Original Code is Lore I modified this code with faster inference and higher accuracy by removing locations regression and add a simple and efficient postprocess
export PATH="/usr/local/cuda-11.8/bin:$PATH"
export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
source ~/.bashrc
conda create --name Lore python=3.7
conda activate Lore
pip install -r requirements.txt
pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
chmod +x *.sh
cd lib/models/networsk/DCNv2
python3 setup.py build develop
Note: if It causes error : Unknown CUDA arch (8.9,9.0,...) or GPU not supported --> export TORCH_CUDA_ARCH_LIST=7.5 or export TORCH_CUDA_ARCH_LIST=6.0 .The rerun setup.py
Available model weights (using dla-34 backbone):
| Name | Backbone | Regressor Arc | Image Size | Checkpoint |
|---|---|---|---|---|
| ckpt_wtw | DLA-34 | 4+4 | 1024 | Trained on WTW |
| ckpt_ptn | DLA-34 | 3+3 | 512 | Trained on PubTabNet |
| ckpt_wireless | ResNet-18 | 4+4 | 768 | Trained on Wireless Tables* |
python inference.py
Availabel Dataset for training Yolo in Here