DeepHGCal

Note: For the standalone version for particle track reconstruction (CMS Tracker), use the branch tntuples along with its specific instructions.

conda activate deephgcalenv
cd DeepHGCal/python
python standalone-models/tntuples.py configs/Jan19cfg.ini tntuples

If you would like to run other analyses for the CMS HGCal, follow the instructions given below.

Based on the DeepJetCore framework (https://github.com/DL4Jets/DeepJetCore) [CMS-AN-17-126] for HGCal reconstruction purposes.

The framework consists of two parts:

1. HGCal ntuple converter (Dependencies: root5.3/6)
2. DNN training/evaluation (Dependencies: DeepJetCore and all therein).

The DeepJetCore framework and the DeepHGCal framework should be checked out to the same parent directory. Before usage, always set up the environment by sourcing XXX_env.sh

Installation

git clone https://github.com/DL4Jets/DeepJetCore.git
git clone https://github.com/jkiesele/DeepHGCal.git
cd DeepHGCal
conda env create -f deephgcal.yml -n deephgcalenv
source lxplus_env.sh
pip install -r requirements.txt

• You will need the correct version of tensorflow, tensorflow-gpu(1.8.0) and cudnn in order to run on the CMG GPU.

conda install -c anaconda cudnn=7.1.2  # also installs cudatoolkit=9.0

• You will (probably) need to symlink (ln -s) your libstdc++.so.6 from libstdc++.so.6.0.19 to 6.0.25

unlink $CONDA_PREFIX/lib/libstdc++.so.6
ln -s $CONDA_PREFIX/lib/libstdc++.so.6.0.25 $CONDA_PREFIX/lib/libstdc++.so.6

Usage

The experiments are usually conducted in three steps:

1. Training
2. Testing (dumping of inference result somewhere on disk)
3. Plotting and analysis

Training

• When training for the first time, ensure that the from_scratch configuration value is set to 1.
• Check using the gpustat command and accordingly set the environment variable CUDA_VISIBLE_DEVICES to the available GPU ID.
• Check out off-the-shelf combinations of training configurations:

Path to Train File	Path to Config File	Config Name
bin/train/sparse_conv_clustering	configs/Jan19cfg.ini	single_neighbours
bin/train/sparse_conv_clustering	configs/Jan19cfg.ini	hidden_aggregators_plusmean
bin/train/sparse_conv_clustering	configs/Jan19cfg.ini	tntuples

cd DeepHGCal/python  # Training must be launched from this directory
python <path to train file> <path to config file> config_name

TrackingNTuples

• If you only want to use the repository for training on particle tracks, based on TrackingNtuples
• Refer to the Instructions

git clone https://github.com/SwapneelM/DeepHGCal
cd DeepHGCal 
conda env create -f deephgcal.yml -n deephgcalenv
conda activate deephgcalenv
pip install -r requirements.txt
cd python
python standalone-models/tntuples.py configs/Jan19cfg.ini tntuples

Testing

python <path to train file> <path to config file> config_name --test True

Plotting and analysis

python bin/plot/plot_file.py <path to config file> config_name

For clustering, the plot_file can be plot_inference_clustering.py

It will plot the resolution histogram as well as output mean and variance of resolution on stdout.

Issues

1. Tensorflow and CUDA compatibility issue: Check thread on Tensorflow's Github Issues for your specific version of libcublas.so

2. No module named <module_name>: This means you have not initialized the $PYTHONPATH, $LD_PRELOAD or $PATH and the script cannot find the modules you want to import. (Note that DeepHGCal/python must be placed on the $PYTHONPATH for the internal modules to be found)

3. KeyError: Ensure that you are using the correct configuration file with the model of your choice and it has all the necessary keys initialized.

4. PermissionError: [Errno 1] Operation not permitted: This is most likely a file-writing error and is expected for an off-the-shelf run of the model training. Modify the config to point to your own output directory where you do have write permission.