predict_eao_ensemble.py

"""
Inference + Codabench prediction.zip generator.

For each disease target:
  - Load all best_*.pth and swa_*.pth checkpoints in the per-target results dir
  - Run each on the {split} embeddings of that target
  - Average softmax probabilities across checkpoints (multi-seed/multi-LR/SWA ensemble)
  - Write {split}_per_sample_predictions.csv in the format the organizer's
    cvpr26_organize_eval_metrics_and_predictions.py expects

Then concatenate all targets into predictions.csv and zip → prediction.zip for
direct Codabench submission.
"""
import argparse
import os
import sys
import zipfile

import h5py
import numpy as np
import pandas as pd
import torch
from torch.utils.data import DataLoader, Dataset

THIS = os.path.dirname(os.path.abspath(__file__))
ROOT = os.path.abspath(os.path.join(THIS, ".."))
sys.path.insert(0, os.path.join(ROOT, "starter"))
from models.attention_pooling_multilayers import MultiLayersCrossAttentionPooling  # noqa: E402

# Per-disease decision thresholds, derived from val labels on the v4 multi-seed
# ensemble (threshold_optimize_v2 unique-prob sweep). Baked in so this single
# script reproduces the 0.7086 BalAcc result without needing a follow-up step.
# Disease not in this dict falls back to 0.5.
THRESHOLDS = {
    "hydronephrosis":         0.7685199,
    "lymphadenopathy":        0.5737428,
    "kidney_stone":           0.80407256,
    "covid":                  0.6222638,
    "gallstone":              0.7481811,
    "liver_calcifications":   0.64198047,
    "colorectal_cancer":      0.35786006,
    "liver_lesion":           0.79084086,
    "renal_cyst":             0.10136525,
    "liver_cyst":             0.11666061,
    "adrenal_hyperplasia":    0.5463961,
    "splenomegaly":           0.37268373,
    "lung_nodule_malignancy": 0.44977823,
    "cholecystitis":          0.52176595,
    "atherosclerosis":        0.5064166,
    "fatty_liver":            0.48598397,
    "ascites":                0.5023216,
}


class SpatialFeaturesDataset(Dataset):
    def __init__(self, embeds_dir, csv_path, split, target_column):
        df = pd.read_csv(csv_path)
        split_df = df[df["split"] == split].copy()
        self.paths, self.label_mapping = [], {}
        for _, row in split_df.iterrows():
            case_id = str(row["case_id"])
            base = case_id.split(".nii.gz")[0] if ".nii.gz" in case_id else case_id
            base = base.replace(".h5", "")
            path = os.path.join(embeds_dir, base + ".h5")
            if os.path.exists(path):
                self.paths.append(path)
                self.label_mapping[base] = int(row[target_column])

    def __len__(self):
        return len(self.paths)

    def __getitem__(self, i):
        path = self.paths[i]
        base = os.path.basename(path).replace(".h5", "")
        with h5py.File(path, "r") as hf:
            x = torch.tensor(hf["y_hat"][:]).float()
        return x, torch.tensor(self.label_mapping[base]).long(), base


def discover_target_dirs(results_root):
    """Find target subdirs that contain at least one .pth checkpoint."""
    out = []
    for name in sorted(os.listdir(results_root)):
        d = os.path.join(results_root, name)
        if not os.path.isdir(d):
            continue
        if any(f.endswith(".pth") for f in os.listdir(d)):
            out.append(name)
    return out


def parse_head_hparams(ckpt):
    """The state dict keys look like `heads.head_lr_1e_03.<...>`. We rebuild
    a head with the same architecture as training (defaults match starter)."""
    sd = ckpt["state_dict"]
    # Strip "heads.<head_name>." prefix and detect dimensions
    stripped = {}
    for k, v in sd.items():
        if k.startswith("heads."):
            parts = k.split(".", 2)
            if len(parts) >= 3:
                stripped[parts[2]] = v

