examples/ort_usage_example.py

#!/usr/bin/env python3
"""
granite-docling ONNX Usage Example with ONNX Runtime
Demonstrates how to use the converted granite-docling model for document processing
"""

import onnxruntime as ort
import numpy as np
from PIL import Image
import json

def load_granite_docling_onnx(model_path: str):
    """Load granite-docling ONNX model"""
    print(f"Loading granite-docling ONNX model from: {model_path}")

    session = ort.InferenceSession(model_path)

    # Print model information
    print("Model Information:")
    print(f"  Inputs:")
    for inp in session.get_inputs():
        print(f"    {inp.name}: {inp.shape} ({inp.type})")

    print(f"  Outputs:")
    for out in session.get_outputs():
        print(f"    {out.name}: {out.shape} ({out.type})")

    return session

def preprocess_document_image(image_path: str) -> np.ndarray:
    """Preprocess document image for granite-docling inference"""

    # Load and resize image to 512x512 (SigLIP2 requirement)
    image = Image.open(image_path).convert('RGB')
    image = image.resize((512, 512))

    # Convert to numpy array and normalize
    pixel_values = np.array(image).astype(np.float32) / 255.0

    # Normalize using SigLIP2 parameters (from preprocessor_config.json)
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.225])

    pixel_values = (pixel_values - mean) / std

    # Reshape to [batch_size, channels, height, width]
    pixel_values = pixel_values.transpose(2, 0, 1)  # HWC -> CHW
    pixel_values = pixel_values[np.newaxis, :]      # Add batch dimension

    return pixel_values

def create_text_inputs(prompt: str = "Convert this document to DocTags:") -> tuple:
    """Create text inputs for granite-docling"""

    # Simple tokenization (in practice, use proper tokenizer)
    # This is a simplified example - use actual granite-docling tokenizer
    tokens = [1] + [i for i in range(2, len(prompt.split()) + 2)] + [2]  # Simple token mapping

    input_ids = np.array([tokens], dtype=np.int64)
    attention_mask = np.ones((1, len(tokens)), dtype=np.int64)

    return input_ids, attention_mask

def run_granite_docling_inference(session, image_path: str):
    """Run complete granite-docling inference"""

    print(f"Processing document: {image_path}")

    # Prepare inputs
    pixel_values = preprocess_document_image(image_path)
    input_ids, attention_mask = create_text_inputs()

    print(f"Input shapes:")
    print(f"  pixel_values: {pixel_values.shape}")
    print(f"  input_ids: {input_ids.shape}")
    print(f"  attention_mask: {attention_mask.shape}")

    # Run inference
    outputs = session.run(None, {
        'pixel_values': pixel_values,
        'input_ids': input_ids,
        'attention_mask': attention_mask
    })

    logits = outputs[0]
    print(f"Output logits shape: {logits.shape}")

    # Decode logits to tokens (simplified)
    predicted_tokens = np.argmax(logits, axis=-1)
    print(f"Predicted tokens shape: {predicted_tokens.shape}")

    # In practice, decode tokens to DocTags markup using proper tokenizer
    print("✅ Inference completed successfully")

    return predicted_tokens

def main():
    """Main example usage"""

    model_path = "model.onnx"  # Path to downloaded ONNX model

    try:
        # Load model
        session = load_granite_docling_onnx(model_path)

        # Run inference on example document
        # (Replace with actual document image path)
        image_path = "example_document.png"

        if os.path.exists(image_path):
            result = run_granite_docling_inference(session, image_path)
            print("✅ granite-docling ONNX inference successful!")
        else:
            print("⚠️ No example document provided")
            print("  Create a test document image to run inference")

    except Exception as e:
        print(f"❌ Example failed: {e}")
        import traceback
        traceback.print_exc()

if __name__ == "__main__":
    import os
    main()