MLHEP2020: Sparsification competition

Please log in.

    • MLHEP 2020: Variational sparsification for PID classification

      In this challenge, we are asking you to train a classifier to identify the type of a particle. There are six particle types: electron, proton, muon, kaon, pion and ghost. Ghost is a particle with another type than the first five or detector noise.

      Your task is to achieve a high quality of classification but also sparsify your network as much as possible. You are provided with the LinearSVDO-layer realization that you should use in this challenge.

      In order to submit your solution please follow the instructions below:

      1. Register in CodaLab, with the same email you have on github.

      2. Fork baseline repository.

      3. Play with the architecture of the network in and play with the optimization routine sparse_particle_identification.ipynb to improve your score.

      4. Create zip-archive with weights of your network( and model architecture( 



    • MLHEP 2020: Sparsification competition. Evaluation

      The submission will be evaluated using this code.

      The formula of the score:

      score = max(min((1.2 - NLL) / 0.9, 1), 0) · log(1 + 50000 effecive_number_parameters),
      • NLL is a Negative Log-Likelihood of the classifier,
      • effecive_number_parameters is a number of parameters in your neural network. 


    • MLHEP 2020 PID SparseVD: Rules

      Zipped results must be submitted before the 2020-09-01 20:09:00+00:00. You may submit 15 submissions every day and no more than 1000 in total.

      You can’t use datasets other than provided. You can use any code you find on the Internet, just acknowledge it.


    • MLHEP 2020 PID SparseVD: Data

      Dataset is in csv-format and could be opened with `pandas.read_csv`. For an example of how to work with it take a look on the baseline.

    • Development

      starting_kit public_data
  • Make your submission using github

    ID Status Inputs