    cls_w = stripped.get("classifier.weight")
    cq = stripped.get("class_query")
    if cls_w is None or cq is None:
        raise RuntimeError(f"Checkpoint missing classifier/class_query: keys={list(stripped.keys())[:5]}")
    num_classes, q_times_d = cls_w.shape
    query_num, embed_dim = cq.shape
    assert q_times_d == query_num * embed_dim, (
        f"Mismatch: classifier in_features={q_times_d} vs query_num*embed_dim={query_num*embed_dim}"
    )
    # num_layers = number of cross-attention layers we can find in the keys
    num_layers = 1 + max(
        (int(k.split(".")[1]) for k in stripped.keys() if k.startswith("layers.")),
        default=-1,
    )
    if num_layers < 1:
        num_layers = 2  # fall back to starter default
    return stripped, dict(
        embed_dim=embed_dim, query_num=query_num, num_classes=num_classes,
        num_layers=num_layers, num_heads=4, dropout=0.0, ffn_mult=1,
    )


def load_head(ckpt_path, device):
    ckpt = torch.load(ckpt_path, map_location="cpu")
    stripped, hp = parse_head_hparams(ckpt)
    head = MultiLayersCrossAttentionPooling(**hp)
    head.load_state_dict(stripped, strict=True)
    head.to(device).eval()
    return head, hp


@torch.no_grad()
def predict_one_head(head, loader, device):
    all_probs, all_labels, all_filenames = [], [], []
    for xb, yb, fns in loader:
        xb = xb.to(device)
        logits = head(xb)
        probs = torch.softmax(logits, dim=1).cpu()
        all_probs.append(probs)
        all_labels.append(yb)
        all_filenames.extend(list(fns))
    return torch.cat(all_probs), torch.cat(all_labels), all_filenames


def write_per_sample_csv(probs_avg, labels, filenames, out_path, threshold=None):
    """Format expected by the organizer's cvpr26_organize_eval_metrics_and_predictions.py:
       columns = filename, label, prediction, logit_class_0..C-1, prob_class_0..C-1

    If `threshold` is provided and the head is binary (num_classes==2), use
    prob_class_1 >= threshold for the prediction. Otherwise fall back to argmax.
    """
    num_classes = probs_avg.shape[1]
    if threshold is not None and num_classes == 2:
        preds = (probs_avg[:, 1] >= float(threshold)).long()
    else:
        preds = probs_avg.argmax(1)
    # We didn't track raw logits across the ensemble; use log-prob as a stand-in
    # (the organizer's metrics never read these — only label/prediction/probs).
    log_probs = torch.log(probs_avg.clamp_min(1e-12))
    cols = {"filename": filenames, "label": labels.numpy(), "prediction": preds.numpy()}
    for c in range(num_classes):
        cols[f"logit_class_{c}"] = log_probs[:, c].numpy()
    for c in range(num_classes):
        cols[f"prob_class_{c}"] = probs_avg[:, c].numpy()
    df = pd.DataFrame(cols)
    os.makedirs(os.path.dirname(out_path), exist_ok=True)
    df.to_csv(out_path, index=False)
    return df


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--embeds_root", required=True,
                    help="Root with {target}/embeddings/ subdirs")
    ap.add_argument("--labels_root", required=True,
                    help="Dir with {target}.csv label files")
    ap.add_argument("--results_root", required=True,
                    help="Dir with {target}/ subdirs containing .pth ckpts (output of run_EAO_improved.py)")
    ap.add_argument("--split", default="val", choices=["train", "val", "test"])
    ap.add_argument("--out_zip", default=None,
                    help="Where to write the final prediction.zip (default: results_root/prediction.zip)")
    ap.add_argument("--batch_size", type=int, default=64)
    ap.add_argument("--num_workers", type=int, default=2)
    ap.add_argument("--targets", nargs="*", default=None,
                    help="Subset to predict (default: all subdirs with checkpoints)")
    ap.add_argument("--top_k_ckpts", type=int, default=0,
                    help="If >0, only use top-K checkpoints per target by filename score")
    args = ap.parse_args()

    device = "cuda" if torch.cuda.is_available() else "cpu"
    targets = args.targets or discover_target_dirs(args.results_root)
    if not targets:
        raise SystemExit(f"No target subdirs with .pth found in {args.results_root}")

