💮 Imagenetto

imagenetto is several small scale subset of 10 carefully chosen classes from ImageNet containing 3,000~5,000 training and validation images.

Quick usage

Download

You can download the dataset here: GitHub Release - Imagenetto.

• imagenetto_*.tar.gz contains gzipped files of this dataset with a 80/20 train/val split ratio.
• imagenetto_*_no_split.tar.gz contains gzipped files of the non-split version of this dataset.

Demo

Using the dataset

You can load the dataset with PyTorch and torchvision.

import os
import torch
import torchvision
from torchvision.transforms import transforms

# The root folder of the imagenetto dataset
dataset_root = "./imagenetto_a"

# Define a transform with random cropping and normalization
transform = {
    "train": transforms.Compose(
        [
            transforms.RandomResizedCrop(244),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ]
    ),
    "val": transforms.Compose(
        [
            transforms.Resize(256),
            transforms.CenterCrop(244),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ]
    ),
}

# Load the dataset with torchvision ImageFolder
image_datasets = {
    x: torchvision.datasets.ImageFolder(
        root=os.path.join(dataset_root, x), transform=transform[x]
    )
    for x in ["train", "val"]
}
# Create a dataloader iterative
data_loaders = {
    x: torch.utils.data.DataLoader(
        image_datasets[x], batch_size=4, shuffle=True
    )
    for x in ["train", "val"]
}
# Calculate the size of each split of the dataset
data_sizes = {x: len(image_datasets[x]) for x in ["train", "val"]}

...

Specifications

Labels

If you are using this dataset for tasks other than classification, use ImagenettoA, as it has some shared labels with the VOC dataset. Otherwise, use ImagenettoB.

ImagenettoA

Index	ImageNet Classes	ImageNet Labels	VOC Labels
1	n02109961	Eskimo dog	dog
2	n02123394	Persian cat	cat
3	n02690373	airliner	aeroplane
4	n02917067	bullet train	train
5	n03376595	folding chair	chair
6	n01833805	hummingbird	bird
7	n03782006	monitor	tvmonitor
8	n04146614	school bus	bus
9	n04273569	speedboat	boat
10	n04344873	studio couch	sofa

ImagenettoB

Index	ImageNet Classes	ImageNet Labels
1	n02092002	Scottish deerhound
2	n02134084	assault rifle
3	n07745940	bakery
4	n04389033	espresso
5	n04252077	ice bear
6	n04579432	ski
7	n02749479	snowmobile
8	n04228054	strawberry
9	n02776631	tank
10	n07920052	whistle

Reference ImageNet class - label list: aaronpolhamus/map_clsloc.txt.

Training / Validation

This dataset has a 80%/20% train/validation split.

Caveats

There's no bounding-box annotations available yet, which means that you won't be able to train detectors and such on this dataset. The purpose of this dataset is more of an evaluation dataset rather than a training dataset.

Clone with submodule

git clone --recurse-submodules -j8 git@github.com:daisylab-bit/imagenetto.git

Behind the scenes

• See: Behind the scenes.

Thanks

• ImageNet Downloader
• jfilter/split-folders
• fastai/imagenette

License

MIT.