imagenet-classes 0.2.0

A Python package for managing and retrieving ImageNet-1k mappings among integer class IDs, string class IDs, and human-readable class names.

imagenet-classes

A Python package for managing and retrieving ImageNet-1k (ImageNet2012) class names and 1k/21k mappings.

v0.2.0 — breaking rename: The package was previously distributed as class_mapping with imports from class_mapping. Starting in 0.2.0 the package is renamed to imagenet_classes. All imports must be updated accordingly (see Migration from v0.1.x).

Installation

pip install imagenet-classes

Or from source:

git clone https://github.com/gonikisgo/imagenet-classes.git
cd imagenet-classes
pip install -e .

Quick Start

import imagenet_classes as ic

# Get the raw synset label string for class 388
print(ic.get_1k_class_name(388))
# → 'giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca'

# Get a single curated human-readable name for class 388
print(ic.get_1k_clean_name(388))
# → 'giant panda'

# 1k/21k key conversions
print(ic.imagenet1k_to_21k(388))             # → 'n02510455'
print(ic.imagenet21k_to_1k('n02510455'))     # → 388

Class name methods

There are two methods for retrieving human-readable names. They differ in the source data and the level of curation:

get_1k_class_name(key: int) -> str | None

Returns the original ImageNet-1k label string exactly as it appears in the dataset annotation — a comma-separated list of WordNet synonyms for the synset. Many entries contain five or more synonyms and a scientific name.

ic.get_1k_class_name(0)   # → 'tench, Tinca tinca'
ic.get_1k_class_name(2)   # → 'great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias'

Use this when you need the canonical ImageNet label as-is (e.g. for reproducibility with prior work that used the original synset strings).

get_1k_clean_name(key: int) -> str | None

Returns a single, manually curated human-readable name — one short English phrase per class, with no synonyms or scientific names.

ic.get_1k_clean_name(0)   # → 'tench'
ic.get_1k_clean_name(2)   # → 'great white shark'

Use this when displaying class names in a UI or feeding them to a model as plain text, where a clean, unambiguous single name is preferable.

GPT descriptions and guidelines

Each class has a GPT-generated textual description and classification guideline based on the class content.

ic.get_gpt_class_description(0)
# → 'A thick-bodied freshwater fish with a smooth-looking body covered in very
#    small scales, usually colored olive-green, bronze, or brownish gold...'

ic.get_gpt_class_guidelines(0)
# → 'Classify as tench when the fish is deep-bodied and olive or bronze colored,
#    with tiny scales giving a slick matte appearance...'

1k/21k mappings

ImageNet-1k ↔ ImageNet-21k

Each ImageNet-1k integer label (0–999) corresponds to a WordNet synset key (e.g. 'n01440764') that is also used in ImageNet-21k.

ic.imagenet1k_to_21k(0)              # → 'n01440764'
ic.imagenet21k_to_1k('n01440764')    # → 0

Validation image → ImageNet-21k key

Maps a validation split image filename to its ImageNet-21k class index.

ic.val_image_to_21k_key('ILSVRC2012_val_00015416.JPEG')   # → 13

Using ClassDictionary directly

All module-level functions above delegate to a shared ClassDictionary instance. You can also instantiate ClassDictionary directly, which is useful when you want to swap in a different curated names file:

from imagenet_classes import ClassDictionary

cd = ClassDictionary(clean_names_json_filename='my_clean_names.json')
print(cd.get_1k_clean_name(0))

ClassDictionary also provides two static utilities for building label ↔ name dictionaries from an arbitrary class list:

classes = ['tench', 'goldfish', 'great white shark']

ClassDictionary.create_label_to_name_dict(classes)
# → {0: 'tench', 1: 'goldfish', 2: 'great white shark'}

ClassDictionary.create_name_to_label_dict(classes)
# → {'tench': 0, 'goldfish': 1, 'great white shark': 2}

Migration from v0.1.x

Before (v0.1.x)	After (v0.2.0+)
from class_mapping import ClassDictionary	from imagenet_classes import ClassDictionary or import imagenet_classes as ic
class_dict.get_class_name(key)	ic.get_1k_class_name(key)
class_dict.get_custom_class_name(key)	ic.get_1k_clean_name(key)
class_dict.get_class_1k(key)	ic.imagenet21k_to_1k(key)

If you previously instantiated ClassDictionary() yourself, you can drop it and use the module-level functions (ic.get_1k_class_name, ic.get_1k_clean_name, etc.) — they delegate to a shared instance internally. The only reason to keep using ClassDictionary directly is if you need to supply a custom clean names file:

from imagenet_classes import ClassDictionary

cd = ClassDictionary(clean_names_json_filename='/path/to/my_clean_names.json')
cd.get_1k_clean_name(0)   # uses your file instead of the bundled one

The file must be a plain JSON list of 1000 strings — no keys, no mapping, just names in label order (index 0 = label 0, index 1 = label 1, …):

["tench", "goldfish", "great white shark", ...]

License

MIT — see LICENSE.