ImageNet-1k-256

A preprocessed dataset of ImageNet-1k for image generation.

• Each image is resized to 256x256.
• Resized images are processed by stabilityai/sd-vae-ft-mse to produce latents.

The scripts used for data preprocessing are from REPA & EDM.

The commands used to generate this dataset:

python dataset_tools.py convert --source=/path/to/imagenet/train --dest=/path/to/imagenet_256/images.zip --resolution=256x256 --transform=center-crop-dhariwal
python dataset_tools.py encode --source=/path/to/imagenet_256/images.zip --dest=/path/to/imagenet_256/vae-sd.zip

# Chunk .zip files into multiple smaller segments
split -b 40G --numeric-suffixes=1 --suffix-length=3 images.zip images.zip.chunk_
split -b 20G --numeric-suffixes=1 --suffix-length=3 vae-sd.zip vae-sd.zip.chunk_

The original ImageNet-1k dataset is supposed to be stored in /path/to/imagenet/train, with the following file structure:

/path/to/imagenet/train
  + ...
  + n12768682/
     + n12768682_9750.JPEG
     + n12768682_9768.JPEG
     ...
  + n12985857/
  + n12998815/
  ...
  + n15075141/

How to use

1. Download the dataset.

huggingface-cli download --local-dir /path/to/imagenet_256 --repo-type dataset Yuehao/imagenet-1k-256

2. Join chunks into .zip files:

cat images.zip.chunk_* > images.zip
cat vae-sd.zip.chunk_* > vae-sd.zip

3. Data loading example:

'''
reference:

https://github.com/NVlabs/edm2/blob/main/training/dataset.py#L165
https://github.com/sihyun-yu/REPA/blob/main/dataset.py#L17
'''

import os
import zipfile
import json
import numpy as np
import PIL.Image

data_dir = '/path/to/imagenet_256'
images_dir = os.path.join(data_dir, 'images.zip')
features_dir = os.path.join(data_dir, 'vae-sd.zip')

images_zipfile = zipfile.ZipFile(images_dir)
features_zipfile = zipfile.ZipFile(features_dir)

# Get all file names
image_fnames = sorted(fname for fname in set(images_zipfile.namelist()))
features_fnames = sorted(fname for fname in set(features_zipfile.namelist()))


## Load labels
with features_zipfile.open('dataset.json', 'r') as f:
  labels = json.load(f)['labels']
labels = dict(labels)
labels = [labels[fname.replace('\\', '/')] for fname in feature_fnames]
labels = np.array(labels)
# all labels
labels = labels.astype({1: np.int64, 2: np.float32}[labels.ndim])


## Load images
idx = 0
image_fname = image_fnames[idx]
feature_fname = feature_fnames[idx]

with images_zipfile.open(image_fname, 'r') as f:
  image = np.array(PIL.Image.open(f))

feature = np.load(features_zipfile.open(feature_fname, 'r'))

# RGB image, image latent, image class
# (image, feature, labels[idx])