ViT-Tiny Classifier for RVL-CDIP Document Classification (Distilled)
This model is a compressed Vision Transformer (ViT-Tiny) trained using knowledge distillation from DiT-Large on the RVL-CDIP dataset for document image classification. This model was developed as part of a research internship at the Laboratory of Complex Systems, Ecole Centrale Casablanca
Model Details
- Student Model: ViT-Tiny (Vision Transformer)
- Teacher Model: microsoft/dit-large-finetuned-rvlcdip
- Training Method: Knowledge Distillation
- Parameters: ~5.5M (55x smaller than teacher)
- Dataset: RVL-CDIP (320k document images, 16 classes)
- Task: Document Image Classification
- Accuracy: 0.9210
- Compression Ratio: ~55x parameter reduction from teacher model
Document Classes
The model classifies documents into 16 categories:
- letter - Personal or business correspondence
- form - Structured forms and applications
- email - Email communications
- handwritten - Handwritten documents
- advertisement - Marketing materials and ads
- scientific_report - Research reports and studies
- scientific_publication - Academic papers and journals
- specification - Technical specifications
- file_folder - File folders and organizational documents
- news_article - News articles and press releases
- budget - Financial budgets and planning documents
- invoice - Bills and invoices
- presentation - Presentation slides
- questionnaire - Surveys and questionnaires
- resume - CVs and resumes
- memo - Internal memos and notices
Usage
from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
# Load model
processor = AutoImageProcessor.from_pretrained("HAMMALE/vit-tiny-classifier-rvlcdip")
model = AutoModelForImageClassification.from_pretrained("HAMMALE/vit-tiny-classifier-rvlcdip")
# Load and classify an image
image = Image.open("path_to_your_document_image.jpg")
inputs = processor(image, return_tensors="pt")
# Get predictions
outputs = model(**inputs)
predicted_class_id = outputs.logits.argmax(-1).item()
# Get class names
class_names = [
"letter", "form", "email", "handwritten", "advertisement",
"scientific_report", "scientific_publication", "specification",
"file_folder", "news_article", "budget", "invoice",
"presentation", "questionnaire", "resume", "memo"
]
predicted_class = class_names[predicted_class_id]
print("Predicted class:", predicted_class)
Performance
| Metric | Value |
|---|---|
| Accuracy | 0.9210 |
| Parameters | ~5.5M |
| Model Size | ~22 MB |
| Input Size | 224x224 pixels |
Training Details
- Student Architecture: Vision Transformer (ViT-Tiny)
- Teacher Model: microsoft/dit-large-finetuned-rvlcdip
- Distillation Method: Knowledge Distillation
- Input Resolution: 224x224
- Preprocessing: Standard ImageNet normalization
- Framework: Transformers/PyTorch
- Distillation Benefits: Maintains high accuracy with 55x fewer parameters
Dataset
The RVL-CDIP (Ryerson Vision Lab Complex Document Information Processing) dataset contains:
- 400,000 grayscale document images
- 16 document categories
- Images collected from truth tobacco industry documents
- Standard train/validation/test splits
Citation
@misc{hammale2025vit_tiny_rvlcdip_distilled,
title={ViT-Tiny Classifier for RVL-CDIP Document Classification (Distilled)},
author={Hammale, Mourad},
year={2025},
howpublished={\url{https://huggingface.co/HAMMALE/vit-tiny-classifier-rvlcdip}},
note={Knowledge distilled from microsoft/dit-large-finetuned-rvlcdip}
}
Acknowledgments
This model was created by HAMMALE (Mourad) through knowledge distillation from the larger DiT-Large model (microsoft/dit-large-finetuned-rvlcdip), achieving significant compression while maintaining competitive performance for document classification tasks.
License
This model is released under the Apache 2.0 license.