scripts/make_interpret.py

"""P11 interpretability: train TriFuse-AD on one representative fold, then run
Grad-CAM on the shared CNN encoder for a few correct + incorrect test cases.

Qualitative only. We do NOT claim the model localizes specific structures; we
report that attended regions overlap with anatomy known to be relevant in AD
(medial temporal lobe, ventricles, cortical atrophy).

Needs the GPU, so run this AFTER the main grid finishes (TriFuse-AD is the last
config in the grid and holds the GPU until then).
"""
from __future__ import annotations

import sys
from pathlib import Path

import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt  # noqa: E402

ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "src"))
RESULTS = ROOT / "results"
FIGURES = RESULTS / "figures"
SUBJECTS_CSV = ROOT / "data/metadata/subjects_clean.csv"

from trifuse.data.datasets import SliceDataset, collate, fit_tab_stats  # noqa: E402
from trifuse.data.splits import make_folds, split_for  # noqa: E402
from trifuse.models.registry import build_model  # noqa: E402
from trifuse.training.config import TrainConfig  # noqa: E402
from trifuse.training.trainer import fit  # noqa: E402
from trifuse.analysis.gradcam import GradCAM, overlay_heatmap  # noqa: E402
from trifuse.eval.metrics import CLASS_NAMES  # noqa: E402
from torch.utils.data import DataLoader  # noqa: E402

TRI = ("axial", "coronal", "sagittal")
SEED, FOLD = 7, 0


class _SlicesOnly(nn.Module):
    """Adapt TriFuse-AD's (slices, tab) forward to the single-input API Grad-CAM expects."""

    def __init__(self, model: nn.Module, tab: torch.Tensor):
        super().__init__()
        self.model = model
        self.tab = tab

    def forward(self, slices: torch.Tensor) -> torch.Tensor:
        return self.model(slices, self.tab)


def _target_layer(model: nn.Module) -> nn.Module:
    """Last conv stage of the timm ConvNeXt encoder (outputs B,C,H,W)."""
    enc = model.encoder
    if hasattr(enc, "stages"):
        return enc.stages[-1]
    # fallback: last Conv2d
    convs = [m for m in enc.modules() if isinstance(m, nn.Conv2d)]
    return convs[-1]


def _train_fold(device: str) -> tuple[nn.Module, pd.DataFrame, dict, TrainConfig]:
    df = pd.read_csv(SUBJECTS_CSV)
    folds = make_folds(df)
    tr, va, te = split_for(folds, df, SEED, FOLD)
    train_df = df[df.subject_id.isin(tr)]
    val_df = df[df.subject_id.isin(va)]
    test_df = df[df.subject_id.isin(te)]

    cfg = TrainConfig(name="trifuse_interp", model="trifuse", modality="multimodal",
                      n_slices=9, planes=TRI, epochs=70, batch_size=6,
                      lr=3e-4, backbone_lr=3e-5, freeze_epochs=6, seed=SEED)
    tab_stats = fit_tab_stats(train_df, cfg.tab_features)
    counts = np.bincount(train_df["class_id"], minlength=3).tolist()

    def loader(sub, shuffle, aug):
        ds = SliceDataset(sub, cfg.tab_features, tab_stats, planes=TRI,
                          n_slices=cfg.n_slices, modality="multimodal", augment=aug)
        return DataLoader(ds, batch_size=cfg.batch_size, shuffle=shuffle,
                          num_workers=cfg.num_workers, collate_fn=collate, pin_memory=True)

    model = build_model(cfg, n_tab_features=len(cfg.tab_features), pretrained=True)
    fit(model, loader(train_df, True, True), loader(val_df, False, False),
        cfg, counts, device=device)
    return model, test_df, tab_stats, cfg


def _gradcam_for_subject(model, target, batch, device, n_slices, planes):
    """Return {plane: (gray_center, heat_center)} for the argmax class, center slice per plane.

