Files
hypertower/scripts/output_analysis/visualizations/rebuild_run_best_plots.py
T
2026-03-04 09:40:09 +01:00

533 lines
20 KiB
Python
Executable File

#!/usr/bin/env python3
"""
Recompute per-fold ROC plots for a completed multifold run using the saved
best checkpoints instead of the final epoch.
Example:
python scripts/rebuild_run_best_plots.py \
--run-dir analysis_data/1029_Baseline_Balanced_Resnet/1029_Baseline_Balanced_Resnet_20251029_163906 \
--head image
"""
import argparse
import logging
import json
from pathlib import Path
from types import SimpleNamespace
import matplotlib
import sys
REPO_ROOT = Path(__file__).resolve().parents[3]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
matplotlib.use("Agg")
import matplotlib.pyplot as plt # noqa: E402
import numpy as np # noqa: E402
import pandas as pd # noqa: E402
import torch # noqa: E402
from sklearn.metrics import auc, roc_auc_score, roc_curve # noqa: E402
from classes import build_papila_clinical # noqa: E402
from classes.hypertower import HyperTower # noqa: E402
try: # Allow checkpoints that stored pandas DataFrames in their args.
from torch.serialization import add_safe_globals # type: ignore
add_safe_globals([pd.DataFrame])
except (ImportError, AttributeError):
pass
def parse_args() -> argparse.Namespace:
ap = argparse.ArgumentParser(description="Rebuild ROC plots for an existing multifold run.")
ap.add_argument("--run-dir", required=True, type=Path, help="Path to the run directory under analysis_data.")
ap.add_argument("--head", default="image", choices=["image", "fused", "metadata"], help="Which prediction head to plot.")
ap.add_argument("--class-names", nargs="*", default=None, help="Optional class names to control plot labels.")
ap.add_argument("--overwrite", action="store_true", help="Overwrite existing .npy probability dumps if present.")
ap.add_argument(
"--use-holdout",
action="store_true",
help="Evaluate checkpoints on the saved holdout set instead of the fold validation splits.",
)
return ap.parse_args()
def load_cli_args(run_dir: Path) -> dict:
cli_path = run_dir / "cli_args.json"
if not cli_path.exists():
raise FileNotFoundError(f"Missing cli_args.json in {run_dir}")
with cli_path.open("r", encoding="utf-8") as fh:
return json.load(fh)
def load_summary(run_dir: Path) -> dict:
summary_path = run_dir / "summary.json"
if not summary_path.exists():
raise FileNotFoundError(f"Missing summary.json in {run_dir}")
with summary_path.open("r", encoding="utf-8") as fh:
return json.load(fh)
def resolve_data_dir(raw_dir: str) -> str:
"""
Resolve dataset paths saved in legacy cli_args.json.
Older runs often store "ClinicalData"/"FundusImages" relative to a
dataset root, while current repo layout uses "Papila/<dir>".
"""
p = Path(raw_dir)
if p.exists():
return str(p)
candidates = [
REPO_ROOT / p,
REPO_ROOT / "Papila" / p,
]
for cand in candidates:
if cand.exists():
return str(cand)
return str(p)
def prepare_clinical(cli_args: dict, run_dir: Path) -> tuple:
image_dir = resolve_data_dir(cli_args["image_dir"])
clinical_dir = resolve_data_dir(cli_args["clinical_dir"])
clinical = build_papila_clinical(
image_dir,
clinical_dir,
cli_args["label_col"],
cli_args["cat_cols"],
n_splits=cli_args["n_splits"],
random_seed=cli_args["fold_seed"],
)
holdout_path = run_dir / "holdout.csv"
holdout_df = pd.read_csv(holdout_path) if holdout_path.exists() else None
if holdout_df is not None:
if cli_args["eval_mode"] == "binary":
holdout_df = holdout_df[holdout_df[cli_args["label_col"]].isin([0, 1])].reset_index(drop=True)
join_cols = [c for c in holdout_df.columns if c in clinical.df.columns]
if not join_cols:
raise RuntimeError("Holdout CSV found but no overlapping columns with clinical dataframe.")
