scripts/analysis/evaluate_robustness.py

"""Robustness evaluation for CTA: in-memory perturb video frames at
inference time, recompute AUROC / AP / Acc / Acc@EER / fairness.

Mirrors X-AVDT/train/evaluate_robustness.py (DeeperForensics-style 6
frame-level perturbations × 5 levels), adapted to CTA's data layer:

    Input difference  : CTA reads mp4 directly via decord/pyav (not pre-extracted .pt)
    Perturbation point: between load_video_clip()  and  video_transform()
                        (i.e., on a (T, 3, H, W) float tensor in [0, 1])
    Audio             : NEVER perturbed (consistent with CTA's "audio is real" framing)

Perturbation set & level table copied verbatim from
/apdcephfs_gy4/share_303628665/joywu/research/X-AVDT/train/evaluate_robustness.py
to keep the comparison apples-to-apples.

Output:
  * console summary per (perturbation, level)
  * append-only JSON to --save_json (one run per CLI invocation)

Usage:
    /opt/conda/envs/pytorch/bin/python scripts/analysis/evaluate_robustness.py \\
        --ckpt outputs/cta_diffusion_combined_20260604_205145/checkpoints/epoch16-valauc1.0000.ckpt \\
        --data fairtalking_diffusion_only \\
        --perturbation gaussian_noise --level 3 \\
        --save_json outputs/analysis/robustness/cta_runs.json

  Or run the whole sweep (6 × 5 = 30 settings) by wrapping in a loop:
    for p in color_saturation color_contrast block_wise gaussian_noise gaussian_blur pixelate; do
      for L in 1 2 3 4 5; do
        python scripts/analysis/evaluate_robustness.py --ckpt ... --data ... \\
          --perturbation $p --level $L --save_json ...
      done
    done
"""
from __future__ import annotations

import argparse
import collections
import csv as _csv
import json
import math
import os
import random as _rng_mod
import sys
import time
from pathlib import Path
from typing import Dict, List, Optional

import cv2
import numpy as np
import torch
import torch.nn.functional as F
from omegaconf import OmegaConf
from sklearn.metrics import (
    accuracy_score, average_precision_score, classification_report,
    confusion_matrix, roc_auc_score, roc_curve,
)
from torch.utils.data import DataLoader
from tqdm import tqdm

# --- silence 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

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


# ============================================================================
# Perturbation table (copy-paste from X-AVDT/train/evaluate_robustness.py)
# ============================================================================
SEVERITY = {
    "color_saturation":  [1.0,   0.8,   1.2,   1.5,   2.0],
    "color_contrast":    [1.0,   0.85,  1.2,   1.4,   1.6],
    "block_wise":        [0,     8,     16,    24,    32],
    "gaussian_noise":    [0.0,   0.001, 0.005, 0.01,  0.05],
    "gaussian_blur":     [1,     3,     7,     11,    15],
    "pixelate":          [1,     2,     4,     6,     8],
    "jpeg_quality":      [100,   85,    70,    50,    30],   # JPEG quality factor (lower = stronger compression)
}
PERTURBATIONS = list(SEVERITY.keys())
DEMO_DIMS = ("gender", "race4", "age_group")


# ----- frame-level primitives (operate on uint8 BGR, copied verbatim) -----
def _bgr2ycbcr(img_bgr):
    img = img_bgr.astype(np.float32) / 255.0
    M = np.array([
        [ 0.299,    0.587,    0.114],
        [-0.16874, -0.33126,  0.5],
        [ 0.5,     -0.41869, -0.08131],
    ], dtype=np.float32)
    yuv = img @ M.T
    yuv[..., 1:] += 0.5
    return yuv


def _ycbcr2bgr(ycbcr):
    yuv = ycbcr.copy()
    yuv[..., 1:] -= 0.5
    M = np.array([
        [1.0,  0.0,      1.402],
        [1.0, -0.34414, -0.71414],
        [1.0,  1.772,    0.0],
    ], dtype=np.float32)
    bgr = yuv @ M.T
    return np.clip(bgr * 255.0, 0, 255)


