inference.py

# First try to import with fallbacks
try:
    import torch
    from transformers import AutoTokenizer, AutoModel
    import joblib
    from huggingface_hub import hf_hub_download
    import json
except ImportError as e:
    print(f"Import error: {e}")
    # Try to install missing packages (this might not work in Spaces)
    import subprocess
    import sys
    subprocess.check_call([sys.executable, "-m", "pip", "install", "torch", "transformers", "joblib", "huggingface-hub"])
    import torch
    from transformers import AutoTokenizer, AutoModel
    import joblib
    from huggingface_hub import hf_hub_download
    import json

class DrugInteractionClassifier(torch.nn.Module):
    def __init__(self, n_classes, bert_model_name="emilyalsentzer/Bio_ClinicalBERT"):
        super(DrugInteractionClassifier, self).__init__()
        self.bert = AutoModel.from_pretrained(bert_model_name)
        self.classifier = torch.nn.Sequential(
            torch.nn.Linear(self.bert.config.hidden_size, 256),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.3),
            torch.nn.Linear(256, n_classes)
        )

    def forward(self, input_ids, attention_mask):
        bert_output = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = bert_output[0][:, 0, :]
        return self.classifier(pooled_output)

class DDIPredictor:
    def __init__(self, repo_id="Fredaaaaaa/drug_interaction_severity"):
        self.repo_id = repo_id
        print(f"🚀 Loading model from: {repo_id}")
        
        try:
            # Download model files from Hugging Face
            print("📥 Downloading config.json...")
            self.config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
            
            print("📥 Downloading pytorch_model.bin...")
            self.model_path = hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin")
            
            print("📥 Downloading label_encoder.joblib...")
            self.label_encoder_path = hf_hub_download(repo_id=repo_id, filename="label_encoder.joblib")
            
            # Load config
            with open(self.config_path, "r") as f:
                self.config = json.load(f)
            
            # Load tokenizer from repo
            print("🔤 Loading tokenizer...")
            self.tokenizer = AutoTokenizer.from_pretrained(repo_id)
            
            # Load label encoder
            print("🏷️ Loading label encoder...")
            self.label_encoder = joblib.load(self.label_encoder_path)
            
            # Initialize model
            print("🧠 Initializing model...")
            self.model = DrugInteractionClassifier(
                n_classes=self.config["num_labels"],
                bert_model_name=self.config["bert_model_name"]
            )
            
            # Load weights
            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
            print(f"⚙️ Loading weights on {device}...")
            self.model.load_state_dict(
                torch.load(self.model_path, map_location=device)
            )
            self.model.to(device)
            self.model.eval()
            
            self.device = device
            print(f"✅ Model loaded successfully from {repo_id} on {device}")
            
        except Exception as e:
            print(f"❌ Error loading model: {e}")
            raise e
    
    def predict(self, text, confidence_threshold=0.0):
        """Predict drug interaction severity"""
        if not text or not text.strip():
            return {
                "prediction": "Invalid Input",
                "confidence": 0.0,
                "probabilities": {label: 0.0 for label in self.label_encoder.classes_}
            }
        
        try:
            # Tokenize
            inputs = self.tokenizer(
                text,
                max_length=self.config.get("max_length", 128),
                padding=True,
                truncation=True,
                return_tensors="pt"
            )
            
            inputs = {k: v.to(self.device) for k, v in inputs.items()}
            
            # Predict
            with torch.no_grad():
                outputs = self.model(inputs["input_ids"], inputs["attention_mask"])
                probabilities = torch.softmax(outputs, dim=1)
                confidence, predicted_idx = torch.max(probabilities, dim=1)
            
            predicted_label = self.label_encoder.inverse_transform([predicted_idx.item()])[0]
            
            # Get all probabilities
            all_probs = {
                self.label_encoder.inverse_transform([i])[0]: prob.item() 
                for i, prob in enumerate(probabilities[0])
            }
            
            return {
                "prediction": predicted_label,
                "confidence": confidence.item(),
                "probabilities": all_probs
            }
            
        except Exception as e:
            return {
                "prediction": f"Error: {str(e)}",
                "confidence": 0.0,
                "probabilities": {label: 0.0 for label in self.label_encoder.classes_}
            }

# Global predictor instance
try:
    predictor = DDIPredictor("Fredaaaaaa/drug_interaction_severity")
    MODEL_LOADED = True
except Exception as e:
    print(f"Failed to load model: {e}")
    predictor = None
    MODEL_LOADED = False