marker = holdout_df.assign(_holdout_marker=1)
merged = clinical.df.merge(marker, on=join_cols, how="left")
train_df = merged[merged["_holdout_marker"].isna()].drop(columns=["_holdout_marker"]).reset_index(drop=True)
clinical.frames = [train_df.copy()]
clinical.df = train_df.copy()
clinical._infer_or_validate_feature_types()
clinical._compute_numeric_stats()
clinical._build_cat_maps()
clinical._compute_feature_dim()
clinical._build_kfold_indices()
return clinical, holdout_df
def build_ht_args(cli_args: dict, fold: int, run_dir: Path, models_dir: Path, holdout_df):
image_dir = resolve_data_dir(cli_args["image_dir"])
clinical_dir = resolve_data_dir(cli_args["clinical_dir"])
# Copy of the training-time namespace so HyperTower can be re-instantiated.
return SimpleNamespace(
image_dir=image_dir,
clinical_dir=clinical_dir,
label_col=cli_args["label_col"],
cat_cols=cli_args["cat_cols"],
batch_size=cli_args["batch_size"],
epochs=cli_args["epochs"],
lr=cli_args["lr"],
num_classes=cli_args["num_classes"],
img_augment=cli_args.get("img_augment", True),
focal_gamma=cli_args.get("focal_gamma", 0.0),
eval_mode=cli_args["eval_mode"],
fold=fold,
run_dir=str(run_dir),
models_dir=str(models_dir),
backbone=cli_args["backbone"],
freeze_ratio=cli_args["freeze_ratio"],
fusion_mode=cli_args["fusion_mode"],
use_se=cli_args.get("use_se", True),
se_reduction=cli_args.get("se_reduction", 16),
se_pre_norm=cli_args.get("se_pre_norm", True),
se_where=cli_args.get("se_where", "bridge"),
se_reduction_tower=cli_args.get("se_reduction_tower", cli_args.get("se_reduction", 16)),
se_pre_norm_tower=cli_args.get("se_pre_norm_tower", cli_args.get("se_pre_norm", True)),
warmup_tower_epochs=cli_args.get("warmup_tower_epochs", 0),
warmup_fused_epochs=cli_args.get("warmup_fused_epochs", 0),
gradual_thaw=cli_args.get("gradual_thaw", False),
thaw_phase_duration=cli_args.get("thaw_phase_duration", 5),
thaw_ratio=cli_args.get("thaw_ratio", 0.33),
thaw_target=cli_args.get("thaw_target", "image"),
thaw_start_epoch=cli_args.get("thaw_start_epoch", -1),
initial_freeze=cli_args.get("initial_freeze", False),
bcd_prob=0.5,
bcd_p0=0.20,
bcd_min=0.05,
bcd_max=0.30,
bcd_k=0.4,
bcd_metric="auc",
bcd_alpha_batch=0.2,
bcd_alpha_tower=0.3,
bcd_explore_floor=0.15,
aux_img=0.05,
aux_md=0.05,
aux_detach=True,
ema_alpha=0.9,
entropy_ema=0.7,
early_stop=cli_args.get("early_stop", False),
early_metric=cli_args.get("early_metric"),
early_mode=cli_args.get("early_mode", "auto"),
early_patience=cli_args.get("early_patience", 7),
early_min_delta=cli_args.get("early_min_delta", 0.0),
checkpoint_best=cli_args.get("checkpoint_best", False),
holdout_df=holdout_df,
img_crop_manifest=cli_args.get("img_crop_manifest"),
img_crop_weights=cli_args.get("img_crop_weights"),
img_crop_normalize=cli_args.get("img_crop_normalize"),
img_crop_threshold=cli_args.get("img_crop_threshold"),
img_crop_scale=cli_args.get("img_crop_scale", 2.5),
img_crop_size=cli_args.get("img_crop_size", 224),
img_crop_cache=cli_args.get("img_crop_cache"),
img_crop_tta=cli_args.get("img_crop_tta", False),
img_crop_gt=cli_args.get("img_crop_gt", False),
img_geometry_features=cli_args.get("img_geometry_features", False),
balanced_sampler=cli_args.get("balanced_sampler", False),
)
def collect_logits(ht, loader):
"""Mirror Multifold.eval_collect_logits but for a provided loader."""