def _apply_color_saturation(frame_bgr, param):
    if abs(param - 1.0) < 1e-6:
        return frame_bgr
    ycbcr = _bgr2ycbcr(frame_bgr)
    ycbcr[:, :, 1] = 0.5 + (ycbcr[:, :, 1] - 0.5) * param
    ycbcr[:, :, 2] = 0.5 + (ycbcr[:, :, 2] - 0.5) * param
    out = _ycbcr2bgr(ycbcr)
    return np.clip(out, 0, 255).astype(np.uint8)


def _apply_color_contrast(frame_bgr, param):
    if abs(param - 1.0) < 1e-6:
        return frame_bgr
    out = frame_bgr.astype(np.float32) * param
    return np.clip(out, 0, 255).astype(np.uint8)


def _apply_block_wise(frame_bgr, param, rng):
    if param <= 0:
        return frame_bgr
    width = 8
    block = np.ones((width, width, 3), dtype=np.uint8) * 128
    n = max(1, min(frame_bgr.shape[0], frame_bgr.shape[1]) // 256 * int(param))
    out = frame_bgr.copy()
    H, W = frame_bgr.shape[:2]
    for _ in range(n):
        rw = rng.randint(0, W - 1 - width)
        rh = rng.randint(0, H - 1 - width)
        out[rh:rh + width, rw:rw + width, :] = block
    return out


def _apply_gaussian_noise(frame_bgr, param, rng_np):
    if param <= 0:
        return frame_bgr
    ycbcr = _bgr2ycbcr(frame_bgr)
    h, w, c = ycbcr.shape
    noise = math.sqrt(param) * rng_np.standard_normal((h, w, c)).astype(np.float32)
    noisy = ycbcr + noise
    out = _ycbcr2bgr(noisy)
    return np.clip(out, 0, 255).astype(np.uint8)


def _apply_gaussian_blur(frame_bgr, ksize):
    ksize = int(ksize)
    if ksize <= 1:
        return frame_bgr
    if ksize % 2 == 0:
        ksize += 1
    sigma = ksize / 6.0
    return cv2.GaussianBlur(frame_bgr, (ksize, ksize), sigma)


def _apply_pixelate(frame_bgr, factor):
    factor = int(factor)
    if factor <= 1:
        return frame_bgr
    h, w = frame_bgr.shape[:2]
    sw, sh = max(1, w // factor), max(1, h // factor)
    small = cv2.resize(frame_bgr, (sw, sh), interpolation=cv2.INTER_AREA)
    return cv2.resize(small, (w, h), interpolation=cv2.INTER_LINEAR)


def _apply_jpeg_quality(frame_bgr, quality):
    """JPEG-encode then decode the BGR frame to simulate compression.

    quality is the JPEG quality factor in [1, 100]; higher = better quality
    (less compression). 100 is effectively a no-op.
    """
    quality = int(quality)
    if quality >= 100:
        return frame_bgr
    quality = max(1, min(100, quality))
    encode_params = [int(cv2.IMWRITE_JPEG_QUALITY), quality]
    ok, buf = cv2.imencode(".jpg", frame_bgr, encode_params)
    if not ok:
        return frame_bgr
    decoded = cv2.imdecode(buf, cv2.IMREAD_COLOR)
    return decoded if decoded is not None else frame_bgr


def perturb_video_tensor(video01: torch.Tensor, perturbation: str, level: int,
                         base_seed: int = 42) -> torch.Tensor:
    """video01: (T, 3, H, W) float tensor in [0, 1]. Returns same shape/dtype.

