Add Support for GNN
- Add configs for GNN trained at UCL.
- Add "track labels" (
Y_trk_train
in final output file) for GNN training. - Add option to choose compression for final output files in config (default to
None
). - Add option to choose float precision in final train files. (default to
float16
). I didn't add this as an arg to theTrainSampleWriter
class since it's already carried in the config. - Write jet label as (5,4,0) in a
flavour
field in the final train files. - Allow for non chunked storage of final train files.
Small things:
- Make sure scaling uses
float32
- Some logging improvements to keep track of long steps.
- Ensure out directory exists in resampling stage.
- Increase default chunk size to
1e6