app.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
import torch
import torch.nn as nn
import numpy as np
from typing import Optional, List
import time
from datetime import datetime, timezone
import os
import warnings
from huggingface_hub import hf_hub_download
from contextlib import asynccontextmanager
import uvicorn
from dotenv import load_dotenv
import shutil
import joblib
from pathlib import Path
from transformers import BertTokenizer, BertModel
from utils.model_classes import MHSA_GRU, MultiHeadSelfAttention

load_dotenv()
warnings.filterwarnings('ignore')

# ========================= CONFIGURATION =========================
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

API_VERSION = "1.0.0"
MODEL_VERSION = "MHSA-GRU-Transformer-v1.0"

# Model repository configuration
MODEL_REPO = {
    "repo_id": "camlas/toxicity",
    "files": {
        "classifier": "mhsa_gru_classifier.pth",
        "scaler": "scaler.pkl",
        "config": "config.json",
        "model_weights": "model.safetensors",
        "vocab": "vocab.txt",
        "tokenizer_config": "tokenizer_config.json",
        "special_tokens_map": "special_tokens_map.json"
    }
}

# Global model variables
classifier = None
scaler = None
transformer_model = None
transformer_tokenizer = None
EMBEDDING_TYPE = "Bert"
MODEL_NAME = "ProtBERT"


# ========================= PYDANTIC MODELS =========================
class SequenceRequest(BaseModel):
    sequence: str


class BatchSequenceRequest(BaseModel):
    sequences: List[str]


class PredictionResponse(BaseModel):
    status_code: int
    status: str
    success: bool
    data: Optional[dict] = None
    error: Optional[str] = None
    error_code: Optional[str] = None
    timestamp: str
    api_version: str
    processing_time_ms: float


class HealthResponse(BaseModel):
    status_code: int
    status: str
    service: str
    api_version: str
    model_version: str
    models_loaded: bool
    models_loaded_count: int
    total_models_required: int
    model_sources: dict
    repository_info: dict
    device: str
    timestamp: str


# ========================= HELPER FUNCTIONS =========================
def create_kmers(sequence, k=6):
    """Convert DNA sequence to k-mer tokens (for DNABERT)"""
    kmers = []
    for i in range(len(sequence) - k + 1):
        kmer = sequence[i:i+k]
        kmers.append(kmer)
    return ' '.join(kmers)


def ensure_models_directory():
    models_dir = "models"
    if not os.path.exists(models_dir):
        os.makedirs(models_dir)
        print(f"✅ Created {models_dir} directory")
    return models_dir


def download_model_from_hub(model_name: str) -> Optional[str]:
    """Download individual model files from HuggingFace Hub"""
    try:
        if model_name not in MODEL_REPO["files"]:
            raise ValueError(f"Unknown model: {model_name}")

        filename = MODEL_REPO["files"][model_name]
        repo_id = MODEL_REPO["repo_id"]
        models_dir = ensure_models_directory()
        local_path = os.path.join(models_dir, filename)

        if os.path.exists(local_path):
            print(f"✅ Found {model_name} in local models directory: {local_path}")
            return local_path

        print(f"📥 Downloading {model_name} ({filename}) from {repo_id}...")
        token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")

        if not token:
            print("⚠️ Warning: No HF token found. This may fail for private repositories.")
        
        temp_model_path = hf_hub_download(
            repo_id=repo_id,
            filename=filename,
            repo_type="model",
            token=token
        )

        shutil.copy2(temp_model_path, local_path)
        print(f"✅ {model_name} downloaded and stored!")
        return local_path

    except Exception as e:
        print(f"❌ Error downloading {model_name}: {e}")
        return None


def extract_features_from_sequence(sequence: str):
    """Extract features from sequence using ProtBERT"""
    global transformer_model, transformer_tokenizer
    
    if transformer_model is None or transformer_tokenizer is None:
        raise ValueError("ProtBERT model not loaded")
    
    # ProtBERT expects sequences with spaces between amino acids
    # Convert "MKTAYIAKQR" to "M K T A Y I A K Q R"
    processed_seq = ' '.join(list(sequence.upper()))
    
