deeppurpose002.py

import os
import json
import math
import time
import argparse
from datetime import datetime
import matplotlib.pyplot as plt
import torch

import numpy as np
import pandas as pd

from tdc.multi_pred import DTI
from DeepPurpose import utils
from DeepPurpose import DTI as dp_models


# -------------------------
# Metrics (régression)
# -------------------------
def mse(y_true, y_pred):
    y_true = np.asarray(y_true, dtype=float).reshape(-1)
    y_pred = np.asarray(y_pred, dtype=float).reshape(-1)
    return float(np.mean((y_true - y_pred) ** 2))

def mae(y_true, y_pred):
    y_true = np.asarray(y_true, dtype=float).reshape(-1)
    y_pred = np.asarray(y_pred, dtype=float).reshape(-1)
    return float(np.mean(np.abs(y_true - y_pred)))

def pearson(y_true, y_pred):
    y_true = np.asarray(y_true, dtype=float).reshape(-1)
    y_pred = np.asarray(y_pred, dtype=float).reshape(-1)
    if y_true.size < 2 or np.std(y_true) == 0 or np.std(y_pred) == 0:
        return float("nan")
    return float(np.corrcoef(y_true, y_pred)[0, 1])

def spearman(y_true, y_pred):
    a = pd.Series(np.asarray(y_true, dtype=float).reshape(-1)).rank(method="average").to_numpy()
    b = pd.Series(np.asarray(y_pred, dtype=float).reshape(-1)).rank(method="average").to_numpy()
    return pearson(a, b)

def concordance_index_approx(y_true, y_pred, max_n=2000, seed=0):
    """
    CI approximatif (échantillonné si trop grand).
    """
    y_true = np.asarray(y_true, dtype=float).reshape(-1)
    y_pred = np.asarray(y_pred, dtype=float).reshape(-1)
    n = len(y_true)
    if n < 2:
        return float("nan")

    if n > max_n:
        rng = np.random.default_rng(seed)
        idx = rng.choice(n, size=max_n, replace=False)
        y_true = y_true[idx]
        y_pred = y_pred[idx]
        n = max_n

    conc = 0.0
    total = 0.0
    for i in range(n):
        for j in range(i + 1, n):
            if y_true[i] == y_true[j]:
                continue
            total += 1.0
            dt = y_true[i] - y_true[j]
            dp = y_pred[i] - y_pred[j]
            prod = dt * dp
            if prod > 0:
                conc += 1.0
            elif prod == 0:
                conc += 0.5
    if total == 0:
        return float("nan")
    return float(conc / total)


# -------------------------
# Utils
# -------------------------
def log_section(title):
    bar = "=" * 70
    print("\n" + bar)
    print(title)
    print(bar)

def detect_cols(df):
    cols = df.columns.tolist()
    # TDC DTI: souvent Drug, Target, Y
    drug_col = "Drug" if "Drug" in cols else None
    target_col = "Target" if "Target" in cols else None
    y_col = "Y" if "Y" in cols else None

    # fallback "best effort"
    if drug_col is None:
        for c in cols:
            if "smiles" in c.lower() or "drug" in c.lower():
                drug_col = c
                break
    if target_col is None:
        for c in cols:
            if "sequence" in c.lower() or "target" in c.lower() or "protein" in c.lower():
                target_col = c
                break
    if y_col is None:
        for c in cols:
            if c.lower() in {"y", "label", "labels", "affinity"}:
                y_col = c
                break

    # dernier fallback: 0,1,2
    if drug_col is None and len(cols) >= 1:
        drug_col = cols[0]
    if target_col is None and len(cols) >= 2:
        target_col = cols[1]
    if y_col is None and len(cols) >= 3:
        y_col = cols[2]

    return drug_col, target_col, y_col

def label_transform(y, mode, dataset_name):
    y = np.asarray(y, dtype=float)

