Pose Transformer
The implementation is based on this implementation of Deformable DETR.
Because the original implementation structures the MSDeformAttention
module as a separate Python package with its own Python, CUDA, and C++ scripts, one needs to build this before running the training.
First create a dedicated Conda environment with Python 3.7:
conda create -n deformable_potr python=3.7 pip
conda activate deformable_potr
Then install all the dependencies using:
conda install pytorch=1.5.1 torchvision=0.6.1 cudatoolkit=9.2 -c pytorch
pip install -r requirements.txt
And finally, build the MSDeformAttention
module (and all other custom CUDA operators) using:
cd ./deformable_potr/models/ops
sh ./make.sh
You can then test the CUDA operators by running python test.py
, where all tests should pass.
cd ./utils
python crops2h5.py ./nyu/train_comrefV2V_3Dproj.json