    # Tokenize
    inputs = transformer_tokenizer(
        processed_seq,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=512
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    # Extract features
    with torch.no_grad():
        outputs = transformer_model(**inputs)
        # Use [CLS] token embedding
        cls_embeddings = outputs.last_hidden_state[:, 0, :]
    
    return cls_embeddings.cpu().numpy()


def load_all_models():
    """Load all models from HuggingFace Hub"""
    global classifier, scaler, transformer_model, transformer_tokenizer
    
    models_dir = ensure_models_directory()
    models_loaded = {
        "classifier": False,
        "scaler": False,
        "transformer_model": False,
        "transformer_tokenizer": False
    }

    print(f"🚀 Loading models from {MODEL_REPO['repo_id']}...")
    print("=" * 60)

    try:
        # Download all necessary files
        print("📥 Downloading ProtBERT model files...")
        
        files_to_download = ["config", "model_weights", "vocab", 
                            "tokenizer_config", "special_tokens_map"]
        
        for file_key in files_to_download:
            download_model_from_hub(file_key)
        
        # Load ProtBERT Tokenizer
        print("🔄 Loading ProtBERT tokenizer...")
        try:
            transformer_tokenizer = BertTokenizer.from_pretrained(
                models_dir,
                do_lower_case=False,
                local_files_only=True
            )
            models_loaded["transformer_tokenizer"] = True
            print("✅ ProtBERT tokenizer loaded!")
        except Exception as e:
            print(f"❌ Error loading tokenizer: {e}")
            # Try loading from HuggingFace directly
            print("🔄 Trying to load tokenizer directly from HuggingFace...")
            token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
            transformer_tokenizer = BertTokenizer.from_pretrained(
                MODEL_REPO["repo_id"],
                do_lower_case=False,
                token=token
            )
            models_loaded["transformer_tokenizer"] = True
            print("✅ ProtBERT tokenizer loaded from HuggingFace!")
        
        # Load ProtBERT Model
        print("🔄 Loading ProtBERT model...")
        try:
            transformer_model = BertModel.from_pretrained(
                models_dir,
                local_files_only=True
            )
            models_loaded["transformer_model"] = True
            print("✅ ProtBERT model loaded!")
        except Exception as e:
            print(f"❌ Error loading model: {e}")
            # Try loading from HuggingFace directly
            print("🔄 Trying to load model directly from HuggingFace...")
            token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
            transformer_model = BertModel.from_pretrained(
                MODEL_REPO["repo_id"],
                token=token
            )
            models_loaded["transformer_model"] = True
            print("✅ ProtBERT model loaded from HuggingFace!")
        
        transformer_model.to(device)
        transformer_model.eval()

        # Load Classifier
        print("🔄 Loading classifier (MHSA-GRU)...")
        clf_path = os.path.join(models_dir, MODEL_REPO["files"]["classifier"])
        
        if not os.path.exists(clf_path):
            print("📥 Classifier not found locally, downloading...")
            clf_path = download_model_from_hub("classifier")
        
        if clf_path and os.path.exists(clf_path):
            checkpoint = torch.load(clf_path, map_location=device, weights_only=False)
            
            # Handle different checkpoint formats
            if 'input_dim' in checkpoint:
                input_dim = checkpoint['input_dim']
            else:
                # ProtBERT embedding size is 1024
                input_dim = 1024
            
            classifier = MHSA_GRU(input_dim, hidden_dim=256)
            
            # Load state dict
            if 'model_state_dict' in checkpoint:
                classifier.load_state_dict(checkpoint['model_state_dict'])
            else:
                classifier.load_state_dict(checkpoint)
            
            classifier.to(device)
            classifier.eval()
            models_loaded["classifier"] = True
            print(f"✅ Classifier loaded! (input_dim: {input_dim})")