    Internally:
      1) (T,3,H,W) float [0,1]  ->  (T,H,W,3) uint8 RGB
      2) Per frame: RGB->BGR, apply perturbation, BGR->RGB
      3) (T,H,W,3) uint8 RGB    ->  (T,3,H,W) float [0,1]
    """
    if level == 1:
        return video01     # short-circuit: SEVERITY[*][0] is a no-op
    param = SEVERITY[perturbation][level - 1]

    seed = (hash((base_seed, perturbation, level)) & 0xFFFFFFFF)
    py_rng = _rng_mod.Random(seed)
    np_rng = np.random.RandomState(seed)

    # tensor -> numpy uint8 RGB (T,H,W,3)
    arr = video01.detach().cpu().numpy()           # (T,3,H,W) float
    arr = (arr * 255.0).clip(0, 255).astype(np.uint8)
    arr = np.transpose(arr, (0, 2, 3, 1))           # (T,H,W,3) RGB

    out = np.empty_like(arr)
    for t in range(arr.shape[0]):
        bgr = cv2.cvtColor(arr[t], cv2.COLOR_RGB2BGR)
        if perturbation == "color_saturation":
            bgr = _apply_color_saturation(bgr, param)
        elif perturbation == "color_contrast":
            bgr = _apply_color_contrast(bgr, param)
        elif perturbation == "block_wise":
            bgr = _apply_block_wise(bgr, param, py_rng)
        elif perturbation == "gaussian_noise":
            bgr = _apply_gaussian_noise(bgr, param, np_rng)
        elif perturbation == "gaussian_blur":
            bgr = _apply_gaussian_blur(bgr, param)
        elif perturbation == "pixelate":
            bgr = _apply_pixelate(bgr, param)
        elif perturbation == "jpeg_quality":
            bgr = _apply_jpeg_quality(bgr, param)
        else:
            raise ValueError(f"unsupported perturbation: {perturbation}")
        out[t] = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)

    # numpy uint8 RGB -> tensor [0,1]
    out = np.transpose(out, (0, 3, 1, 2))           # (T,3,H,W)
    return torch.from_numpy(out.astype(np.float32) / 255.0).to(video01.device)


# ============================================================================
# Wrapper transform: insert perturbation BEFORE the existing transform
# ============================================================================
class PerturbedVideoTransform:
    """Wrap an existing VideoTransform. Apply perturbation in [0,1] domain
    first, then delegate to the underlying transform (which crops/normalizes)."""

    def __init__(self, base_transform, perturbation: str, level: int,
                 base_seed: int = 42):
        self.base = base_transform
        self.perturbation = perturbation
        self.level = level
        self.base_seed = base_seed

    def __call__(self, video):
        if self.level != 1:
            video = perturb_video_tensor(
                video, self.perturbation, self.level, self.base_seed,
            )
        if self.base is not None:
            video = self.base(video)
        return video


# ============================================================================
# Metric helpers (copy from X-AVDT)
# ============================================================================
def _tpr_at_fpr(y_true, y_score, fpr_target):
    fpr, tpr, _ = roc_curve(y_true, y_score)
    if (fpr <= fpr_target).any():
        return float(tpr[fpr <= fpr_target].max())
    return 0.0


def _compute_eer_threshold(y_true, y_score):
    fpr, tpr, thresholds = roc_curve(y_true, y_score)
    fnr = 1 - tpr
    return float(thresholds[int(np.argmin(np.abs(fpr - fnr)))])


def _metrics_block(y_true, y_score, threshold=0.5):
    y_true = np.asarray(y_true)
    y_score = np.asarray(y_score)
    y_pred = (y_score >= threshold).astype(int)

    out = {}
    try:    out["AUROC"] = float(roc_auc_score(y_true, y_score))
    except: out["AUROC"] = None
    try:    out["AP"] = float(average_precision_score(y_true, y_score))
    except: out["AP"] = None
    out[f"Accuracy@{threshold:.2f}"] = float(accuracy_score(y_true, y_pred))
    out["Confusion Matrix"] = confusion_matrix(y_true, y_pred, labels=[0, 1]).tolist()
    out["Classification Report"] = classification_report(
        y_true, y_pred, labels=[0, 1], output_dict=True, zero_division=0
    )
    try:
        thr = _compute_eer_threshold(y_true, y_score)
        out["EER_threshold"] = thr
        out["Acc@EER"] = float(accuracy_score(y_true, (y_score >= thr).astype(int)))
    except:
        out["EER_threshold"] = None; out["Acc@EER"] = None
    try:
        out["TPR@FPR=1%"] = _tpr_at_fpr(y_true, y_score, 0.01)
        out["TPR@FPR=0.1%"] = _tpr_at_fpr(y_true, y_score, 0.001)
    except:
        out["TPR@FPR=1%"] = None; out["TPR@FPR=0.1%"] = None
    return out


