Skip to content

ZCls

Data

trick-data

zcls provides three ways to speed up data processing time

lmdb
prefetcher
zoom

lmdb

Use lmdb can speed up the read from disk to memory

First, use pnno to create .lmdb file from dataset
Second, see How to use LMDBDataset

prefetcher

Using prefetcher can speed up the reading from CPU memory to GPU memory

just open config item like this:

DATALOADER:
  ...
  ...
  ...
  PREFETCHER: True

zoom

Before training, the data can be prescaled. zcls provides a script tools/zoom.py to realize it