scripts/analysis/ablation_swap_audio.py

"""Ablation-Swap: swap audio between different real identities and measure s_asym.

Motivation
----------
Our OOD-asymmetry claim says that s_asym < 0 on talking-head fakes because the
video is off-manifold while the audio stays in-distribution. To pressure-test
that account, this script constructs a "swap" ablation:

  * both video and audio come from REAL clips
  * but the audio is shuffled to belong to a DIFFERENT identity than the video

This puts the joint (v, a) pair off-manifold at the pairing level while each
marginal remains in-distribution. The predictor sees an ambiguous case that
is neither the "video-OOD" fake regime nor the fully in-distribution real
regime.

Prediction (falsifiable)
------------------------
If s_asym is genuinely tracking "which side is OOD", the swap distribution
should sit BETWEEN the real distribution (≈ 0) and the fake distribution
(≈ -0.23) — not indistinguishable from real.

Output
------
A .npz file with per-sample l_av, l_va, s_asym, score, label, meta, plus a
`condition` field ∈ {"real_paired", "real_swapped", "fake"} for downstream
plotting. Also emits a comparison histogram.

Run
---
Same hydra flags as dump_cta_features.py; example:

    python3 scripts/analysis/ablation_swap_audio.py \
        +ckpt=outputs/cta_ablation_diffusion_A1_full_20260611_161857/checkpoints/epoch02-valauc0.9999.ckpt \
        method=cta_ablation method.ablation_variant=A1_full \
        data=fairtalking +split=val \
        +out_dir=outputs/analysis/ablation_swap_diffusion \
        +seed=0
"""
from __future__ import annotations

import os
import sys
import warnings
from pathlib import Path
from typing import Any, Dict, List

import hydra
import numpy as np
import torch
import torch.nn.functional as F
from omegaconf import DictConfig, OmegaConf
from torch.utils.data import DataLoader

# --- silence torch.load weights_only restriction (mirror src/train.py) -----
import lightning_fabric.utilities.cloud_io as _lf_cloud_io
_orig_torch_load = torch.load
def _unsafe_torch_load(*args, **kwargs):
    kwargs["weights_only"] = False
    return _orig_torch_load(*args, **kwargs)
_lf_cloud_io.torch.load = _unsafe_torch_load
torch.load = _unsafe_torch_load

# ensure src/ is importable
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))

from src.data import FairTalkingDataModule  # noqa: E402
from src.methods import build_method        # noqa: E402


def _forward_batch(model, video, audio, device):
    v = model.model.video(video)
    a = model.model.audio(audio)
    v_pred = model.model.av_pred(src_tokens=a["tokens"], tgt_query=v["tokens"])
    a_pred = model.model.va_pred(src_tokens=v["tokens"], tgt_query=a["tokens"])
    l_av = F.mse_loss(v_pred, v["tokens"], reduction="none").mean(dim=[1, 2])
    l_va = F.mse_loss(a_pred, a["tokens"], reduction="none").mean(dim=[1, 2])
    asym = l_va - l_av
    logits = model.model.classify(v["pooled"], a["pooled"], l_av, l_va)
    score = torch.sigmoid(logits.squeeze(-1))
    return (l_av.cpu().float().numpy(),
            l_va.cpu().float().numpy(),
            asym.cpu().float().numpy(),
            score.cpu().float().numpy())


def _permute_audio_between_reals(audio: torch.Tensor, is_real: torch.Tensor,
                                  rng: np.random.Generator) -> torch.Tensor:
    """Swap audio ONLY between real samples in the batch.

    Fake samples keep their audio (we don't want to mutate the fake condition).
    Real samples get their audio permuted with a derangement so that no real
    sample retains its own audio.
    """
    real_idx = torch.nonzero(is_real, as_tuple=False).squeeze(-1)
    if real_idx.numel() < 2:
        return audio  # can't derange
    order = np.arange(real_idx.numel())
    for _ in range(20):  # try up to 20 times to find a derangement
        rng.shuffle(order)
        if not np.any(order == np.arange(real_idx.numel())):
            break
    else:
        # fall back to a rotation by 1
        order = np.roll(np.arange(real_idx.numel()), 1)

    swapped = audio.clone()
    swapped[real_idx] = audio[real_idx[order]]
    return swapped


@hydra.main(version_base=None, config_path="../../configs", config_name="train")
def main(cfg: DictConfig) -> None:
    ckpt_path = cfg.get("ckpt", None)
    if ckpt_path is None:
        raise SystemExit("Missing +ckpt=<path> override.")
    ckpt_path = str(Path(ckpt_path).resolve())
    out_dir = Path(cfg.get("out_dir", "outputs/analysis/ablation_swap")).resolve()
    out_dir.mkdir(parents=True, exist_ok=True)
    split = cfg.get("split", "val")
    seed = int(cfg.get("seed", 0))
    max_batches = cfg.get("max_batches", None)
    max_batches = None if max_batches in (None, "null", "None") else int(max_batches)

    print(f"[swap] ckpt      = {ckpt_path}")
    print(f"[swap] data cfg  = {cfg.data.name}")
    print(f"[swap] split     = {split}")
    print(f"[swap] out_dir   = {out_dir}")
    print(f"[swap] seed      = {seed}")

    model = build_method(
        method_name=cfg.method.name,
        method_cfg=cfg.method,
        backbone_cfg=cfg.backbone,
        data_cfg=cfg.data,
    )
    state = torch.load(ckpt_path, map_location="cpu")
    sd = state.get("state_dict", state)
    missing, unexpected = model.load_state_dict(sd, strict=False)
    if missing:    print(f"[swap]  {len(missing)} missing keys (first 5): {missing[:5]}")
    if unexpected: print(f"[swap]  {len(unexpected)} unexpected keys (first 5): {unexpected[:5]}")