def _fmt4(v):
    try:    return f"{float(v):.4f}"
    except: return "n/a"


def compute_fairness(y_true, y_score, groups):
    y_true = np.asarray(y_true); y_score = np.asarray(y_score); groups = np.asarray(groups)
    pred = (y_score > 0.5).astype(int)
    def _g_fpr(g):
        m = (groups == g) & (y_true == 0); return float(pred[m].mean()) if m.sum() else 0.0
    def _g_tpr(g):
        m = (groups == g) & (y_true == 1); return float(pred[m].mean()) if m.sum() else 0.0
    def _g_acc(g):
        m = (groups == g);                  return float((pred[m] == y_true[m]).mean()) if m.sum() else 0.0
    def _g_dp(g):
        m = (groups == g);                  return float(pred[m].mean()) if m.sum() else 0.0
    uniq = sorted(set(groups.tolist()))
    if not uniq:
        return None
    fprs = [_g_fpr(g) for g in uniq]
    tprs = [_g_tpr(g) for g in uniq]
    accs = [_g_acc(g) for g in uniq]
    dps  = [_g_dp(g)  for g in uniq]
    ns   = [int((groups == g).sum()) for g in uniq]
    return {
        "F_FPR": float(np.std(fprs)) * 100,
        "F_MEO": (max(max(fprs) - min(fprs), max(tprs) - min(tprs))) * 100,
        "F_DP":  float(np.std(dps))  * 100,
        "F_OAE": float(np.std(accs)) * 100,
        "groups": {g: {"n": n, "fpr": f, "tpr": t, "acc": a, "dp": d}
                   for g, n, f, t, a, d in zip(uniq, ns, fprs, tprs, accs, dps)},
    }


def load_demographics(csv_path: Optional[str]) -> Dict[str, Dict[str, str]]:
    if not csv_path or not os.path.exists(csv_path):
        return {}
    out = {}
    with open(csv_path, newline="") as f:
        for row in _csv.DictReader(f):
            base = (row.get("basename") or "").strip()
            if base:
                out[base] = {
                    "gender":    (row.get("gender") or "").strip(),
                    "race4":     (row.get("race4") or "").strip(),
                    "age_group": (row.get("age_group") or "").strip(),
                }
    return out


def _set_seed(seed):
    np.random.seed(seed); torch.manual_seed(seed); _rng_mod.seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False


# ============================================================================
# main
# ============================================================================
def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument("--ckpt", type=str, required=True)
    p.add_argument("--data", type=str, default="fairtalking_diffusion_only",
                   help="hydra data config name (without .yaml)")
    p.add_argument("--perturbation", type=str, required=True, choices=PERTURBATIONS)
    p.add_argument("--level", type=int, required=True, choices=[1, 2, 3, 4, 5])
    p.add_argument("--batch_size", type=int, default=8)
    p.add_argument("--num_workers", type=int, default=4)
    p.add_argument("--seed", type=int, default=42)
    p.add_argument("--demographics_csv", type=str, default=None,
                   help="optional CSV with basename,race4,gender,age_group "
                        "for fairness; if omitted, fairness section is skipped")
    p.add_argument("--save_json", type=str, default=None)
    p.add_argument("--verbose", action="store_true",
                   help="Also print per-fake-vs-real and fairness blocks to "
                        "stdout. JSON always contains them regardless.")
    return p.parse_args()