        # Load Scaler
        print("🔄 Loading feature scaler...")
        scaler_path = os.path.join(models_dir, MODEL_REPO["files"]["scaler"])
        
        if not os.path.exists(scaler_path):
            print("📥 Scaler not found locally, downloading...")
            scaler_path = download_model_from_hub("scaler")
        
        if scaler_path and os.path.exists(scaler_path):
            scaler = joblib.load(scaler_path)
            models_loaded["scaler"] = True
            print("✅ Scaler loaded!")

        loaded_count = sum(models_loaded.values())
        total_count = len(models_loaded)

        print(f"\n📊 Model Loading Summary:")
        print(f"   • Successfully loaded: {loaded_count}/{total_count}")
        print(f"   • Repository: {MODEL_REPO['repo_id']}")
        print(f"   • Embedding Model: {MODEL_NAME}")
        print(f"   • Device: {device}")
        
        critical_models = ["classifier", "scaler", "transformer_model", "transformer_tokenizer"]
        critical_loaded = all(models_loaded[m] for m in critical_models)

        if critical_loaded:
            print("🎉 All critical models loaded successfully!")
            return True
        else:
            print("⚠️ Some critical models failed to load")
            print(f"   Models status: {models_loaded}")
            return False

    except Exception as e:
        print(f"❌ Error loading models: {e}")
        import traceback
        traceback.print_exc()
        return False


# ========================= FASTAPI APPLICATION =========================
@asynccontextmanager
async def lifespan(app: FastAPI):
    # Startup
    print("🚀 Starting Toxicity Prediction API...")
    success = load_all_models()
    if not success:
        print("⚠️ Warning: Not all models loaded successfully")
    yield
    # Shutdown
    print("🔄 Shutting down API...")


app = FastAPI(
    title="Toxicity Prediction API",
    description="API for toxicity prediction using MHSA-GRU with Transformer embeddings",
    version="1.0.0",
    lifespan=lifespan
)


@app.get("/")
async def root():
    return {
        "message": "Toxicity Prediction API",
        "version": API_VERSION,
        "endpoints": {
            "/predict": "POST - Predict toxicity for a single sequence",
            "/predict/batch": "POST - Predict toxicity for multiple sequences",
            "/example": "GET - Try the API with a hardcoded example sequence",
            "/health": "GET - Check API health and model status"
        },
        "example_usage": {
            "single": {
                "method": "POST",
                "url": "/predict",
                "body": {"sequence": "MKTAYIAKQRQISFVKSHFSRQLE"}
            },
            "batch": {
                "method": "POST",
                "url": "/predict/batch",
                "body": {
                    "sequences": [
                        "MLLPATMSDKPDMAEIEKFDKSKLKKTETQEKNPLPSKETIEQEKQAGES",
                        "MFGLPQQEVSEEEKRAHQEQTEKTLKQAAYVAAFLWVSPMIWHLVKKQWK"
                    ]
                }
            },
            "example": {
                "method": "GET",
                "url": "/example",
                "description": "No input needed - just call this endpoint"
            }
        }
    }


@app.post("/predict", response_model=PredictionResponse)
async def predict(request: SequenceRequest):
    start_time = time.time()
    timestamp = datetime.now(timezone.utc).isoformat()

    try:
        if not request.sequence or len(request.sequence) == 0:
            raise HTTPException(
                status_code=400,
                detail={
                    "status_code": 400,
                    "status": "error",
                    "success": False,
                    "error": "No sequence provided",
                    "error_code": "MISSING_SEQUENCE",
                    "timestamp": timestamp,
                    "api_version": API_VERSION,
                    "processing_time_ms": round((time.time() - start_time) * 1000, 2)
                }
            )

        # Check if models are loaded
        if classifier is None or scaler is None or transformer_model is None:
            raise HTTPException(
                status_code=503,
                detail={
                    "status_code": 503,
                    "status": "error",
                    "success": False,
                    "error": "Models not loaded properly",
                    "error_code": "MODEL_NOT_LOADED",
                    "timestamp": timestamp,
                    "api_version": API_VERSION,
                    "processing_time_ms": round((time.time() - start_time) * 1000, 2)
                }
            )

