Skip to content

Draft: Config for MLP compatible with hls4ml

Fotis Giasemis requested to merge dev-fgiasemi into main

Changes:

  • Disable the use of latex by default
  • Add config velo-query-long_h8-hls4ml based on velo-query-long_h8, which has a (3,8,8,8,3)-architecture MLP
  • Replace Tanh with ReLU activation, for compatibility with hls4ml
  • Disable layernorm, for compatibility with hls4ml, allowing to choose layernorm with a Boolean variable in the embedding config
  • Add batch script sbatch_a100_40_script.sh for running full_pipeline.ipynb on the cluster, while keeping track of the output from the .err/.out files
  • Create folder called artifact_archive, where important runs, reports and checkpoints can be archived
  • Run full pipeline with config velo-query-long_h8-hls4ml and save in artifact_archive

Merge request reports