Summary
The docs/source/use_dataset.mdx documentation uses AutoFeatureExtractor from transformers in its image preprocessing example. For vision models, AutoImageProcessor is now the recommended API and better reflects current transformers conventions.
Problem
In the [Apply data augmentations] section of the documentation, the image augmentation example imports and uses AutoFeatureExtractor:
from transformers import AutoFeatureExtractor
from datasets import load_dataset, Image
feature_extractor = AutoFeatureExtractor.from_pretrained("google/vit-base-patch16-224-in21k")
dataset = load_dataset("AI-Lab-Makerere/beans", split="train")Fix
Replace AutoFeatureExtractor with AutoImageProcessor in the code snippet:
from transformers import AutoImageProcessor
from datasets import load_dataset, Image
image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
dataset = load_dataset("AI-Lab-Makerere/beans", split="train")This fix is consistent with the direction taken in the transformers library itself — see [PR #20111] and [PR #20501], where vision feature extractor references were systematically replaced with image processors across the transformers codebase and documentation.
Note: the variable name feature_extractor can be kept as-is for a minimal diff, but renaming it to image_processor better reflects the updated class name and matches current transformers conventions.
References
Summary
The
docs/source/use_dataset.mdxdocumentation usesAutoFeatureExtractorfromtransformersin its image preprocessing example. For vision models,AutoImageProcessoris now the recommended API and better reflects currenttransformersconventions.Problem
In the [Apply data augmentations] section of the documentation, the image augmentation example imports and uses
AutoFeatureExtractor:Fix
Replace
AutoFeatureExtractorwithAutoImageProcessorin the code snippet:This fix is consistent with the direction taken in the
transformerslibrary itself — see [PR #20111] and [PR #20501], where vision feature extractor references were systematically replaced with image processors across thetransformerscodebase and documentation.References
AutoImageProcessordocsAutoImageProcessoras the dedicated class for vision modelstransformersdocumentation. Thedatasetsdocs were not in scope for that PR, making this fix a natural follow-up.