        # Validate sequence
        sequence = request.sequence.upper().strip()
        if len(sequence) < 10:
            raise HTTPException(
                status_code=400,
                detail={
                    "status_code": 400,
                    "status": "error",
                    "success": False,
                    "error": "Sequence too short (minimum 10 characters)",
                    "error_code": "SEQUENCE_TOO_SHORT",
                    "timestamp": timestamp,
                    "api_version": API_VERSION,
                    "processing_time_ms": round((time.time() - start_time) * 1000, 2)
                }
            )

        # Step 1: Extract features using ProtBERT
        features = extract_features_from_sequence(sequence)

        # Step 2: Scale features
        scaled_features = scaler.transform(features)

        # Step 3: Predict using MHSA-GRU
        features_tensor = torch.FloatTensor(scaled_features).to(device)
        
        with torch.no_grad():
            probability = classifier(features_tensor).cpu().numpy()[0, 0]

        # Determine prediction
        prediction_class = 1 if probability > 0.5 else 0
        predicted_label = "Toxic" if prediction_class == 1 else "Non-Toxic"
        confidence = float(abs(probability - 0.5) * 2)

        # Determine confidence level
        if confidence > 0.8:
            confidence_level = "high"
        elif confidence > 0.6:
            confidence_level = "medium"
        else:
            confidence_level = "low"

        processing_time = round((time.time() - start_time) * 1000, 2)

        return PredictionResponse(
            status_code=200,
            status="success",
            success=True,
            data={
                "sequence": sequence[:100] + "..." if len(sequence) > 100 else sequence,
                "sequence_length": len(sequence),
                "prediction": {
                    "predicted_class": predicted_label,
                    "confidence": confidence,
                    "confidence_level": confidence_level,
                    "toxicity_score": float(probability),
                    "non_toxicity_score": float(1 - probability)
                },
                "metadata": {
                    "embedding_model": MODEL_NAME,
                    "embedding_type": EMBEDDING_TYPE,
                    "model_version": MODEL_VERSION,
                    "device": str(device)
                }
            },
            timestamp=timestamp,
            api_version=API_VERSION,
            processing_time_ms=processing_time
        )

    except HTTPException:
        raise
    except Exception as e:
        processing_time = round((time.time() - start_time) * 1000, 2)
        raise HTTPException(
            status_code=500,
            detail={
                "status_code": 500,
                "status": "error",
                "success": False,
                "error": f"Internal server error: {str(e)}",
                "error_code": "INTERNAL_ERROR",
                "timestamp": timestamp,
                "api_version": API_VERSION,
                "processing_time_ms": processing_time
            }
        )


@app.post("/predict/batch", response_model=PredictionResponse)
async def predict_batch(request: BatchSequenceRequest):
    """
    Predict toxicity for multiple sequences at once.
    
    Example request body:
    {
        "sequences": [
            "MLLPATMSDKPDMAEIEKFDKSKLKKTETQEKNPLPSKETIEQEKQAGES",
            "MFGLPQQEVSEEEKRAHQEQTEKTLKQAAYVAAFLWVSPMIWHLVKKQWK"
        ]
    }
    """
    start_time = time.time()
    timestamp = datetime.now(timezone.utc).isoformat()

    try:
        if not request.sequences or len(request.sequences) == 0:
            raise HTTPException(
                status_code=400,
                detail={
                    "status_code": 400,
                    "status": "error",
                    "success": False,
                    "error": "No sequences provided",
                    "error_code": "MISSING_SEQUENCES",
                    "timestamp": timestamp,
                    "api_version": API_VERSION,
                    "processing_time_ms": round((time.time() - start_time) * 1000, 2)
                }
            )

        # Check if models are loaded
        if classifier is None or scaler is None or transformer_model is None:
            raise HTTPException(
                status_code=503,
                detail={
                    "status_code": 503,
                    "status": "error",
                    "success": False,
                    "error": "Models not loaded properly",
                    "error_code": "MODEL_NOT_LOADED",
                    "timestamp": timestamp,
                    "api_version": API_VERSION,
                    "processing_time_ms": round((time.time() - start_time) * 1000, 2)
                }
            )

        results = []
        
        for idx, seq in enumerate(request.sequences, 1):
            try:
                sequence = seq.upper().strip()
                