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device).eval()

    dm = FairTalkingDataModule(data_cfg=cfg.data, return_paired=False)
    stage = "fit" if split in {"train", "val"} else "test"
    dm.setup(stage=stage)
    if split == "train":
        loader = dm.train_dataloader()
    elif split == "val":
        loader = dm.val_dataloader()
    else:
        loader = dm.test_dataloader()

    rng = np.random.default_rng(seed)
    n_batches = len(loader) if max_batches is None else min(max_batches, len(loader))
    print(f"[swap] forwarding {n_batches} batches (paired + swapped) …")

    rows: List[Dict[str, Any]] = []
    with torch.no_grad():
        for bi, batch in enumerate(loader):
            if batch is None:
                continue
            if max_batches is not None and bi >= max_batches:
                break
            video = batch["video"].to(device, non_blocking=True)
            audio = batch["audio"].to(device, non_blocking=True)
            labels = batch["label"].long()
            metas = batch.get("meta", [{}] * video.size(0))
            is_real = (labels == 0)

            # ---- pass 1: original pairing (real_paired + fake) --------------
            l_av_o, l_va_o, asym_o, score_o = _forward_batch(model, video, audio, device)

            # ---- pass 2: audio swapped among reals -------------------------
            audio_sw = _permute_audio_between_reals(audio, is_real, rng)
            l_av_s, l_va_s, asym_s, score_s = _forward_batch(model, video, audio_sw, device)

            for i in range(video.size(0)):
                m = metas[i] if isinstance(metas[i], dict) else {}
                base = {
                    "basename":  str(m.get("basename",  "")),
                    "generator": str(m.get("generator", "")),
                    "num":       str(m.get("num",       "")),
                    "label":     int(labels[i].item()),
                }
                # original condition
                cond_orig = "real_paired" if is_real[i] else "fake"
                rows.append({
                    **base,
                    "condition": cond_orig,
                    "l_av":  float(l_av_o[i]),
                    "l_va":  float(l_va_o[i]),
                    "asym":  float(asym_o[i]),
                    "score": float(score_o[i]),
                })
                # swapped condition ONLY for real samples (fake pass-2 is redundant)
                if is_real[i]:
                    rows.append({
                        **base,
                        "condition": "real_swapped",
                        "l_av":  float(l_av_s[i]),
                        "l_va":  float(l_va_s[i]),
                        "asym":  float(asym_s[i]),
                        "score": float(score_s[i]),
                    })

            if (bi + 1) % 20 == 0:
                print(f"[swap]   batch {bi+1}/{n_batches}   collected {len(rows)} rows")

    # ---- summary + save ----------------------------------------------------
    import pandas as pd
    df = pd.DataFrame(rows)
    csv_path = out_dir / "swap_features.csv"
    df.to_csv(csv_path, index=False)
    print(f"[swap] wrote {csv_path}   ({len(df)} rows)")

    print("\n[swap] Per-condition summary")
    for cond in ("real_paired", "real_swapped", "fake"):
        sub = df[df["condition"] == cond]
        if len(sub) == 0:
            continue
        print(f"  {cond:<14}  n={len(sub):4d}   "
              f"l_av={sub['l_av'].mean():.4f}   "
              f"l_va={sub['l_va'].mean():.4f}   "
              f"asym={sub['asym'].mean():+.4f}   "
              f"score={sub['score'].mean():.3f}")

    # ---- histogram ---------------------------------------------------------
    import matplotlib
    matplotlib.use("Agg")
    import matplotlib.pyplot as plt

    palette = {
        "real_paired":  ("#1E8449", "real (paired)"),
        "real_swapped": ("#F1C40F", "real (audio-swapped)"),
        "fake":         ("#C0392B", "fake"),
    }
    fig, ax = plt.subplots(figsize=(7.6, 4.6))
    all_asym = df["asym"].values
    lo, hi = np.quantile(all_asym, 0.005), np.quantile(all_asym, 0.995)
    bins = np.linspace(lo, hi, 55)
    for cond, (color, label) in palette.items():
        vals = df.loc[df["condition"] == cond, "asym"].values
        if len(vals) == 0:
            continue
        ax.hist(vals, bins=bins, density=True, alpha=0.55,
                color=color, label=f"{label}  n={len(vals)}   μ={vals.mean():+.3f}",
                edgecolor="none")
        ax.axvline(vals.mean(), color=color, lw=1.2, linestyle="--", alpha=0.9)
    ax.set_xlabel(r"$s_{\rm asym}$ = $L_{V \to A}$ − $L_{A \to V}$")
    ax.set_ylabel("density")
    ax.set_title("Ablation-Swap:  swapping audio between real identities")
    ax.legend(loc="best", fontsize=9)
    ax.grid(alpha=0.25, linestyle=":")
    plt.tight_layout()
    for ext in ("png", "pdf"):
        p = out_dir / f"hist_asym_swap.{ext}"
        plt.savefig(p, dpi=240, bbox_inches="tight", facecolor="white")
        print(f"[swap] wrote {p}")
    plt.close(fig)


if __name__ == "__main__":
    main()