ClusterFinder

Project aiming to find clusters in data streaming from CMS experiment CERN

To produced data for 1000 events, use the command:

python3 data_generator.py 1000

Starting the Setup:

Currently, we are using the existing setup at BASE_PATH mentioned below. We can run the notebooks and models present here using the information below.

Using [email protected]

conda activate ml-env
export BASE_PATH=/nfs_scratch/hsharma/MachineLearning/ClusterFinder
jupyter notebook list

if no notebook is already running, then start the jupyter notebook server using the command below:

cd $BASE_PATH
nohup jupyter notebook --no-browser --port 3000
jupyter notebook list

Use the command below to connect to the jupyter notebook locally in your machine. If the ports are busy, then try using a different set of port values."

ssh -L 1234:localhost:9999 [email protected] ssh -L 9999:localhost:3000 -N cmsgpu01

To shut down any jupyter notebook server running on a specific port, run the following command in cmsgpu01 machine:

# For notebook running in background on port 3000
fuser -k 3000/tcp

Using Docker:

Coming Soon

HLS4ML Conversion

For performing your experiment with different settings of HLS4ML, duplicate the HLS4ML_Interface_Template and run the notebook accordingly. The project output folder can be set by setting the desired folder name in cfg['OutputDir'] in hls4ml config part.

The list of models trained, along with their specifications can be found in models.json. The JSON is also loaded in the Template for anyone to look at.

Once the Vivado project folder is produced using the interface notebook, it can be synthesized using the command below:

nohup vivado_hls -f build_prj.tcl "reset=1 synth=1 csim=0 cosim=0 validation=0 export=0 vsynth=0" > vhls_synth.out &

Note:

• The current template uses io_stream method for passing data between layers. We must switch to io_parallel.

For more information on the the documentation and development of HLS4ML visit: https://fastmachinelearning.org/hls4ml

Future works:

• Build a rectangular I/P network for the working problem statement
• From Classification to Regression Task
• Regenerate i/p - o/p pairs for new task problem