tuanhl's picture
Upload folder using huggingface_hub
d903901 verified
|
Raw History Blame Contribute Delete
4.77 kB
metadata
language: en
license: other
license_name: imagenet
license_link: https://www.image-net.org/download-images.php
pretty_name: ImageNet_10cls_small
task_categories:
  - image-classification
tags:
  - tinyML
  - imagenet
  - open-set-classification
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: calib
        path: calib_images/**
      - split: test
        path: test_images/**
    features:
      - name: image
        dtype: image
      - name: label
        dtype:
          class_label:
            names:
              '0': pekinese
              '1': tabby
              '2': african elephant
              '3': sports car
              '4': airliner
              '5': coffee mug
              '6': pizza
              '7': castle
              '8': electric guitar
              '9': cucumber
              '10': others
dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype:
        class_label:
          names:
            '0': pekinese
            '1': tabby
            '2': african elephant
            '3': sports car
            '4': airliner
            '5': coffee mug
            '6': pizza
            '7': castle
            '8': electric guitar
            '9': cucumber
            '10': others

Dataset Card for ImageNet_10cls_small

This dataset is a compact downstream variant of ILSVRC/ImageNet, curated for the MobileNetV3-Small 0.35 open-set classifier in TinyML applications. It uses 10 selected ImageNet classes and an others class.

Dataset Description

Dataset Structure

Split Images Directory
calib 500 calib_images/
test 1 test_images/

The calib split contains 20 training images per known class and up to 300 training images labeled others. The test split contains the simulator inputs selected at indices [255] from the locally prepared open-set validation dataset (seed 42). These few test images are for checking compilation and simulator execution; their metrics do not represent a full validation benchmark.

Full ImageNet training and validation datasets are not included. Obtain them separately from the upstream source for QAT training or validation benchmarking. The class-count component of the dataset name describes the known classes; others is an additional output label. Different class selections require their own dataset name and class mapping.

Each image directory includes a metadata.jsonl file containing file_name and label. The root dataset.json also records the sample order, original indices, source WNIDs, sampling settings and SHA-256 image checksums.

Images retain their original bytes and dimensions. For the target classifier, resize the shorter side to 256 pixels using bilinear interpolation, center crop to 224 × 224, and apply ImageNet normalization. The Quartz example feeds raw 0–255 RGB to its inserted input standardizer, which performs the normalization.

Class Labels

Label Class
0 pekinese
1 tabby
2 african elephant
3 sports car
4 airliner
5 coffee mug
6 pizza
7 castle
8 electric guitar
9 cucumber
10 others

Licensing and Attribution Information

This subset inherits the ImageNet Terms of Access. ImageNet provides access for non-commercial research and educational purposes.

  1. Source images: Copyright remains with the respective image owners. ImageNet and the curators of this subset do not own those image copyrights. This subset grants no additional rights to the images.
  2. Labels and metadata: Labels derive from ILSVRC/ImageNet. The manifest records the subset selection and remaps non-selected classes to others; it does not replace or relax the upstream terms.

Users must comply with the upstream terms and applicable image copyrights, including the requirement that research associates and colleagues agree to those terms before receiving access. This card does not grant unrestricted redistribution rights.

Upstream Attribution

Please acknowledge this derivative dataset and the original ImageNet dataset.

Citation

@inproceedings{deng2009imagenet,
  title={ImageNet: A Large-Scale Hierarchical Image Database},
  author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
  booktitle={2009 IEEE Conference on Computer Vision and Pattern Recognition},
  pages={248--255},
  year={2009}
}