device = ht.device
ht.img_tower.eval()
ht.md_tower.eval()
outputs = []
with torch.no_grad():
if ht.mode == "vote":
ht.head_img.eval()
ht.head_md.eval()
ht.vote.eval()
else:
ht.bridge.eval()
for batch in loader:
if len(batch) == 4:
imgs, metas, geometry, labels = batch
else:
imgs, metas, labels = batch
geometry = None
imgs = imgs.to(device)
metas = metas.to(device)
labels = labels.to(device)
if geometry is not None and geometry.numel() > 0:
geometry = geometry.to(device)
else:
geometry = None
if ht.mode == "vote":
img_feats = ht.img_tower(imgs, geometry)
md_feats = ht.md_tower(metas)
out_img = ht.head_img(img_feats)
out_md = ht.head_md(md_feats)
out_fused = ht.vote(out_img, out_md)
else:
img_feats = ht.img_tower(imgs, geometry)
md_feats = ht.md_tower(metas)
result = ht.bridge(img_feats, md_feats)
if isinstance(result, tuple):
out_fused, out_img, out_md = result
else:
out_fused, out_img, out_md = result, None, None
outputs.append(
(
labels.detach().cpu().numpy(),
torch.softmax(out_fused, dim=1).detach().cpu().numpy() if out_fused is not None else None,
torch.softmax(out_img, dim=1).detach().cpu().numpy() if out_img is not None else None,
torch.softmax(out_md, dim=1).detach().cpu().numpy() if out_md is not None else None,
)
)
if not outputs:
return np.array([]), None, None, None
y_all, pf, pi, pm = zip(*outputs)
y_true = np.concatenate(y_all, axis=0)
probs_f = np.concatenate([p for p in pf if p is not None], axis=0) if any(p is not None for p in pf) else None
probs_i = np.concatenate([p for p in pi if p is not None], axis=0) if any(p is not None for p in pi) else None
probs_m = np.concatenate([p for p in pm if p is not None], axis=0) if any(p is not None for p in pm) else None
return y_true, probs_f, probs_i, probs_m
def compute_per_class_curves(y_true, probs):
if probs is None:
return {}
num_classes = probs.shape[1]
curves = {}
for k in range(num_classes):
y_bin = (y_true == k).astype(np.uint8)
fpr, tpr, _ = roc_curve(y_bin, probs[:, k])
curves[k] = {"fpr": fpr, "tpr": tpr, "auc": auc(fpr, tpr) if len(fpr) > 1 else np.nan}
return curves
def choose_head_probs(head: str, probs_f, probs_i, probs_m):
if head == "fused":
return probs_f
if head == "metadata":
return probs_m
return probs_i
def legacy_probs_suffix(head: str) -> str:
if head == "image":
return "img"
if head == "metadata":
return "md"
return "fused"
def ensure_binary_slice(y_true, *arrays):
mask = np.isin(y_true, [0, 1])
filtered = [y_true[mask]]
for arr in arrays:
if arr is None:
filtered.append(None)
else:
filtered.append(arr[mask])
return filtered
def plot_overlays(
per_fold_curves,
out_dir: Path,
class_names: list[str],
head: str,
eval_mode: str,
suffix: str = "",
):
keys = sorted({k for _, curves in per_fold_curves for k in curves.keys()})
if eval_mode == "binary":
keys = [k for k in keys if k == 1]
if not keys:
return
name_map = {k: (class_names[k] if k < len(class_names) else f"class_{k}") for k in keys}
out_dir.mkdir(parents=True, exist_ok=True)
for k in keys:
fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(111)
ax.plot([0, 1], [0, 1], linestyle="--", linewidth=1, color="grey")
for fold_idx, curves in per_fold_curves:
if k not in curves:
continue
fpr = curves[k]["fpr"]
tpr = curves[k]["tpr"]
auc_val = curves[k]["auc"]
label = f"Fold {fold_idx} (AUC {auc_val:.3f})" if auc_val == auc_val else f"Fold {fold_idx}"
ax.plot(fpr, tpr, linewidth=1.5, label=label)
ax.set_xlabel("False Positive Rate")
ax.set_ylabel("True Positive Rate")
ax.set_title(f"{head} head — {name_map[k]} ROC per fold")
ax.legend(loc="lower right")
fig.tight_layout()
safe_name = name_map[k].replace(" ", "_")