    aggregate_dfs = []
    for target in targets:
        ck_dir = os.path.join(args.results_root, target)
        ckpts = sorted([f for f in os.listdir(ck_dir) if f.endswith(".pth")])
        if not ckpts:
            print(f"[skip] {target}: no checkpoints")
            continue
        if args.top_k_ckpts > 0:
            # Score from filename: best_*acc{score}_*.pth
            def score_of(fn):
                for tag in ("balanced_acc", "auroc"):
                    if tag in fn:
                        try:
                            return float(fn.split(tag)[1].split("_")[0])
                        except Exception:
                            pass
                return -1.0
            ckpts = sorted(ckpts, key=score_of, reverse=True)[: args.top_k_ckpts]

        embeds_dir = os.path.join(args.embeds_root, target, "embeddings")
        labels_csv = os.path.join(args.labels_root, target + ".csv")
        if not os.path.isdir(embeds_dir):
            print(f"[skip] {target}: missing {embeds_dir}")
            continue
        if not os.path.exists(labels_csv):
            print(f"[skip] {target}: missing {labels_csv}")
            continue

        df = pd.read_csv(labels_csv)
        # Only the target column; if the CSV uses a slightly different name, infer it.
        if target not in df.columns:
            cand = [c for c in df.columns if c not in ("case_id", "split")]
            if len(cand) != 1:
                raise RuntimeError(f"Cannot infer target col for {target}: {df.columns.tolist()}")
            target_col = cand[0]
        else:
            target_col = target
        ds = SpatialFeaturesDataset(embeds_dir, labels_csv, args.split, target_col)
        if len(ds) == 0:
            print(f"[skip] {target}: empty {args.split} split (no .h5 files matched)")
            continue
        loader = DataLoader(ds, batch_size=args.batch_size, shuffle=False,
                            num_workers=args.num_workers, pin_memory=True)

        # Average probs across all selected checkpoints
        probs_sum = None
        labels_keep, filenames_keep = None, None
        for ck in ckpts:
            head, hp = load_head(os.path.join(ck_dir, ck), device)
            probs, labels, filenames = predict_one_head(head, loader, device)
            if probs_sum is None:
                probs_sum = probs
                labels_keep, filenames_keep = labels, filenames
            else:
                probs_sum = probs_sum + probs
        probs_avg = probs_sum / len(ckpts)

        thr = THRESHOLDS.get(target, 0.5)
        out_csv = os.path.join(ck_dir, f"{args.split}_per_sample_predictions.csv")
        df_out = write_per_sample_csv(probs_avg, labels_keep, filenames_keep, out_csv, threshold=thr)
        df_out["disease_name"] = target
        # Quick val metric for reporting
        from sklearn.metrics import balanced_accuracy_score, roc_auc_score
        try:
            bal = balanced_accuracy_score(df_out["label"], df_out["prediction"])
        except Exception:
            bal = float("nan")
        try:
            auroc = roc_auc_score(df_out["label"], df_out["prob_class_1"])
        except Exception:
            auroc = float("nan")
        print(f"[{target}]  ckpts={len(ckpts)}  n={len(df_out)}  bal_acc={bal:.4f}  auroc={auroc:.4f}  thr={thr:.4f}")
        aggregate_dfs.append(df_out)

    if not aggregate_dfs:
        raise SystemExit("No predictions written.")
    df_all = pd.concat(aggregate_dfs, ignore_index=True)

    # Write aggregated predictions.csv + zip it
    pred_csv = os.path.join(args.results_root, "predictions.csv")
    df_all.to_csv(pred_csv, index=False)
    out_zip = args.out_zip or os.path.join(args.results_root, "prediction.zip")
    with zipfile.ZipFile(out_zip, "w", zipfile.ZIP_DEFLATED) as zf:
        zf.write(pred_csv, arcname="predictions.csv")
    print(f"\nWrote {pred_csv}  ({len(df_all)} rows, {df_all['disease_name'].nunique()} diseases)")
    print(f"Wrote {out_zip}")


if __name__ == "__main__":
    main()