run.py

import argparse
import numpy as np
import pandas as pd
from datetime import datetime
from tqdm import trange
import torch
import torch.nn as nn
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.compose import make_column_selector, ColumnTransformer
import wandb
import torchtuples as tt

# models
from sksurv.ensemble import ComponentwiseGradientBoostingSurvivalAnalysis
from lifelines.fitters.weibull_aft_fitter import WeibullAFTFitter
from model import CoxPH, MTLR, CQRNN, LogNormalNN
from pycox.models import DeepHitSingle, CoxTime
from pycox.models.cox_time import MLPVanillaCoxTime

# Conformality
from icp import ConformalSurvDist, CSDiPOT
from icp.scorer import QuantileRegressionNC, SurvivalPredictionNC

from utils import save_params, set_seed, print_performance
from utils.util_survival import survival_data_split, xcal_from_hist, make_time_bins
from args import generate_parser
from data import make_survival_data
from data.cond_features import get_cond_functions
from SurvivalEVAL import QuantileRegEvaluator
from CondCalEvaluation import cond_xcal, wsc_xcal


def main(args=None):
    if isinstance(args, argparse.Namespace):
        wandb.init(
            project="conditionalCSD",
            config=args,
            name=args.model + "_" + args.data
        )
    else:
        wandb.init(config=args)
    wandb.define_metric("C-index", summary="mean")
    wandb.define_metric("IBS", summary="mean")
    wandb.define_metric("MAE_Hinge", summary="mean")
    wandb.define_metric("MAE_PO", summary="mean")
    wandb.define_metric("KM-cal", summary="mean")
    wandb.define_metric("X-cal", summary="mean")
    wandb.define_metric("cond_xcal", summary="mean")
    wandb.define_metric("wsc_xcal", summary="mean")
    wandb.define_metric("train_time", summary="mean")
    wandb.define_metric("infer_time", summary="mean")

    args = wandb.config
    data = make_survival_data(args.data)
    conditions = get_cond_functions(args.data)
    features = data.columns.to_list()
    assert "time" in data.columns and "event" in data.columns, "The event time variable and censor indicator " \
                                                               "variable is missing or need to be renamed."
    # Continuous features: median imputation + scaling
    enc_con = Pipeline(steps=[
        ('impute', SimpleImputer(strategy='median')),
        ('scale', StandardScaler()),
    ])
    sel_con = make_column_selector(pattern='^num_')

    # Non-continuous features: mode imputation only
    enc_cat = Pipeline(steps=[
        ('impute', SimpleImputer(strategy='most_frequent')),
    ])
    sel_cat = make_column_selector(
        pattern='^(?!num_|time$|event$).*'  # everything NOT num_, time, or event
    )

    enc_df = ColumnTransformer(
        transformers=[
            ('num', enc_con, sel_con),
            ('cat', enc_cat, sel_cat),
        ],
        remainder='passthrough',
        verbose_feature_names_out=False
    )
    enc_df.set_output(transform='pandas')

    if args.early_stop:
        pct_train = 0.8
        pct_val = 0.1
        pct_test = 0.1
    else:
        pct_train = 0.9
        pct_val = 0.0
        pct_test = 0.1

    args.n_features = len(features) - 2     # excluding time and event
    args.timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    args.device = "cuda:0" if torch.cuda.is_available() else "cpu"
    device = torch.device(args.device)

    path = save_params(args)

    ci = []
    mae_hinge = []
    mae_po = []
    ibs = []
    km_cal = []
    xcal_stats = []
    cond_xcal_stats = []
    wsc_xcal_stats = []
    train_times = []
    infer_times = []

    pbar_outer = trange(args.n_exp, disable=not args.verbose, desc='Experiment')
    for i in pbar_outer:
        seed_i = args.seed + i
        set_seed(seed_i, device)
        data_train, data_val, data_test = survival_data_split(data, stratify_colname='both', frac_train=pct_train,
                                                              frac_val=pct_val, frac_test=pct_test, random_state=seed_i)
        # standardize the data
        data_train = enc_df.fit_transform(data_train).astype('float32')
        data_val = enc_df.transform(data_val).astype('float32') if not data_val.empty else data_val
        data_test = enc_df.transform(data_test).astype('float32')

        # get the labels for evaluation
        t_train, e_train = data_train["time"].values, data_train["event"].values
        t_val, e_val = data_val["time"].values, data_val["event"].values if not data_val.empty else None
        x_test = data_test.drop(['time', 'event'], axis=1).values
        t_test, e_test = data_test["time"].values, data_test["event"].values
        t_train_val = np.concatenate((t_train, t_val)) if not data_val.empty else t_train
        e_train_val = np.concatenate((e_train, e_val)) if not data_val.empty else e_train

        # this is make sure MTLR and DeepHit have the same number of bins, but the bins locations are different
        # -- MTLR uses the uniformly-divided quantiles, while DeepHit uses the uniformly-divided times.
        if args.model in ["MTLR", "DeepHit"]:
            discrete_bins = make_time_bins(t_train, event=e_train)
            if args.model == "DeepHit":
                # the first bin of DeepHit must be smaller than the smallest time in the data
                discrete_bins[0] = max(t_train_val.min() - 1e-5, 0)