    # Force auto to paffinity_nm for standard datasets
    if mode == "auto":
        if dataset_name.lower() in ["davis", "kiba"] or dataset_name.startswith("BindingDB"):
            mode = "paffinity_nm"
        else:
            mode = "none"

    if mode == "none":
        return y, "none"

    if mode == "paffinity_nm":
        # SAFETY CHECK: Clip values to avoid log(0) or log(negative)
        y = np.where(y < 1e-9, 1e-9, y) 
        # Convert nM to pM ( -log10( Molar ) )
        # Value 100 nM = 100e-9 M = 1e-7 M -> -log10(1e-7) = 7.0
        # Formula: 9 - log10(nM)
        y = 9.0 - np.log10(y)
        return y, "paffinity_nm"

    raise ValueError(f"Unknown label_transform: {mode}")

def make_run_dir(base_dir, dataset):
    ts = datetime.now().strftime("%Y%m%d_%H%M%S")
    run_id = f"{ts}_{dataset}"
    run_dir = os.path.join(base_dir, run_id)
    os.makedirs(run_dir, exist_ok=True)
    return run_id, run_dir

def check_gpu():
    """Check GPU availability and return CUDA ID for DeepPurpose."""
    if torch.cuda.is_available():
        device_name = torch.cuda.get_device_name(0)
        gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1e9
        print(f"\n[SYSTEM] ✅ CUDA GPU Detected: {device_name}")
        print(f"[SYSTEM]    Memory: {gpu_memory:.1f} GB")
        print(f"[SYSTEM]    CUDA Version: {torch.version.cuda}")
        return 0  # Use GPU 0
    else:
        print("\n[SYSTEM] ⚠️ No GPU detected. Running on CPU (will be slow).")
        return -1  # Use CPU


def main():
    # --- GPU DETECTION FIRST ---
    use_cuda = check_gpu()
    
    ap = argparse.ArgumentParser()
    ap.add_argument("--dataset", default="DAVIS", help="DAVIS | KIBA | BindingDB_Kd | BindingDB_Ki | BindingDB_IC50")
    ap.add_argument("--drug_enc", default="Morgan")
    ap.add_argument("--target_enc", default="CNN")
    ap.add_argument("--epochs", type=int, default=10)
    ap.add_argument("--batch", type=int, default=256)
    ap.add_argument("--lr", type=float, default=1e-4)
    ap.add_argument("--seed", type=int, default=1)
    ap.add_argument("--split", default="random", help="DeepPurpose split_method (random recommandé ici)")
    ap.add_argument("--frac_train", type=float, default=0.8)
    ap.add_argument("--frac_val", type=float, default=0.1)
    ap.add_argument("--frac_test", type=float, default=0.1)
    ap.add_argument("--label_transform", default="paffinity_nm", help="Force log transform! (auto | none | paffinity_nm)")
    ap.add_argument("--harmonize", default="none", help="none | mean | max_affinity (utile surtout BindingDB_*)")
    ap.add_argument("--runs_dir", default="runs")
    ap.add_argument("--dry_run", action="store_true", help="Charge dataset + prints info, sans training")
    ap.add_argument("--gpu", type=int, default=None, help="Override GPU ID (default: auto-detect)")
    args = ap.parse_args()

    # Override GPU if specified
    if args.gpu is not None:
        use_cuda = args.gpu
        print(f"[SYSTEM] Using specified GPU: {use_cuda}")

    np.random.seed(args.seed)

    run_id, run_dir = make_run_dir(args.runs_dir, args.dataset)
    pred_path = os.path.join(run_dir, "predictions_test.csv")
    summary_path = os.path.join(run_dir, "run_summary.json")

    # -------------------------
    # 1) LOAD DATA (TDC)
    # -------------------------
    log_section("[1] LOAD DATASET (TDC)")
    print(f"[RUN] run_id={run_id}")
    print(f"[RUN] dataset={args.dataset}")
    print(f"[RUN] cache=PyTDC (download auto si nécessaire)")

    t0 = time.time()
    data = DTI(name=args.dataset)

    if args.harmonize != "none" and args.dataset.startswith("BindingDB_"):
        mode = args.harmonize
        print(f"[TDC] harmonize_affinities(mode='{mode}')")
        data.harmonize_affinities(mode=mode)

    # PyTDC renvoie un DataFrame via get_data() (selon version, get_data(format='df') existe aussi)
    try:
        df = data.get_data()
    except TypeError:
        df = data.get_data(format="df")

    print(f"[INPUT] raw_shape={df.shape}")
    drug_col, target_col, y_col = detect_cols(df)
    print(f"[INPUT] detected_cols: drug='{drug_col}', target='{target_col}', label='{y_col}'")

