---
license: apache-2.0
pipeline_tag: image-classification
tags:
- image-classification
- multi-label-classification
- onnx
- openvino
- pdf
- document-understanding
- rag
datasets:
- Wikit/PdfVisClassif
---

# PDF Page Classifier

Multi-label classifier for PDF page images. Determines whether a PDF page
requires image embedding (vs. text-only) in RAG pipelines.

Backbone: EfficientNet-Lite0. Exported to ONNX and OpenVINO INT8 via
Quantization-Aware Training (QAT). **No PyTorch required at inference time.**

## Classes

- `Complex Table`
- `Simple Table`
- `Visual - Essential`
- `Visual - Supportive`

Pages matching any of the following classes should trigger image embedding:

- `Complex Table`
- `Visual - Essential`

Default threshold: `0.5`

## Usage

### With [chunknorris](https://github.com/wikit-ai/chunknorris) (recommended)

```bash
pip install "chunknorris[ml-onnx]"       # ONNX backend
pip install "chunknorris[ml-openvino]"   # OpenVINO INT8, fastest on CPU
```

```python
from chunknorris.ml import load_classifier

clf = load_classifier("Wikit/pdf-pages-classifier")   # auto-selects best available backend
result = clf.predict("page.png")
# {"needs_image_embedding": True, "predicted_classes": [...], "probabilities": {...}}
```

### Standalone (no chunknorris)

```bash
git clone https://huggingface.co/Wikit/pdf-pages-classifier
cd pdf-pages-classifier
pip install onnxruntime Pillow numpy   # or: openvino Pillow numpy
```

```python
from classifiers import load_classifier

clf = load_classifier(".")            # auto-selects available backend
result = clf.predict("page.png")
```

## Files

| File | Format | Notes |
|------|--------|-------|
| `model.onnx` | ONNX FP32 | Cross-platform CPU/GPU inference |
| `openvino_model.xml/.bin` | OpenVINO INT8 | Fastest CPU inference (QAT) |
| `pytorch_model.bin` | PyTorch | Raw checkpoint; requires `torch` + `timm` |
| `config.json` | JSON | Preprocessing config and class names |
| `classifiers/` | Python | Standalone inference scripts (no chunknorris needed) |

## Dataset

Trained on [Wikit/PdfVisClassif](https://huggingface.co/datasets/Wikit/PdfVisClassif).