Vision Calorimeter (ViC) and Anti-Neutron Transformer (ANT)

Vision Calorimeter for High-Energy Particle Detection

In high-energy physics, estimating anti-neutron parameters (position and momentum) using the electromagnetic calorimeter (EMC) is crucial but challenging. To conquer this challenge, we propose Vision Calorimeter (ViC), a framework that migrates visual object detectors to analyze particle images. The motivation lies in introducing a physics-inspired heat-conduction operator (HCO) into the detector's backbone and head to handle the discrete and sparse patterns of these images. Implemented via the Discrete Cosine Transform, HCO extracts frequency-domain features, bridging the distribution gap between natural and particle images. Experiments demonstrate that ViC significantly outperforms conventional methods, reducing the incident position prediction error by 46.16% (from 17.31° to 9.32°) and providing the first baseline result with an incident momentum regression error of 21.48%. This study underscores ViC's great potential as a reliable particle detector for high-energy physics. Code is available at https://github.com/yuhongtian17/ViC.

Anti-Neutron Transformer for High-Energy Particle Detection

In the field of high-energy particle physics, anti-neutron ($\bar{n}$) serves as an important probe for studying fundamental units of the material world. Due to the discrete and sparse nature of data, it is challenging to measure the $\bar{n}$ features using machine learning algorithms. Recent studies have tackled the $\bar{n}$ detection problem using computer vision techniques, which, however, are puzzled by representation redundancy and model generalizability when processing discrete and sparse detector readouts. In this study, we propose the Anti‑Neutron Transformer (ANT), by handling the high-energy detection problem with inspiration from language processing. The motivation lies in the fact that Transformer models enjoy not only serialized representation consistent with discrete deposition patterns of $\bar{n}$ but also higher information density applicable to sparse data. We treat each $\bar{n}$ event as a "sentence" and each deposited energy point as a "word", constructing a token sequence for a Transformer model. The energy points also serve as spatial "anchors" for prediction, enabling the model to learn their correlations with the incident position and momentum of $\bar{n}$. ANT undergoes extensive pre-training on quantities of unlabeled $\bar{n}$ events to recover masked position values, providing underlying prior knowledge about spatial radiation patterns of energy. Experiments on high-energy particle data show that ANT achieves state-of-the-art performance, while outperforming visual detectors in terms of statistical analysis and physical application. Code is available at https://github.com/yuhongtian17/ViC.

Dataset

A valset of $\bar{n}$ dataset is available.

Install MMDetection Step by Step

Yes indeed, it depends on PyTorch, MMCV, MMEngine and MMDetection:

# Also: cann-8.0.rc2, torch-2.1.0, torch_npu-2.1.0.post6, mmengine-0.10.5, mmcv-2.2.0, mmdet-3.3.0

wget https://developer.download.nvidia.com/compute/cuda/12.4.1/local_installers/cuda_12.4.1_550.54.15_linux.run
chmod +x ./cuda_12.4.1_550.54.15_linux.run
sudo sh cuda_12.4.1_550.54.15_linux.run

# Add CUDA path
echo "export PATH=/usr/local/cuda-12.4/bin:\$PATH" >> ~/.bashrc
echo "export LD_LIBRARY_PATH=/usr/local/cuda-12.4/lib64:\$LD_LIBRARY_PATH" >> ~/.bashrc
echo "" >> ~/.bashrc
source ~/.bashrc
nvcc -V

# NO sudo when install anaconda
wget https://mirror.tuna.tsinghua.edu.cn/anaconda/archive/Anaconda3-2024.10-1-Linux-x86_64.sh
chmod +x ./Anaconda3-2024.10-1-Linux-x86_64.sh
./Anaconda3-2024.10-1-Linux-x86_64.sh

conda create -n openmmlab241b python=3.9 -y
conda activate openmmlab241b
# ref: https://pytorch.org/get-started/previous-versions/#v241
conda install pytorch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 pytorch-cuda=12.4 -c pytorch -c nvidia

pip install pip==24.3.1 --index https://pypi.tuna.tsinghua.edu.cn/simple/
pip install numpy==1.26.4 ninja==1.11.1.1 psutil==6.1.0 --index https://pypi.tuna.tsinghua.edu.cn/simple/
pip install matplotlib==3.9.3 opencv-python==4.10.0.84 cython==3.0.11 --index https://pypi.tuna.tsinghua.edu.cn/simple/

pip install -U openmim
mim install mmengine==0.10.5
mim install mmcv==2.2.0
# mim install mmdet==3.3.0

# git clone https://github.com/open-mmlab/mmdetection.git
# cd mmdetection
wget https://github.com/open-mmlab/mmdetection/archive/refs/tags/v3.3.0.zip -O mmdetection-3.3.0.zip
unzip mmdetection-3.3.0.zip
cd mmdetection-3.3.0/

# Modify Line 9 in "./mmdet/__init__.py" to >>> mmcv_maximum_version = '3.0.0'
pip install -v -e . --index https://pypi.tuna.tsinghua.edu.cn/simple/

# IF MSCOCO2017 Dataset does not exist
mkdir -p "./data/coco/"
cd "./data/coco/"
wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip train2017.zip
unzip val2017.zip
unzip annotations_trainval2017.zip
cd ../../

# Try using MMDetection to train RetinaNet with MSCOCO2017
CUDA_VISIBLE_DEVICES=0,1,2,3 ./tools/dist_train.sh ./configs/retinanet/retinanet_r50_fpn_1x_coco.py 4
CUDA_VISIBLE_DEVICES=0,1,2,3 ./tools/dist_train.sh ./configs/swin/retinanet_swin-t-p4-w7_fpn_1x_coco.py 4

