datadings.sets.ImageNet21k_write module

usage: ImageNet21k_write.py [-h] [-o PATH] [-y] [-s] [-t 0-2]

[–compress [quality 0-100]] [–subsampling {444,422,420,440,411,Gray}] INDIR

Create ImageNet21k winter release data set files.

This tool will look for the following files in the input directory:

  • winter21_whole.tar.gz

See also

http://image-net.org/download https://github.com/Alibaba-MIIL/ImageNet21K

Note

Registration is required to download this dataset. Please visit the website to download it. If you experience issues downloading you may consider using bittorrent: https://academictorrents.com/details/8ec0d8df0fbb507594557bce993920442f4f6477

Important

For performance reasons samples are read in same order as they are stored in the source tar files. It is recommended to use the datadings-shuffle command to create a shuffled copy.

Important

Samples have the following keys:

  • "key"

  • "image"

  • "label"

  • "label_tree"

Positional arguments

INDIR Directory that contains dataset source files.

options:
-h, --help

show this help message and exit

-o PATH, --outdir PATH

Output directory. Defaults to indir.

-y, --no-confirm

Don’t require user interaction.

-s, --skip-verification

Skip verification of source files.

-t 0-2, –threads 0-2

Number of threads for conversion. 0 uses all available CPUs (default 1).

–compress [quality 0-100]

Use JPEG compression with optional quality. Default quality is 85. Big images are resized to roughly fit 500x375.

–subsampling {444,422,420,440,411,Gray}

Color subsampling factor used with compress option. 444 is forced for small images to preserve details.

datadings.sets.ImageNet21k_write.main()[source]
datadings.sets.ImageNet21k_write.write_sets(files, outdir, args)[source]
datadings.sets.ImageNet21k_write.yield_samples(infile)[source]