suffix_str = suffix if suffix else ""
fig.savefig(out_dir / f"roc_{head}_{safe_name}_perfold{suffix_str}.png", dpi=160)
plt.close(fig)
def plot_mean_sd(
per_fold_curves,
out_dir: Path,
class_names: list[str],
head: str,
eval_mode: str,
suffix: str = "",
):
keys = sorted({k for _, curves in per_fold_curves for k in curves.keys()})
if eval_mode == "binary":
keys = [k for k in keys if k == 1]
if not keys:
return
grid = np.linspace(0, 1, 501)
fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(111)
ax.plot([0, 1], [0, 1], linestyle="--", linewidth=1, color="grey")
for k in keys:
tprs = []
aucs = []
for _, curves in per_fold_curves:
if k not in curves:
continue
fpr = curves[k]["fpr"]
tpr = curves[k]["tpr"]
aucs.append(curves[k]["auc"])
tprs.append(np.interp(grid, fpr, tpr))
if not tprs:
continue
tprs = np.vstack(tprs)
mean = tprs.mean(axis=0)
std = tprs.std(axis=0)
label = class_names[k] if k < len(class_names) else f"class_{k}"
label = f"{label} (AUC {np.nanmean(aucs):.3f}±{np.nanstd(aucs):.3f})"
ax.plot(grid, mean, linewidth=2, label=label)
ax.fill_between(grid, np.maximum(mean - std, 0), np.minimum(mean + std, 1), alpha=0.15)
ax.set_xlabel("False Positive Rate")
ax.set_ylabel("True Positive Rate")
ax.set_title(f"Mean OVR ROC (±1 SD) — {head} head")
ax.legend(loc="lower right")
fig.tight_layout()
suffix_str = suffix if suffix else ""
out_dir.mkdir(parents=True, exist_ok=True)
fig.savefig(out_dir / f"roc_{head}_mean_ovr{suffix_str}.png", dpi=160)
plt.close(fig)
def main():
args = parse_args()
run_dir = args.run_dir.resolve()
cli_args = load_cli_args(run_dir)
summary = load_summary(run_dir)
class_names = (
args.class_names
if args.class_names
else (cli_args.get("class_names") or (["Healthy", "Glaucoma"] if cli_args["eval_mode"] == "binary" else ["Healthy", "Glaucoma", "Suspect"]))
)
shortname = cli_args.get("shortname") or run_dir.parent.name
run_id = cli_args.get("run_id") or run_dir.name
base_models_dir = Path("models") / shortname / run_id
clinical, holdout_df = prepare_clinical(cli_args, run_dir)
if args.use_holdout and holdout_df is None:
raise SystemExit("Holdout metrics requested but no holdout.csv found for this run.")
per_fold_curves = []
fold_aucs = []
head = args.head
file_suffix = "_holdout" if args.use_holdout else ""
for fold_entry in summary.get("fold_metrics", []):
fold_idx = int(fold_entry["fold"])
best_epoch = fold_entry.get("best_epoch")
if not best_epoch:
print(f"[skip] Fold {fold_idx}: no best_epoch recorded.")
continue
# Prefer saved fold arrays when available. This avoids reconstructing
# HyperTower for legacy runs whose external weight paths no longer exist.
base = run_dir / f"fold{fold_idx}{file_suffix}"
y_path = Path(f"{base}_y_true.npy")
p_path = Path(f"{base}_probs_{legacy_probs_suffix(head)}.npy")
if y_path.exists() and p_path.exists():
y_true = np.load(y_path)
head_probs = np.load(p_path)
if cli_args["eval_mode"] == "binary" and head_probs.shape[1] >= 2:
head_probs = head_probs[:, :2]
curves = compute_per_class_curves(y_true, head_probs)
per_fold_curves.append((fold_idx, curves))
try:
if head_probs.shape[1] > 2:
fold_auc = roc_auc_score(y_true, head_probs, multi_class="ovr", average="macro")
else:
target_scores = head_probs[:, 1] if head_probs.shape[1] > 1 else head_probs[:, 0]
fold_auc = roc_auc_score(y_true, target_scores)
fold_aucs.append(fold_auc)
print(
f"[info] Fold {fold_idx}: using saved arrays "
f"({y_path.name}, {p_path.name}), AUC={fold_auc:.4f}"
)
except Exception:
print(
f"[warning] Fold {fold_idx}: using saved arrays "
f"({y_path.name}, {p_path.name}) but AUC failed."