        if args.model == "CoxPH":
            model = CoxPH(
                n_features=args.n_features,
                hidden_size=args.neurons,
                norm=args.norm,
                activation=args.activation,
                dropout=args.dropout
            )
        elif args.model == "MTLR":
            model = MTLR(
                n_features=args.n_features,
                time_bins=discrete_bins,
                hidden_size=args.neurons,
                norm=args.norm,
                activation=args.activation,
                dropout=args.dropout
            )
        elif args.model == "CQRNN":
            model = CQRNN(
                n_features=args.n_features,
                hidden_size=args.neurons,
                n_quantiles=args.n_quantiles,
                norm=args.norm,
                activation=args.activation,
                dropout=args.dropout
            )
        elif args.model == "LogNormalNN":
            model = LogNormalNN(
                n_features=args.n_features,
                hidden_size=args.neurons,
                norm=args.norm,
                activation=args.activation,
                dropout=args.dropout,
                lam=args.lam
            )
        elif args.model == "DeepHit":
            labtrans = DeepHitSingle.label_transform(discrete_bins.numpy())
            net = tt.practical.MLPVanilla(in_features=args.n_features, num_nodes=args.neurons,
                                          out_features=labtrans.out_features, batch_norm=args.norm,
                                          dropout=args.dropout, activation=getattr(nn, args.activation))
            model = DeepHitSingle(net, tt.optim.Adam, device=args.device, alpha=0.2, sigma=0.1,
                                  duration_index=labtrans.cuts)
            model.label_transform = labtrans
        elif args.model == "CoxTime":
            labtrans = CoxTime.label_transform()
            labtrans.fit(t_train, e_train)
            net = MLPVanillaCoxTime(in_features=args.n_features, num_nodes=args.neurons, batch_norm=args.norm,
                                    dropout=args.dropout, activation=getattr(nn, args.activation))
            model = CoxTime(net, tt.optim.Adam, device=args.device, labtrans=labtrans)
            model.label_transform = labtrans
        elif args.model == "GB":
            model = ComponentwiseGradientBoostingSurvivalAnalysis(loss='coxph', n_estimators=100, random_state=seed_i)
        elif args.model == "AFT":
            model = WeibullAFTFitter(penalizer=0.01)
        else:
            raise ValueError(f"Unknown model name: {args.model}")

        if args.post_process == "CSD":
            nc_model = QuantileRegressionNC(model, args)
            icp = ConformalSurvDist(nc_model, condition=None, decensor_method=args.decensor_method,
                                    n_quantiles=args.n_quantiles)
        elif args.post_process == "CSD-iPOT":
            nc_model = SurvivalPredictionNC(model, args)
            icp = CSDiPOT(nc_model, decensor_method=args.decensor_method, n_percentile=args.n_quantiles)
        else:
            raise ValueError(f"Unknown post-processing method: {args.post_process}")

        # Fit the ICP using the proper training set, and using valset for early stopping
        start_time = datetime.now()
        icp.fit(data_train, data_val)
        # Calibrate the ICP using the calibration set
        if args.use_train:
            data_val = pd.concat([data_train, data_val], ignore_index=True)
        icp.calibrate(data_val)
        mid_time = datetime.now()
        # Produce predictions for the test set
        quan_levels, quan_preds = icp.predict(x_test)
        end_time = datetime.now()

        train_time = (mid_time - start_time).total_seconds()
        infer_time = (end_time - mid_time).total_seconds()

        # evaluate the performance
        evaler = QuantileRegEvaluator(quan_preds, quan_levels, t_test, e_test, t_train_val, e_train_val,
                                      predict_time_method="Median", interpolation='Pchip')
        c_index = evaler.concordance(ties="All")[0]
        ibs_score = evaler.integrated_brier_score(num_points=10)
        hinge_abs = evaler.mae(method='Hinge', verbose=False, weighted=True)
        po_abs = evaler.mae(method='Pseudo_obs', verbose=False, weighted=True)
        km_cal_score = evaler.km_calibration()
        _, dcal_hist = evaler.d_calibration()
        xcal_score = xcal_from_hist(dcal_hist)
        pred_probs = evaler.predict_probability_from_curve(evaler.event_times)
        cond_xcal_score = cond_xcal(x_test, e_test, pred_probs, conditions)
        if data.shape[0] >= 1000:
            wsc_xcal_score = wsc_xcal(x_test, e_test, pred_probs, random_state=seed_i)
        else:
            wsc_xcal_score = 0  # not enough data to compute the WSC

        ci.append(c_index)
        ibs.append(ibs_score)
        mae_hinge.append(hinge_abs)
        mae_po.append(po_abs)
        km_cal.append(km_cal_score)
        xcal_stats.append(xcal_score)
        cond_xcal_stats.append(cond_xcal_score)
        wsc_xcal_stats.append(wsc_xcal_score)
        train_times.append(train_time)
        infer_times.append(infer_time)

        wandb.log({'C-index': c_index,
                   'IBS': ibs_score,
                   'MAE_Hinge': hinge_abs,
                   'MAE_PO': po_abs,
                   'KM-cal': km_cal_score,
                   'X-cal': xcal_score,
                   'cond_xcal': cond_xcal_score,
                   'wsc_xcal': wsc_xcal_score,
                   'train_time': train_time,
                   'infer_time': infer_time
                   })
    print_performance(
        path=path,
        Cindex=ci,
        IBS=ibs,
        MAE_Hinge=mae_hinge,
        MAE_PO=mae_po,
        KM_cal=km_cal,
        xCal_stats=xcal_stats,
        cond_xCal_stats=cond_xcal_stats,
        wsc_xCal_stats=wsc_xcal_stats,
        train_times=train_times,
        infer_times=infer_times
    )


if __name__ == '__main__':
    # enable for debugging
    # torch.autograd.set_detect_anomaly(True)

    args = generate_parser()
    main(args)
    wandb.finish()