app.py

from fastapi import FastAPI
from pydantic import BaseModel
import torch
from transformers import BertTokenizer, BertForSequenceClassification
import os
from fastapi.middleware.cors import CORSMiddleware
import logging
from contextlib import asynccontextmanager # Import asynccontextmanager

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Global variables for model and tokenizer
tokenizer = None
model = None
device = None

# --- Lifespan Event Handler ---
@asynccontextmanager
async def lifespan(app: FastAPI):
    global tokenizer, model, device
    logger.info("Starting model loading and quantization process (Lifespan event)...")
    try:
        # --- CHANGE STARTS HERE ---
        # Using the smallest available Amharic BERT model for extreme resource efficiency.
        # This model has only 4.18 Million parameters.
        model_name_or_path = "rasyosef/bert-tiny-amharic"
        # --- CHANGE ENDS HERE ---

        logger.info(f"Attempting to load tokenizer from: {model_name_or_path}")
        tokenizer = BertTokenizer.from_pretrained(model_name_or_path)
        logger.info("Tokenizer loaded successfully.")

        logger.info(f"Attempting to load model from: {model_name_or_path}")
        # Load the model
        model = BertForSequenceClassification.from_pretrained(model_name_or_path)
        model.eval() # Set model to evaluation mode
        logger.info("Model loaded successfully.")

        # --- Apply Dynamic Quantization ---
        logger.info("Applying dynamic quantization to the model...")
        model = torch.quantization.quantize_dynamic(
            model, {torch.nn.Linear}, dtype=torch.qint8
        )
        logger.info("Dynamic quantization applied successfully.")
        # --- End Dynamic Quantization ---

        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        model.to(device) # Move quantized model to device
        logger.info(f"Model moved to device: {device}")
        logger.info("All components loaded, quantized, and ready for serving.")

    except Exception as e:
        logger.error(f"Critical Error during model loading or quantization: {e}", exc_info=True)
        # For a truly free deployment, if essential model loading fails,
        # it might be better to let the app crash to avoid silent failures.
        # However, for debugging purposes, we'll keep 'pass' for now.
        pass

    yield # This indicates that the startup logic is complete and the app can serve requests

    # --- Shutdown logic (optional) ---
    logger.info("Shutting down application (Lifespan event)...")
    if torch.cuda.is_available():
        torch.cuda.empty_cache()


# Initialize FastAPI app with lifespan
app = FastAPI(lifespan=lifespan)

# --- CORS Configuration ---
origins = [
    "http://localhost",
    "http://localhost:3000", # We Allow requests from our React development server
    "https://natishanau-amharic-fake-news-frontend.vercel.app", # Your Vercel frontend domain
    "https://natishanau-amharic-fake-news-backend.hf.space", # Allow self-access if needed
]

app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# --- Health Check Endpoint ---
@app.get("/")
async def health_check():
    model_status = "loaded and quantized" if model is not None and tokenizer is not None else "loading, quantizing, or failed"
    return {"status": "ok", "message": "FastAPI backend is running", "model_status": model_status}

# Define request body structure
class TextInput(BaseModel):
    text: str

@app.post("/predict/")
async def predict(input: TextInput):
    """
    Predicts whether the given Amharic text is 'Real News' or 'Fake News'.
    """
    if model is None or tokenizer is None:
        logger.warning("Prediction request received but model is not yet loaded or failed to load.")
        return {"error": "Model not loaded yet. Please try again in a moment.", "status_code": 503}

    text = input.text
    
    if not text.strip():
        return {"text": text, "prediction": "Invalid Input", "details": "Text input cannot be empty."}

    try:
        inputs = tokenizer(
            text,
            return_tensors="pt",
            truncation=True,
            padding=True,
            max_length=512
        ).to(device)

        with torch.no_grad():
            outputs = model(**inputs)

        logits = outputs.logits
        probabilities = torch.softmax(logits, dim=-1)
        
        predicted_class_idx = torch.argmax(logits, dim=-1).item()
        confidence = probabilities[0][predicted_class_idx].item()
        
        if predicted_class_idx == 0:
            prediction_label = "Real News"
        else:
            prediction_label = "Fake News"

        return {
            "original_text": text,
            "prediction": prediction_label,
            "confidence": round(confidence, 4)
        }
    except Exception as e:
        logger.error(f"Error during prediction: {e}", exc_info=True)
        return {"error": f"An error occurred during prediction: {e}", "status_code": 500}