)
continue
fold_models_dir = base_models_dir / f"fold{fold_idx}"
best_checkpoint = fold_models_dir / "model_best.pt"
if not best_checkpoint.exists():
print(
f"[warning] Fold {fold_idx}: missing model_best.pt at {best_checkpoint} "
f"and missing fallback arrays {y_path.name}/{p_path.name}"
)
continue
ht_args = build_ht_args(cli_args, fold_idx, run_dir, fold_models_dir, holdout_df)
ht = HyperTower(clinical, ht_args)
for handler in list(ht.logger.handlers):
handler.close()
ht.logger.handlers = [logging.NullHandler()]
train_log_path = Path("train.log")
if train_log_path.exists() and train_log_path.stat().st_size == 0:
try:
train_log_path.unlink()
except OSError:
pass
try:
state = torch.load(best_checkpoint, map_location=ht.device, weights_only=False)
except TypeError:
state = torch.load(best_checkpoint, map_location=ht.device)
ht._restore_from_state(state)
if args.use_holdout:
eval_df = holdout_df.copy()
else:
_, eval_df = clinical.get_split_dfs(fold_idx)
if cli_args["eval_mode"] == "binary":
eval_df = eval_df[eval_df[cli_args["label_col"]].isin([0, 1])].reset_index(drop=True)
if eval_df.empty:
print(f"[warning] Fold {fold_idx}: evaluation dataframe is empty; skipping.")
continue
ht.test_loader = ht._make_loader_for_df(eval_df, is_train=False)
y_true, probs_f, probs_i, probs_m = collect_logits(ht, ht.test_loader)
if cli_args["eval_mode"] == "binary":
y_true, probs_f, probs_i, probs_m = ensure_binary_slice(y_true, probs_f, probs_i, probs_m)
head_probs = choose_head_probs(head, probs_f, probs_i, probs_m)
if head_probs is None:
print(f"[skip] Fold {fold_idx}: head '{head}' not available.")
continue
if head_probs.shape[1] >= 2:
head_probs = head_probs[:, :2]
if args.overwrite:
base = run_dir / f"fold{fold_idx}{file_suffix}"
np.save(f"{base}_y_true.npy", y_true)
if probs_f is not None:
np.save(f"{base}_probs_fused.npy", probs_f)
if probs_i is not None:
np.save(f"{base}_probs_img.npy", probs_i)
if probs_m is not None:
np.save(f"{base}_probs_md.npy", probs_m)
curves = compute_per_class_curves(y_true, head_probs)
per_fold_curves.append((fold_idx, curves))
try:
if head_probs.shape[1] > 2:
fold_auc = roc_auc_score(y_true, head_probs, multi_class="ovr", average="macro")
else:
target_scores = head_probs[:, 1] if head_probs.shape[1] > 1 else head_probs[:, 0]
fold_auc = roc_auc_score(y_true, target_scores)
fold_aucs.append(fold_auc)
print(f"[info] Fold {fold_idx}: best epoch {best_epoch}, AUC={fold_auc:.4f}")
except Exception:
print(f"[warning] Fold {fold_idx}: unable to compute AUC.")
if not per_fold_curves:
raise SystemExit("No folds processed; nothing to plot.")
plots_dir = run_dir / "plots"
plot_overlays(per_fold_curves, plots_dir, class_names, head, cli_args["eval_mode"], file_suffix)
plot_mean_sd(per_fold_curves, plots_dir, class_names, head, cli_args["eval_mode"], file_suffix)
if fold_aucs:
print(f"[info] {head} head mean AUC across folds: {np.mean(fold_aucs):.4f} ± {np.std(fold_aucs):.4f}")
print(f"Plots regenerated under {plots_dir}")
if __name__ == "__main__":
main()