Datasets:

---

ILSVRC

/

imagenet-1k

like 748

Follow

Large Scale Visual Recognition Challenge 91

Yes you can, e.g. this will save all the images in a directory of your choice

from pathlib import Path
from datasets import load_dataset, Image

# Login using e.g. `huggingface-cli login` to access this dataset
ds = load_dataset("ILSVRC/imagenet-1k", split="train")
ds = ds.cast_column("image", Image(decode=False))

output_dir = Path("path/to/output/dir")

for image in ds["image"]:
    (output_dir / image["path"]).write_bytes(image["bytes"])

Note that we use Image(decode=False) do access the image path name and bytes instead of getting PIL.Image objects

a simple script for extracting

import os
from datasets import load_dataset
from tqdm import tqdm

dataset = load_dataset("ILSVRC/imagenet-1k")

label_names = dataset["train"].features["label"].names

def export_split(split_name, output_root):
    split = dataset[split_name]
    
    for idx, sample in enumerate(tqdm(split)):
        image = sample["image"]        # PIL Image
        label = sample["label"]        # int
        
        class_name = label_names[label]
        
        class_dir = os.path.join(output_root, split_name, class_name)
        os.makedirs(class_dir, exist_ok=True)
        
        save_path = os.path.join(class_dir, f"{idx}.jpg")
        image.save(save_path)

export_split("train", "./imagenet")

export_split("validation", "./imagenet")