app.py

import torch
import gradio as gr
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image
from torchvision import transforms
from torchvision.transforms import InterpolationMode

# Device configuration
device = "cpu"

# Load processor (for text tokenization/decoding only)
processor = TrOCRProcessor.from_pretrained("microsoft/trocr-small-printed")

# Load and prepare the quantized model
model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-small-printed")
model = torch.quantization.quantize_dynamic(
    model, {torch.nn.Linear}, dtype=torch.qint8
)
model.load_state_dict(torch.load("best_model_int8.pt", map_location="cpu"), strict=False)
model.to(device)
model.eval()

# Define the EXACT same preprocessing used during training (INFERENCE version)
# Critical: Must match the training pipeline's resize method (LANCZOS interpolation)
inference_transform = transforms.Compose([
    # 1. Sharp resizing - same as training (LANCZOS preserves thin strokes)
    transforms.Resize((384, 384), interpolation=InterpolationMode.LANCZOS),
    # 2. Convert to tensor (range [0, 1])
    transforms.ToTensor(),
])

def predict(img: Image.Image):
    """
    Process image with training-matched preprocessing and run OCR inference.
    
    Args:
        img: PIL Image in RGB format
        
    Returns:
        Recognized text string
    """
    # Step 1: Ensure image is in RGB mode (consistent with training)
    if img.mode != 'RGB':
        img = img.convert('RGB')
    
    # Step 2: Apply the SAME transformation as in training
    # This gives us a tensor in [C, H, W] format, range [0, 1]
    pixel_values = inference_transform(img)
    
    # Step 3: Add batch dimension -> [1, C, H, W]
    pixel_values = pixel_values.unsqueeze(0)
    pixel_values = pixel_values.to(device)
    
    # Step 4: Run inference
    with torch.no_grad():
        generated_ids = model.generate(pixel_values)
    
    # Step 5: Decode the generated token IDs to text
    text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
    
    return text

# Create Gradio interface
gr.Interface(
    fn=predict,
    inputs=gr.Image(type="pil", label="Upload word image"),
    outputs=gr.Textbox(label="Recognized Text"),
    title="TrOCR OCR (CPU Optimized)",
    description="Fine-tuned TrOCR on IIIT-5K  | CPU inference"
).launch(share=True)