app.py

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageClassification
import gradio as gr
import pytesseract

def classify_meme(image: Image.Image):
    # OCR: extract text from image
    extracted_text = pytesseract.image_to_string(image)

    # Process image with SigLIP2 model
    inputs = processor(images=image, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predictions = {labels[i]: float(probs[0][i]) for i in range(len(labels))}

    return {
        "Predictions": predictions,
        "Extracted Text": extracted_text.strip()
    }

# Load model and processor from Hugging Face
model = AutoModelForImageClassification.from_pretrained("google/siglip2-base-patch16-naflex")
processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-naflex")
labels = model.config.id2label

def classify_meme(image: Image.Image):
    inputs = processor(images=image, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predictions = {labels[i]: float(probs[0][i]) for i in range(len(labels))}
    return predictions

# Gradio interface
demo = gr.Interface(
    fn=classify_meme,
    inputs=gr.Image(type="pil"),
    outputs=[
        gr.Label(num_top_classes=2, label="Predictions"),
        gr.Textbox(label="Extracted Text")
    ],
    title="Meme Classifier with OCR",
    description="Upload a meme to classify its sentiment and extract text using OCR."
)


if __name__ == "__main__":
    demo.launch(share = True)