# Iterating over Image feature columns is extremely slow

**URL:** <https://discuss.huggingface.co/t/iterating-over-image-feature-columns-is-extremely-slow/149854>\
**Category:** 🤗Datasets\
**Created:** [April 10, 2025, 7:07pm UTC](https://discuss.huggingface.co/t/iterating-over-image-feature-columns-is-extremely-slow/149854 "2025-04-10T19:07:59Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![sohamcheeseburger](https://avatars.discourse-cdn.com/v4/letter/s/7bcc69/32.png) [@sohamcheeseburger](https://discuss.huggingface.co/u/sohamcheeseburger)\
**Post date:** [April 10, 2025, 7:07pm UTC](https://discuss.huggingface.co/t/iterating-over-image-feature-columns-is-extremely-slow/149854/1 "2025-04-10T19:07:59Z")

</div>

We are trying to load datasets where the image column stores `PIL.PngImagePlugin.PngImageFile` images. However, iterating over these datasets is extremely slow.  
What I have found:

1. It is the presence of the image column that causes the slowdown. Removing the column from the dataset results in blazingly fast (as expected) times
2. It is ~2x faster to iterate when the column contains a single image as opposed to a list of images i.e., the feature is a Sequence of Image objects. We often need multiple images per sample, so we need to work with a list of images
3. It is ~17x faster to store paths to PNG files and load them using `PIL.Image.open`, as opposed to iterating over a `Dataset` with an Image column, and ~30x faster compared to `Sequence` of `Image`s. See a simple script below with an openly available dataset.

It would be great to understand the standard practices for storing and loading multimodal datasets (image + text).

[Load image data](https://huggingface.co/docs/datasets/en/image_load) seems a bit underdeveloped? (e.g., `dataset.decode` only works with `IterableDataset`, but it’s not clear from the doc)

Thanks!

```python
from datasets import load_dataset, load_from_disk
from PIL import Image
from pathlib import Path

ds = load_dataset("getomni-ai/ocr-benchmark")

for idx, sample in enumerate(ds["test"]):
    image = sample["image"]
    image.save(f"/tmp/ds_files/images/image_{idx}.png")

ds.save_to_disk("/tmp/ds_columns")

# Remove the 'image' column
ds["test"] = ds["test"].remove_columns(["image"])

# Create image paths for each sample
image_paths = [f"images/image_{idx}.png" for idx in range(len(ds["test"]))]

# Add the 'image_path' column to the dataset
ds["test"] = ds["test"].add_column("image_path", image_paths)

# Save the updated dataset
ds.save_to_disk("/tmp/ds_files")
files_path = Path("/tmp/ds_files")
column_path = Path("/tmp/ds_columns")

# load and benchmark
ds_file = load_from_disk(files_path)
ds_column = load_from_disk(column_path)
import time

images_files = []
start = time.time()
for idx in range(len(ds_file["test"])):
    image_path = files_path / ds_file["test"][idx]["image_path"]
    image = Image.open(image_path)
    images_files.append(image)
end = time.time()
print(f"Time taken to load images from files: {end - start} seconds")

# Time taken to load images from files: 1.2364635467529297 seconds

images_column = []
start = time.time()
for idx in range(len(ds_column["test"])):
    images_column.append(ds_column["test"][idx]["image"])
end = time.time()
print(f"Time taken to load images from columns: {end - start} seconds")

# Time taken to load images from columns: 20.49347186088562 seconds

```

---

<div class="post-metadata">

**Author:** ![lhoestq](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/lhoestq/32/52888_2.png) [@lhoestq](https://discuss.huggingface.co/u/lhoestq)\
**Post date:** [April 11, 2025, 1:20pm UTC](https://discuss.huggingface.co/t/iterating-over-image-feature-columns-is-extremely-slow/149854/2 "2025-04-11T13:20:21Z")

</div>

Hi ! The `Image()` decoding simply uses PIL.Image.open and also calls `image.load().

This step decodes the image and loads it in memory. This step is a bit slower than calling PIL.Image.open alone but it ensures that the python shell doesn’t crash because of too many open files (issue , here: [[image feature] Too many files open error when image feature is returned as a path · Issue #3985 · huggingface/datasets · GitHub](https://github.com/huggingface/datasets/issues/3985))

This part of the codebase is open to contributions though, if you have an idea of how to improve this !

Moreover `.decode()`happens to be a very recent addition, so it’s indeed not yet available for `Dataset` objects (it would be cool though). Anyway thanks for the valuable feedback 🙂

---

<div class="post-metadata">

**Author:** ![sohamcheeseburger](https://avatars.discourse-cdn.com/v4/letter/s/7bcc69/32.png) [@sohamcheeseburger](https://discuss.huggingface.co/u/sohamcheeseburger)\
**Post date:** [April 11, 2025, 4:02pm UTC](https://discuss.huggingface.co/t/iterating-over-image-feature-columns-is-extremely-slow/149854/3 "2025-04-11T16:02:30Z")

</div>

Thanks for the reply, that’s an important point I missed! I’m quite new to the world of images so I don’t know how well I can contribute.  
Wondering if it’s possible to bring more speedup to image loading. For example, loading pixel values from bytes in a memory-mapped file (e.g., using `np.memmap`, `np.frombuffer`) and using `Image.fromarray` (if necessary) is still significantly faster than either methods