A forked version of the pytorch implementation of Detectron from https://github.com/roytseng-tw/Detectron.pytorch Contact [email protected] or [email protected] for questions.

Be sure to use the larcv1_mcc9 branch!

The original implementation is modified to use a custom particle physics dataset, and run on MicroBooNE LArTPC images. For details see https://microboone.fnal.gov/wp-content/uploads/MICROBOONE-NOTE-1081-PUB.pdf

Further modifications to the original network include the sparsification of the ResNet module, this is performed in https://github.com/NuTufts/Detectron.pytorch/blob/larcv1_mcc9/lib/modeling/SparseResNet.py based off of sparse convolutions from Facebook's SparseConvNet https://github.com/facebookresearch/SparseConvNet which is included in this repo. These sparse convolutions serve to speed up the ResNet runtime for low pixel-occupancy LArTPC images.

To train the network edit a config file based off of: https://github.com/NuTufts/Detectron.pytorch/blob/larcv1_mcc9/configs/baselines/mills_config_2_full_image.yaml

Then run

python tools/train_particle.py --dataset particle --cfg configs/baselines/mills_config_2_full_image.yaml --use_tfboard

To resume training from checkpoint:

python tools/train_particle.py --dataset particle --cfg configs/baselines/mills_config_2_full_image.yaml --use_tfboard --resume --load_ckpt path/to/ckpt/file.pth

To inference on events use infer_particle.py

python tools/infer_particle --dataset particle --cfg configs/baselines/mills_config_2_full_image.yaml --load_ckpt path/to/ckpt/file.pth --image_dir path/to/imagefile/ --output_dir path/to/outpngs/ --num_images 10

The tools directory generally contains the training and analysis scripts used to examine the network performance. To modify the network to run on a custom dataset modify lib/datasets/larcvdataset.py