    TriFuse-AD flattens the 27 slices into the CNN batch dim, so the target layer's
    captured activations/grads are (T, C, H', W') — one per slice. We compute a
    per-token Grad-CAM here rather than using GradCAM.__call__'s single-map return.
    """
    import torch.nn.functional as F
    slices = batch["slices"].to(device)          # (1, T, 3, H, W)
    tab = batch["tab"].to(device)
    wrapper = _SlicesOnly(model, tab)
    cam = GradCAM(wrapper, target)
    _ = cam(slices, class_idx=None)               # triggers fwd+bwd, fills cam._acts/_grads
    acts, grads = cam._acts, cam._grads           # (T, C, H', W')
    cam.remove()

    H, W = slices.shape[-2:]
    weights = grads.mean(dim=(2, 3), keepdim=True)         # (T, C, 1, 1)
    heat = F.relu((weights * acts).sum(dim=1))             # (T, H', W')
    heat = F.interpolate(heat.unsqueeze(1), size=(H, W), mode="bilinear",
                         align_corners=False)[:, 0]        # (T, H, W)
    heat = heat.detach().cpu().numpy()

    out = {}
    center = n_slices // 2
    for p_idx, plane in enumerate(planes):
        tok = p_idx * n_slices + center
        gray = slices[0, tok, 0].detach().cpu().numpy()
        h = heat[tok]
        h = (h - h.min()) / (np.ptp(h) + 1e-8)             # per-slice normalize
        out[plane] = (gray, h)
    return out


def main():
    device = "cuda" if torch.cuda.is_available() else "cpu"
    FIGURES.mkdir(parents=True, exist_ok=True)
    print(f"training TriFuse-AD on seed={SEED} fold={FOLD} ({device}) for interpretability...",
          flush=True)
    model, test_df, tab_stats, cfg = _train_fold(device)
    model.eval()
    target = _target_layer(model)

    ds = SliceDataset(test_df, cfg.tab_features, tab_stats, planes=TRI,
                      n_slices=cfg.n_slices, modality="multimodal", augment=False)
    dl = DataLoader(ds, batch_size=1, shuffle=False, collate_fn=collate)

    cases = []  # (subject_id, y_true, y_pred, {plane:(gray,heat)})
    for batch in dl:
        with torch.no_grad():
            logits = model(batch["slices"].to(device), batch["tab"].to(device))
            pred = int(logits.argmax(1)[0])
        yt = int(batch["y"][0])
        maps = _gradcam_for_subject(model, target, batch, device, cfg.n_slices, TRI)
        cases.append((batch["subject_id"][0], yt, pred, maps))

    correct = [c for c in cases if c[1] == c[2]][:6]
    wrong = [c for c in cases if c[1] != c[2]][:3]
    picked = correct + wrong
    if not picked:
        print("no test cases produced; aborting figure")
        return

    n = len(picked)
    fig, axes = plt.subplots(n, 3, figsize=(9, 3 * n))
    if n == 1:
        axes = axes[None, :]
    for r, (sid, yt, yp, maps) in enumerate(picked):
        tag = "OK" if yt == yp else "WRONG"
        for c, plane in enumerate(TRI):
            gray, heat = maps[plane]
            axes[r, c].imshow(overlay_heatmap(gray, heat))
            axes[r, c].axis("off")
            if c == 0:
                axes[r, c].set_ylabel(f"{sid}\nT:{CLASS_NAMES[yt]} P:{CLASS_NAMES[yp]} [{tag}]",
                                      fontsize=8, rotation=0, ha="right", va="center")
            axes[r, c].set_title(plane if r == 0 else "", fontsize=9)
    fig.suptitle("TriFuse-AD Grad-CAM (shared CNN encoder, center slice per plane)")
    fig.tight_layout()
    out = FIGURES / "fig_gradcam_trifuse.png"
    fig.savefig(out, dpi=150)
    plt.close(fig)
    print(f"wrote {out}  ({len(correct)} correct + {len(wrong)} wrong cases)")


if __name__ == "__main__":
    main()