Installation¶
Clone the repository, but make sure to recursively clone the submodules too.
- ::
git clone –recursive git@github.com:bmmtstb/dynamically-gated-similarities
If you cloned the repository already, use git submodule update --init --recursive
.
Packages and Environment¶
Create a new python 3.10
environment and install the requirements.
This will also install the required submodules. This will take a while!
Note
This will install PyTorch with support for CUDA 12.1, you might want to change that for your environment.
- ::
python3.10 -m venv venv
Activate the environment according to you OS, then install the base requirements.
- ::
pip install -r requirements.txt
To properly install the torchreid
-package, we need to run the setup scripts in the submodules / dependencies.
All modules in the ./dependencies
folder are submodules, make sure that they have been cloned properly.
Some of the respective requirements of torchreid
have been installed by this packages requirements file already.
And finally, make sure to install the dgs module itself.
- ::
cd ./dependencies/torchreid python setup.py develop cd ../.. pip install -e .
Now the dgs module is installed, including two evaluation tools:
The
PoseTrack21
evaluation toolkit, namedposetrack21
. Note, that you need to download thePoseTrack21
dataset first. (It might be possible to use it for PT17 or PT18 data, but no guarantees.) Have a look at the respective repository for how to do so.The
poseval
-package has been installed by the requirements and can be used for evaluation too. For more information, visit the respective GitHub <https://github.com/leonid-pishchulin/poseval> page.
Next Steps¶
There are a few examples in the ./scripts/ directory, with their explanations here <scripts_page>.
Check the information in weights <weights> to download (more) example model weights or use your own.
There are multiple example configurations in the ./configs/ directory, with some additional explanation here <configs>.
If you want to train your own models, use custom datasets<dataset>, or validate results, make sure to check out the rest of the docs. There are methods to register new modules for every type of existing module, which simplifies the usage of custom modules in the configuration files.
Backbone Models¶
TODO
Pytorch Keypoint-RCNN¶
See KeypointRCNNImageBackbone
.
AlphaPose Backbone¶
Currently it is only possible to use AP as predictor using a separate AlphaPose installation,
and then import the json prediction-files.
There is a WIP-version of a AlphaPose backbone on the alpha_pose
branch.