Note : re-added missing fmpose3D dependency

Important : trying to run CI with pinned protobuf for functional tests, as functional tests started failing in CI without any repository-side dependency changes. The failure occurs during deeplabcut.export_model() when TensorFlow writes the graph, raising:
TypeError: MessageToString() got an unexpected keyword argument 'float_format'

This is likely due to dependency drift in the runner image / resolver (e.g. -latest / Python minor version changes) causing an incompatible protobuf runtime to be installed alongside TensorFlow, even though DeepLabCut itself is pinned to a commit.

Note; trying with protobuf<7 seems to be successful so far. I tried to advertise the file clearly and only use it where necessary to minimize potential side-effects, but I would not call this future-proof either.

Maybe squashing this when merging would be cleaner.

I'm helping out 3 (so far) other researchers who are having difficulties installing DLC after getting my own copy installed last week on my Windows 11 machine.
They are running into a series of similar dependency issues, so I directed them to this solution (#3184).
However, this is a long and complicated series of entries, so I want to understand what exactly we should be doing for the new installation method. Is this correct?

Install uv

Where do they install pytorch? Before or after any of these uv steps?

uv sync --extra gui --python 3.11
uv run deeplabcut

Also, is this the way to install through a pyproject.toml or am I missing something?

Do you know when the official installation instructions will be updated?
Thanks for working on all this!

Hello @amadorkane,

Sorry for the continued trouble, the best would be to install pytorch between those two steps you listed; make sure to check their docs for GPU/CUDA support.

uv sync --extra gui --python 3.11
uv pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu<CUDA_VER> # where CUDA_VER is the cuda version supported by your GPU
uv run deeplabcut

However you will need to clone the repo and install from the main branch to benefit from the pyproject.toml based install, and likely update napari to 0.6.6 manually (or install from the napari-DLC repo main branch by cloning).

I understand this is complicated, and I apologize for the situation, but we are doing our best to fix this as soon as possible while ensuring we do not worsen the situation either.

We are almost finished merging + releasing fixes across repositories, but currently the installation is problematic due to pinned versions in the napari-DLC dependencies. A release is staged but it will take a little longer for us to align everything; sorry again.

Once we have this and ideally #3225 merged, we can finally make an easier-to-install rc release and update the docs.
In the meantime, please send any errors you encounter and I will assist with the installation.

Best,
Cyril

Thank you very much Cyril! When I saw how many pages of dependencies there were after installation, I got the complexity of the problems you are up against!

I am not sure how to do this: "clone the repo and install from the main branch". In #3219, I was advised to try:
git clone https://github.com/deruyter92/DeepLabCut.git
cd DeepLabCut
git checkout deruyter92/arash_and_jaap/uv_packaging

But that led to an error when doing the last step. The method that finally worked did not involved cloning the forked repo so I'm unclear how to do this part correctly still.

Also, is the following correct for the additional napari steps? Also,

uv sync --extra gui --python 3.11
uv pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu<CUDA_VER> # where CUDA_VER is the cuda version supported by your GPU

uv pip3 install -v "napari==0.6.6" <-- correct?
uv pip3 install napari-deeplabcut <-- correct?

uv run deeplabcut

Hello @amadorkane,

Apologies for the delay, we are now nearly done with the release.

In the meantime, my colleague @arashsm79 kindly suggested the following, which I edited a bit to work on Windows (in PowerShell) :

mkdir dlc_project
cd dlc_project
uv venv -p 3.12 .dlc_venv
& .\.dlc_venv\Scripts\activate.ps1
uv pip install torch torchvision --torch-backend=cu<CUDA_VER>
uv pip install "deeplabcut[gui] @ git+https://github.com/deeplabcut/deeplabcut.git"
uv run -m deeplabcut

No cloning, should work out of the box, hopefully.
(The GUI might take a bit to open on first launch)

For the CUDA_VER, check the nvidia-smi command output, in the upper right you will see something like :

Simply replace <CUDA_VER> with e .g. 131 or 126 or 128 depending on what you see (no separator, no brackets).

Please let me know if this worked,

Best,
Cyril