                # Validate sequence length
                if len(sequence) < 10:
                    results.append({
                        "sequence_index": idx,
                        "sequence": sequence[:100] + "..." if len(sequence) > 100 else sequence,
                        "sequence_length": len(sequence),
                        "error": "Sequence too short (minimum 10 characters)",
                        "predicted_class": None,
                        "toxicity_score": None,
                        "confidence": None
                    })
                    continue
                
                # Extract features using ProtBERT
                features = extract_features_from_sequence(sequence)
                scaled_features = scaler.transform(features)
                features_tensor = torch.FloatTensor(scaled_features).to(device)
                
                with torch.no_grad():
                    probability = classifier(features_tensor).cpu().numpy()[0, 0]
                
                prediction_class = 1 if probability > 0.5 else 0
                predicted_label = "Toxic" if prediction_class == 1 else "Non-Toxic"
                confidence = float(abs(probability - 0.5) * 2)
                
                # Determine confidence level
                if confidence > 0.8:
                    confidence_level = "high"
                elif confidence > 0.6:
                    confidence_level = "medium"
                else:
                    confidence_level = "low"
                
                results.append({
                    "sequence_index": idx,
                    "sequence": sequence[:100] + "..." if len(sequence) > 100 else sequence,
                    "sequence_length": len(sequence),
                    "predicted_class": predicted_label,
                    "toxicity_score": float(probability),
                    "non_toxicity_score": float(1 - probability),
                    "confidence": confidence,
                    "confidence_level": confidence_level,
                    "error": None
                })
                
            except Exception as e:
                # Handle individual sequence errors without stopping the batch
                results.append({
                    "sequence_index": idx,
                    "sequence": seq[:100] + "..." if len(seq) > 100 else seq,
                    "sequence_length": len(seq),
                    "error": f"Error processing sequence: {str(e)}",
                    "predicted_class": None,
                    "toxicity_score": None,
                    "confidence": None
                })

        processing_time = round((time.time() - start_time) * 1000, 2)
        
        # Count successful predictions
        successful_predictions = sum(1 for r in results if r.get("predicted_class") is not None)

        return PredictionResponse(
            status_code=200,
            status="success",
            success=True,
            data={
                "total_sequences": len(request.sequences),
                "successful_predictions": successful_predictions,
                "failed_predictions": len(request.sequences) - successful_predictions,
                "results": results,
                "metadata": {
                    "embedding_model": MODEL_NAME,
                    "embedding_type": EMBEDDING_TYPE,
                    "model_version": MODEL_VERSION,
                    "device": str(device)
                }
            },
            timestamp=timestamp,
            api_version=API_VERSION,
            processing_time_ms=processing_time
        )

    except HTTPException:
        raise
    except Exception as e:
        processing_time = round((time.time() - start_time) * 1000, 2)
        raise HTTPException(
            status_code=500,
            detail={
                "status_code": 500,
                "status": "error",
                "success": False,
                "error": f"Internal server error: {str(e)}",
                "error_code": "INTERNAL_ERROR",
                "timestamp": timestamp,
                "api_version": API_VERSION,
                "processing_time_ms": processing_time
            }
        )


@app.get("/example", response_model=PredictionResponse)
async def predict_example():
    """
    Predict using a hardcoded example protein sequence.
    No input required - just call this endpoint to see how the API works.
    
