Deform-conv

Bachelors theseis project about research and evaluation of modifications to deformable convolution

Note that this repo is still in development

Getting started

If something does not work, please refer to the official MMDetection installation guide and InternImage installation guide

Create a virtual environment

conda create -n deform-conv python=3.10
conda activate deform-conv

Base requirements

pip install -r requirements.txt

MMDetection and friends

pip install -U openmim
mim install mmengine

FORCE_CUDA="1" MMCV_WITH_OPS=1 mim install "mmcv>=2.0.0" -v
# This can take some time, depending on your cuda version, it might compile from source

mim install mmdet -v

Install deformable convolution v3 CUDA op

bash ops_dcnv3/install.sh

Download coco2017 dataset (27G)

sudo apt install unzip
# Or install from your distro's package manager

bash data/download.sh

Training with example model mask-rcnn internimage-tiny

python train.py configs/coco/mask_rcnn_internimage_t_fpn_1x_coco.py