predict.py

import torch
import os
import numpy as np
import logging
import glob
import pickle
from rdkit import Chem
from model import DMPNN, MolGraph, BatchMolGraph
from data import TASKS, get_global_features, SCALER_FILE

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

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

def _sanitize_array(x):
    """
    Ensure JSON-safe numeric output.
    NaN -> 0.0, +inf -> 1.0, -inf -> 0.0
    """
    return np.nan_to_num(x, nan=0.0, posinf=1.0, neginf=0.0)

class Tox21Predictor:
    def __init__(self, model_dir="."):
        self.models = []
        self.scaler = None

        # Load scaler
        if os.path.exists(SCALER_FILE):
            try:
                with open(SCALER_FILE, "rb") as f:
                    self.scaler = pickle.load(f)
                logger.info("Scaler loaded.")
            except Exception as e:
                logger.error(f"Error loading scaler: {e}")

        # Load ensemble checkpoints
        model_paths = glob.glob(os.path.join("checkpoints", "model_seed_*.pt"))
        for path in model_paths:
            try:
                # Fix: Using n_tasks to match the DMPNN __init__ in model.py
                model = DMPNN(
                    hidden_size=300,
                    depth=3,
                    global_feats_size=217,
                    n_tasks=len(TASKS),
                )
                model.load_state_dict(torch.load(path, map_location=DEVICE))
                model.to(DEVICE)
                model.eval()
                self.models.append(model)
                logger.info(f"Loaded checkpoint: {path}")
            except Exception as e:
                logger.error(f"Failed to load {path}: {e}")

    def predict(self, smiles_list):
        # Default results: mapping SMILES to a dict of tasks set to 0.0
        results = {s: {task: 0.0 for task in TASKS} for s in smiles_list}

        if not self.models:
            logger.warning("No models loaded. Returning default zeros.")
            return results

        valid_indices = []
        valid_smiles = []
        mol_graphs = []
        global_features = []

        for i, s in enumerate(smiles_list):
            try:
                mol = Chem.MolFromSmiles(s)
                if mol is None:
                    continue
                
                # Featurization
                feats = get_global_features(mol)
                feats = _sanitize_array(feats)
                
                valid_indices.append(i)
                valid_smiles.append(s)
                mol_graphs.append(MolGraph(s))
                global_features.append(feats)
            except Exception as e:
                logger.debug(f"SMILES error for {s}: {e}")
                continue

        if not mol_graphs:
            return results

        # Scale features
        global_features = np.stack(global_features)
        if self.scaler is not None:
            try:
                global_features = self.scaler.transform(global_features)
            except Exception as e:
                logger.error(f"Scaling error: {e}")

        global_features = _sanitize_array(global_features)
        global_features_tensor = torch.tensor(
            global_features, dtype=torch.float32, device=DEVICE
        )

        # Batch graph processing
        batch_graph = BatchMolGraph(mol_graphs)
        ensemble_preds = []

        # Inference
        with torch.no_grad():
            for model in self.models:
                logits = model(batch_graph, global_features_tensor)
                probs = torch.sigmoid(logits).cpu().numpy()
                ensemble_preds.append(_sanitize_array(probs))

        if not ensemble_preds:
            return results

        # Average ensemble predictions
        avg_preds = np.mean(np.stack(ensemble_preds, axis=0), axis=0)
        avg_preds = _sanitize_array(avg_preds)

        # Map back to results dictionary
        for batch_idx, original_idx in enumerate(valid_indices):
            s = smiles_list[original_idx]
            for t_idx, task in enumerate(TASKS):
                results[s][task] = float(avg_preds[batch_idx, t_idx])

        return results

# ============================
# HF LEADERBOARD ENTRY POINT
# ============================
_predictor = None

def predict(smiles_list):
    """
    Leaderboard Entry Point: Returns a dictionary with "predictions" key
    containing the raw dictionary results from Tox21Predictor.
    """
    global _predictor
    if _predictor is None:
        _predictor = Tox21Predictor()

    raw_results = _predictor.predict(smiles_list)

    return {"predictions": raw_results}