Train and Test

pip install yacs timm uproot openpyxl einops fvcore --index https://pypi.tuna.tsinghua.edu.cn/simple/

# Download our code
cd ../
git clone https://github.com/yuhongtian17/ViC.git
cp -r ViC/mmdetection-main/* mmdetection-3.3.0/
cd mmdetection-3.3.0/

# Prepare dataset
# NOTE: We regret that the release of *.root files requires further data-sharing agreements with BESIII.
cd "./data/"
unzip BESIII_training_sample.zip
cd ../
python ./tools/dataset_converters/root_to_json.py --srcroot "./data/BESIII_training_sample/Nm_1m.root" --df_prefix "Nm_1m" --fn_prefix "Nm_1m"

# ##### ##### ##### ##### ##### ##### ##### ##### ##### ##### #

# Prepare ViC's pre-trained model
mkdir -p "./data/pretrained/"
cd "./data/pretrained/"
wget https://github.com/MzeroMiko/vHeat/releases/download/vheatcls/vHeat_tiny.pth
python ../../vheat_pth_tools/interpolate4downstream.py --pt_pth 'vHeat_tiny.pth' --tg_pth 'vheat_tiny_512.pth'
cd ../../

# Train ViC
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=33010 ./tools/dist_train.sh "./configs/_hep2coco_/abla/hep-retinanet_vheatk-tiny_fpn_1x_hep2coco-1m_8xbs16.py" 8
# Test ViC
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=33020 ./tools/dist_test.sh "./configs/_hep2coco_/abla/hep-retinanet_vheatk-tiny_fpn_1x_hep2coco-1m_8xbs16.py" "./work_dirs/hep-retinanet_vheatk-tiny_fpn_1x_hep2coco-1m_8xbs16/epoch_12.pth" 8 --out "./work_dirs/results_vic_1m_ep12.pkl"
python ./tools/analysis_tools/hep_eval.py --pkl_path "./work_dirs/results_vic_1m_ep12.pkl" --json_path "./data/HEP2COCO/Nm_1m/Nm_1m__b00000001__e00100000.json"

# ##### ##### ##### ##### ##### ##### ##### ##### ##### ##### #

# Prepare ANT's pre-trained model
cd "./data/pretrained/"
wget https://github.com/yuhongtian17/ViC/releases/download/ANT-checkpoints/selfsup_50x-20250721.pth
mv selfsup_50x-20250721.pth selfsup_50x.pth
cd ../../

# Train and test ANT
model_type="abla"
model_i="hepv2-ssd_trans-base-selfsup_nofpn_trans-head-selfsup_384c-50x_1x_hep2seq-1m_8xbs64"
eval_json="./data/HEP2COCO/Nm_1m/Nm_1m__b00000001__e00100000.json"

./tools/dist_train.sh "./configs/_hep2seq_/${model_type}/${model_i}.py" 8
./tools/dist_test.sh "./configs/_hep2seq_/${model_type}/${model_i}.py" "./work_dirs/${model_i}/epoch_12.pth" 8 --out "./work_dirs/${model_i}/results_ant_1m_ep12.pkl"
python ./tools/analysis_tools/hep_eval.py --pkl_path "./work_dirs/${model_i}/results_ant_1m_ep12.pkl" --json_path "${eval_json}"

# NOTE: Performance with our "ANT_Nm-1m_epoch_12.pth": mAB 9.17° mRE 20.78%

Bugs Report

(1) An error may occur on the new server: "ImportError: libGL.so.1: cannot open shared object file: No such file or directory." It can be solved by the following shell command:

sudo apt update
sudo apt install libgl1-mesa-glx

(2) An error may occur on the RTX 4090 server when training ANT: "RuntimeError: received 0 items of ancdata". It can be solved by the following shell command:

echo "ulimit -n 1048576" >> ~/.bashrc
source ~/.bashrc
ulimit -n

The output result is 1048576 instead of 1024, indicating successful modification.

License

ViC is released under the License.

Appendix A: How to use OpenMMLab series with Ascend 910B

Supported: mmengine-0.10.5, mmcv-latest (>2.2.0, main-241212), mmpretrain-latest (>1.2.0, main-241212), mmdetection-3.3.0, mmyolo-0.6.0, mmrotate-1.x.

# mirror: cann8.0.RC2-torch2.1.0-conda24.7.1-vscode4.12.0-ubuntu22.04-ssh-arm64
# Execute this command every time a new console is opened!
source /usr/local/Ascend/ascend-toolkit/set_env.sh

conda create -n openmmlab210p6b python=3.9 -y
mkdir -p "/workspace/all-data/envs/"
mv /home/miniconda3/envs/openmmlab210p6b/ /workspace/all-data/envs/
ln -s /workspace/all-data/envs/openmmlab210p6b/ /home/miniconda3/envs/
conda activate openmmlab210p6b
mkdir -p "/workspace/openmmlab210p6b/"
cd "/workspace/openmmlab210p6b/"

# ref: https://www.hiascend.com/document/detail/zh/Pytorch/60RC3/configandinstg/instg/insg_0001.html
# Download PyTorch installation package
wget https://download.pytorch.org/whl/cpu/torch-2.1.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
# Download torch_npu plugin package
wget https://gitee.com/

ViC

Install / Use

README