machine. Its coded in rust and is make the management of python project very fast and robust. All the example provided in this article are assuming you run on a linux system. it is installable with a simple line in the terminal ```bash curl -LsSf https://astral.sh/uv/install.sh | sh ``` This should make the uv command available through your shell. you can check the python versions already available by running ```bash uv python list ``` if the version of python you want is not already installed you can run the next command (change the 3.14 to the required python version) ```python uv python install 3.14 ``` # Creation on the environment Unlike conda, uv is based on project directories and not on global environments. So each project will have its own uv configuration. Here, you have two options. ## Local quick environement Only use a local environement that will be used only for this project. This option is good for prototyping. In order to manage the part with GPU compatibilities, uv provides an [help page](https://docs.astral.sh/uv/guides/integration/pytorch/) to get the torch versions matching your locally installed cuda : ```bash uv venv --python 3.12 # <1> uv pip install torch torchvision --torch-backend=auto uv pip install -r requirements.txt # <2> ``` 1. Creating the virtual environement in the current directory with the specified python version. 2. Install all packages listed in requirement.txt in the local env. ```txt omegaconf numpy pandas seaborn matplotlib typer pycocotools albumentationsX opencv-python torchmetrics lightning transformers ``` ## Complete and reproducible dev If you want a complete reproducible management of dependencies you should use the project features of uv. ```bash uv init --python 3.14 . # <1> uv pip install torch torchvision --torch-backend=auto uv add -r requirements.txt # <2> ``` Here, uv will create a full pyproject.toml where the python version and all the dependencies whith their versions are saved. ## Simple Notebook setup In the case you do not want to setup a big project with requirements for long term re use, but only a small env for fast and easy data analysis for example. You just need to run this command to open a notebook using the env created in previous steps, in this way, you do not need more setup than this. ```bash uv run --with jupyter jupyter lab ``` The minimal setup requiring absolutly nothing is as follows, uv takes care of installing jupyter and notebooks utils and only the required packages are installed. ```bash uv venv --python 3.12 uv pip install numpy seaborn matplotlib pandas seaborn scipy uv run --with jupyter jupyter lab ```