What is the meta.bin file used by the ImageNet dataset?

[pmeier]

Comment from @kanonjz in #1457

I downloaded imagenet myself and used parse_val_archive to prepare the folders, but got an error below. What is the meta.bin? I didn't find it in the imagenet.

The meta file meta.bin is not present in the root directory or is corrupted. " "This file is automatically created by the ImageNet dataset.

[pmeier]

meta.bin is a binary file which holds 2 things:

1. A dict with a mapping from the WordNet IDs (wnids) to human-readable classes.
2. A list of the wnid of each validation image.

The meta.bin file is created by executing the parse_devkit_archive function:

vision/torchvision/datasets/imagenet.py

Lines 105 to 113 in 598b61d

	def parse_devkit_archive(root, file=None):
	"""Parse the devkit archive of the ImageNet2012 classification dataset and save
	the meta information in a binary file.
	Args:
	root (str): Root directory containing the devkit archive
	file (str, optional): Name of devkit archive. Defaults to
	'ILSVRC2012_devkit_t12.tar.gz'
	"""

Note that this requires scipy to be installed.

---

@kanonjz To resolve your error you have 2 options:

1. Download the devkit archive and parse it with the function mentioned above to generate the meta.bin file.
2. The parse_val_archive function only tries to load the wnids from the meta.bin file if none are given to it. If you have a list, try passing it like this:

parse_val_archive(..., wnids=my_wnid_list)

[kanonjz]

Thanks!!

[pmeier]

@kanonjz Is your problem resolved? If yes, I'm going to close this issue.

[kanonjz]

@pmeier Yes!

[fmassa]

Thanks @pmeier for helping out with this issue!