Skip to content

speedup data loading for HOP

Franck Galpin requested to merge Franck/VVCSoftware_VTM:HOP_training_speedup into VTM-11.0_nnvc
  • the file to read from are open only 1 time instead of being open at each iteration
  • a bug in some version of numpy make fromfile have a large read ahead which increase a lot IO during the training: replaced by low level reading

On our cluster, we observed speedup between 2 and 5 depending on the machine IO load.

Merge request reports