    # -------------------------
    # 2) CLEAN + PREP
    # -------------------------
    log_section("[2] CLEAN + PREP (Input DeepPurpose)")
    df = df[[drug_col, target_col, y_col]].copy()
    df.columns = ["Drug", "Target", "Y"]

    n0 = len(df)
    df = df.dropna()
    df["Drug"] = df["Drug"].astype(str).str.strip()
    df["Target"] = df["Target"].astype(str).str.strip()
    df = df[(df["Drug"] != "") & (df["Target"] != "")]
    df["Y"] = pd.to_numeric(df["Y"], errors="coerce")
    df = df.dropna()
    n1 = len(df)

    y_raw = df["Y"].astype(float).values
    y_trans, used_transform = label_transform(y_raw, args.label_transform, args.dataset)
    df["Y"] = y_trans
    df = df.dropna()
    n2 = len(df)

    # drop duplicates (Drug,Target) -> keep first
    df = df.drop_duplicates(subset=["Drug", "Target"], keep="first").reset_index(drop=True)
    n3 = len(df)

    print(f"[CLEAN] start={n0} -> after_na/parse={n1} -> after_transform={n2} -> after_dedup={n3}")
    print(f"[LABEL] transform={used_transform}")
    y = df["Y"].astype(float).values
    print(f"[LABEL] stats: min={np.min(y):.6f} | mean={np.mean(y):.6f} | median={np.median(y):.6f} | max={np.max(y):.6f}")
    print("[SAMPLE] first_rows:")
    print(df.head(3).to_string(index=False))

    if args.dry_run:
        print(f"\n[DRY_RUN] stop ici. (Aucun training lancé)")
        return

    # -------------------------
    # 3) SPLIT + ENCODE (DeepPurpose)
    # -------------------------
    log_section("[3] DATA_PROCESS (DeepPurpose)")
    frac = [args.frac_train, args.frac_val, args.frac_test]
    print(f"[SPLIT] method={args.split} | frac={frac} | seed={args.seed}")
    X_drugs = df["Drug"].values
    X_targets = df["Target"].values
    y = df["Y"].values

    train, val, test = utils.data_process(
        X_drugs, X_targets, y,
        drug_encoding=args.drug_enc,
        target_encoding=args.target_enc,
        split_method=args.split,
        frac=frac,
        random_seed=args.seed
    )

    # tailles (DeepPurpose objects ont souvent .Label)
    try:
        n_train, n_val, n_test = len(train.Label), len(val.Label), len(test.Label)
    except Exception:
        # fallback: len(obj)
        n_train, n_val, n_test = len(train), len(val), len(test)

    print(f"[SPLIT] sizes: train={n_train} | val={n_val} | test={n_test}")
    print(f"[ENC] drug_encoding={args.drug_enc} | target_encoding={args.target_enc}")

    # -------------------------
    # 4) MODEL INIT + TRAIN (WITH GPU CONFIG)
    # -------------------------
    log_section("[4] MODEL INIT + TRAIN")
    config = utils.generate_config(
        drug_encoding=args.drug_enc,
        target_encoding=args.target_enc,
        cls_hidden_dims=[1024, 1024, 512],
        train_epoch=args.epochs,
        batch_size=args.batch,
        LR=args.lr,
        result_folder=run_dir,
        cuda_id=use_cuda  # <<< GPU CONFIG HERE
    )

    print("[MODEL] config:")
    print(f"  epochs={args.epochs} | batch={args.batch} | lr={args.lr}")
    print(f"  hidden=[1024,1024,512] | result_dir={run_dir}")
    print(f"  cuda_id={use_cuda} {'(GPU)' if use_cuda >= 0 else '(CPU)'}")

    model = dp_models.model_initialize(**config)
    
    # Verify device placement
    if use_cuda >= 0:
        device = next(model.model.parameters()).device
        print(f"[MODEL] ✓ Model loaded on device: {device}")

    t_train0 = time.time()
    try:
        model.train(train, val, test)
    except TypeError:
        model.train(train, val)
    t_train1 = time.time()
    print(f"[TIME] train_seconds={t_train1 - t_train0:.1f}")

    # -------------------------
    # 5) EVAL + EXPORT
    # -------------------------
    log_section("[5] EVAL + EXPORT")
    print("[PREDICT] predicting on test...")
    
