ILSVRC/imagenet-1k · 2552th image has an enormous ICC profile (1.5MB)

Feb 12

Hello, I noticed that the 2552th image in the training set has an enormous ICC profile (1.5MB). This is problematic when modifying (map) the dataset because Hugging Face saves the image as a PNG which also contains the ICC profile and cannot be opened by PIL anymore.

>>> dataset = load_dataset('imagenet-1k')
>>> dataset['train'][2552]['image'].save('hello_kitty.png')
>>> Image.open('hello_kitty.png') 
...
ValueError: Decompressed Data Too Large

Therefore, if we map over the dataset (map) to modify the images (e.g. crop or resize) we get corrupted cache files.

def transform(row):
    row['image'] = row['image'].resize((128, 128))
    return row

temp = dataset['train'].map(transform)
temp[2552]  # break

mariosasko

Feb 13

Hi! You can avoid this error by using the fix from this SO thread. Or, delete the ICC profile from the metadata before saving an image:

def transform(row):
    row['image'] = row['image'].resize((128, 128))
    row["image"].info.pop("icc_profile", None)
    return row

francois-rozet

Feb 22

Thanks @mariosasko , your fix seems to work! An alternative is to transform the image into a NumPy array and then back into an image, which removes all metadata.

def transform(row):
    row['image'] = row['image'].resize((128, 128))
    row['image'] = Image.fromarray(np.asarray(row['image']))
    return row

francois-rozet changed discussion status to closed Feb 22