    Example sequence: MLLPATMSDKPDMAEIEKFDKSKLKKTETQEKNPLPSKETIEQEKQAGES
    """
    start_time = time.time()
    timestamp = datetime.now(timezone.utc).isoformat()

    # Hardcoded example sequence
    EXAMPLE_SEQUENCE = "MLLPATMSDKPDMAEIEKFDKSKLKKTETQEKNPLPSKETIEQEKQAGES"

    try:
        # Check if models are loaded
        if classifier is None or scaler is None or transformer_model is None:
            raise HTTPException(
                status_code=503,
                detail={
                    "status_code": 503,
                    "status": "error",
                    "success": False,
                    "error": "Models not loaded properly",
                    "error_code": "MODEL_NOT_LOADED",
                    "timestamp": timestamp,
                    "api_version": API_VERSION,
                    "processing_time_ms": round((time.time() - start_time) * 1000, 2)
                }
            )

        sequence = EXAMPLE_SEQUENCE.upper().strip()

        # Step 1: Extract features using ProtBERT
        features = extract_features_from_sequence(sequence)

        # Step 2: Scale features
        scaled_features = scaler.transform(features)

        # Step 3: Predict using MHSA-GRU
        features_tensor = torch.FloatTensor(scaled_features).to(device)
        
        with torch.no_grad():
            probability = classifier(features_tensor).cpu().numpy()[0, 0]

        # Determine prediction
        prediction_class = 1 if probability > 0.5 else 0
        predicted_label = "Toxic" if prediction_class == 1 else "Non-Toxic"
        confidence = float(abs(probability - 0.5) * 2)

        # Determine confidence level
        if confidence > 0.8:
            confidence_level = "high"
        elif confidence > 0.6:
            confidence_level = "medium"
        else:
            confidence_level = "low"

        processing_time = round((time.time() - start_time) * 1000, 2)

        return PredictionResponse(
            status_code=200,
            status="success",
            success=True,
            data={
                "note": "This is an example prediction using a hardcoded sequence",
                "sequence": sequence,
                "sequence_length": len(sequence),
                "prediction": {
                    "predicted_class": predicted_label,
                    "confidence": confidence,
                    "confidence_level": confidence_level,
                    "toxicity_score": float(probability),
                    "non_toxicity_score": float(1 - probability)
                },
                "metadata": {
                    "embedding_model": MODEL_NAME,
                    "embedding_type": EMBEDDING_TYPE,
                    "model_version": MODEL_VERSION,
                    "device": str(device),
                    "source": "hardcoded_example"
                }
            },
            timestamp=timestamp,
            api_version=API_VERSION,
            processing_time_ms=processing_time
        )

    except HTTPException:
        raise
    except Exception as e:
        processing_time = round((time.time() - start_time) * 1000, 2)
        raise HTTPException(
            status_code=500,
            detail={
                "status_code": 500,
                "status": "error",
                "success": False,
                "error": f"Internal server error: {str(e)}",
                "error_code": "INTERNAL_ERROR",
                "timestamp": timestamp,
                "api_version": API_VERSION,
                "processing_time_ms": processing_time
            }
        )


@app.get("/health", response_model=HealthResponse)
async def health_check():
    models_loaded = all([
        classifier is not None,
        scaler is not None,
        transformer_model is not None,
        transformer_tokenizer is not None
    ])

    model_sources = {
        "classifier": {
            "loaded": classifier is not None,
            "source": "huggingface_hub",
            "repository": MODEL_REPO["repo_id"]
        },
        "scaler": {
            "loaded": scaler is not None,
            "source": "huggingface_hub",
            "repository": MODEL_REPO["repo_id"]
        },
        "transformer_model": {
            "loaded": transformer_model is not None,
            "model_name": MODEL_NAME,
            "source": "huggingface_hub",
            "repository": MODEL_REPO["repo_id"]
        }
    }

    repository_info = {
        "repository_id": MODEL_REPO["repo_id"],
        "embedding_type": EMBEDDING_TYPE,
        "model_name": MODEL_NAME,
        "total_models": len(MODEL_REPO["files"])
    }

    return HealthResponse(
        status_code=200 if models_loaded else 503,
        status="healthy" if models_loaded else "unhealthy",
        service="Toxicity Prediction API",
        api_version=API_VERSION,
        model_version=MODEL_VERSION,
        models_loaded=models_loaded,
        models_loaded_count=sum(1 for source in model_sources.values() if source["loaded"]),
        total_models_required=3,
        model_sources=model_sources,
        repository_info=repository_info,
        device=str(device),
        timestamp=datetime.now(timezone.utc).isoformat()
    )


if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)