    # [FIX] Reset index to ensure alignment between DataFrame and Model Output
    test = test.reset_index(drop=True)
    
    y_true = np.asarray(test.Label.values, dtype=float).reshape(-1)
    
    # [DEBUG] Check what the model is actually predicting
    raw_pred = model.predict(test)
    y_pred = np.asarray(raw_pred, dtype=float).reshape(-1)

    # [DEBUG] Print first 5 comparisons to verify scaling matches
    print(f"[DEBUG] First 5 True: {y_true[:5]}")
    print(f"[DEBUG] First 5 Pred: {y_pred[:5]}")

    m_mse = mse(y_true, y_pred)
    m_rmse = float(math.sqrt(m_mse))
    m_mae = mae(y_true, y_pred)
    m_p = pearson(y_true, y_pred)
    m_s = spearman(y_true, y_pred)
    m_ci = concordance_index_approx(y_true, y_pred, max_n=2000, seed=args.seed)

    print("[METRICS] test:")
    print(f"  MSE   = {m_mse:.6f}")
    print(f"  RMSE  = {m_rmse:.6f}")
    print(f"  MAE   = {m_mae:.6f}")
    print(f"  Pearson  = {m_p:.6f}")
    print(f"  Spearman = {m_s:.6f}")
    print(f"  CI(approx) = {m_ci:.6f}")

    out_pred = pd.DataFrame({"y_true": y_true, "y_pred": y_pred})
    out_pred.to_csv(pred_path, index=False)
    print(f"[FILE] saved predictions: {pred_path}")

    summary = {
        "run_id": run_id,
        "dataset": args.dataset,
        "drug_encoding": args.drug_enc,
        "target_encoding": args.target_enc,
        "epochs": args.epochs,
        "batch": args.batch,
        "lr": args.lr,
        "seed": args.seed,
        "split_method": args.split,
        "frac": [args.frac_train, args.frac_val, args.frac_test],
        "label_transform": used_transform,
        "harmonize": args.harmonize,
        "n_rows_after_clean": int(len(df)),
        "cuda_id": use_cuda,
        "gpu_used": use_cuda >= 0,
        "gpu_name": torch.cuda.get_device_name(0) if use_cuda >= 0 else "CPU",
        "metrics_test": {
            "mse": m_mse,
            "rmse": m_rmse,
            "mae": m_mae,
            "pearson": m_p,
            "spearman": m_s,
            "ci_approx": m_ci,
        },
        "files": {
            "predictions_test_csv": pred_path,
        },
        "timing": {
            "load_seconds": time.time() - t0,
            "train_seconds": t_train1 - t_train0,
        }
    }
    with open(summary_path, "w", encoding="utf-8") as f:
        json.dump(summary, f, indent=2)
    print(f"[FILE] saved summary: {summary_path}")

    # -------------------------
    # 6) VISUALISATION
    # -------------------------
    log_section("[6] VISUALISATION")
    
    # Scatter plot
    scatter_png = os.path.join(run_dir, "scatter.png")
    plt.figure(figsize=(8, 6))
    plt.scatter(y_true, y_pred, s=8, alpha=0.5)
    plt.plot([y_true.min(), y_true.max()], [y_true.min(), y_true.max()], 'r--', label='Perfect fit')
    plt.xlabel("y_true")
    plt.ylabel("y_pred")
    plt.title("Test: y_true vs y_pred")
    plt.legend()
    plt.tight_layout()
    plt.savefig(scatter_png, dpi=200)
    plt.close()
    print(f"[PLOT] saved: {scatter_png}")

    # Sorted curves
    curves_png = os.path.join(run_dir, "curves_sorted.png")
    order = np.argsort(y_true)
    plt.figure(figsize=(10, 6))
    plt.plot(y_true[order], label="y_true", alpha=0.7)
    plt.plot(y_pred[order], label="y_pred", alpha=0.7)
    plt.xlabel("samples (sorted by y_true)")
    plt.ylabel("value")
    plt.title("Test: curves (sorted)")
    plt.legend()
    plt.tight_layout()
    plt.savefig(curves_png, dpi=200)
    plt.close()
    print(f"[PLOT] saved: {curves_png}")

    # Residuals
    res_png = os.path.join(run_dir, "residuals.png")
    res = y_pred - y_true
    plt.figure(figsize=(8, 6))
    plt.scatter(y_true, res, s=8, alpha=0.5)
    plt.axhline(0, color='r', linestyle='--')
    plt.xlabel("y_true")
    plt.ylabel("y_pred - y_true")
    plt.title("Test: residuals")
    plt.tight_layout()
    plt.savefig(res_png, dpi=200)
    plt.close()
    print(f"[PLOT] saved: {res_png}")

    print("\n[DONE]")


if __name__ == "__main__":
    main()