def main():
    args = parse_args()
    _set_seed(args.seed)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # ---- model ckpt -> infer the method/data/backbone configs ---------------
    print(f"[robustness] ckpt: {args.ckpt}")
    print(f"[robustness] data config: {args.data}")
    print(f"[robustness] perturbation={args.perturbation}  level={args.level}  "
          f"param={SEVERITY[args.perturbation][args.level - 1]}")

    state = torch.load(args.ckpt, map_location="cpu")
    hp = state.get("hyper_parameters", {})
    if not hp:
        raise SystemExit("ckpt has no hyper_parameters; cannot rebuild model")
    method_cfg = OmegaConf.create(hp["method_cfg"])
    backbone_cfg = OmegaConf.create(hp["backbone_cfg"])

    # ---- build target data config (from yaml) ------------------------------
    data_cfg_path = Path("configs/data") / f"{args.data}.yaml"
    if not data_cfg_path.exists():
        raise SystemExit(f"data config not found: {data_cfg_path}")
    data_cfg = OmegaConf.load(data_cfg_path)
    # resolve env vars in cfg (DATA_ROOT, MMDF_ROOT, ...)
    os.environ.setdefault("DATA_ROOT",
        "/apdcephfs_gy4/share_303628665/joywu/dataset/FairTalking-Bench")
    os.environ.setdefault("HDTF_PAIRED_ROOT",
        "/apdcephfs_gy5/share_303628665/joyewu/HDTF-paird")
    os.environ.setdefault("MMDF_ROOT",
        "/apdcephfs_gy4/share_303628665/joywu/dataset/MMDF_test_only")
    data_cfg = OmegaConf.create(OmegaConf.to_container(data_cfg, resolve=True))

    # ---- build model + load ckpt -------------------------------------------
    model = build_method(
        method_name=method_cfg.name,
        method_cfg=method_cfg,
        backbone_cfg=backbone_cfg,
        data_cfg=data_cfg,
    )
    sd = state.get("state_dict", state)
    missing, unexpected = model.load_state_dict(sd, strict=False)
    if missing:    print(f"  missing keys: {len(missing)} (first 3: {missing[:3]})")
    if unexpected: print(f"  unexpected keys: {len(unexpected)}")
    model.to(device).eval()

    # ---- build datamodule, then patch eval_transform with perturbation ----
    dm = FairTalkingDataModule(data_cfg=data_cfg, return_paired=False)
    base_eval_transform = dm.eval_transform
    dm.eval_transform = PerturbedVideoTransform(
        base_eval_transform, args.perturbation, args.level, args.seed,
    )
    dm.setup(stage="test")
    loader = DataLoader(
        dm.test_ds, batch_size=args.batch_size, shuffle=False,
        num_workers=args.num_workers, pin_memory=torch.cuda.is_available(),
        collate_fn=getattr(dm, "_collate_fn",
                           __import__("src.data.datamodule", fromlist=["_collate_drop_none"])._collate_drop_none),
    )

    # ---- demographics (optional) -------------------------------------------
    demographics = load_demographics(args.demographics_csv)

    # ---- forward -----------------------------------------------------------
    scores, labels, basenames, generators = [], [], [], []
    with torch.inference_mode():
        for batch in tqdm(loader, desc=f"{args.perturbation}/L{args.level}", leave=False):
            if batch is None:
                continue
            batch_video = batch["video"].to(device, dtype=torch.float)
            batch_audio = batch["audio"].to(device, dtype=torch.float)
            score = model.score({"video": batch_video, "audio": batch_audio})
            scores.extend(score.detach().cpu().float().numpy().tolist())
            labels.extend(batch["label"].long().tolist())
            metas = batch.get("meta", [])
            for m in metas:
                basenames.append(str((m or {}).get("basename", "")))
                generators.append(str((m or {}).get("generator", "")))

    # ---- metrics: overall + per-fake-vs-real -------------------------------
    result = {
        "perturbation": args.perturbation,
        "level": args.level,
        "param": SEVERITY[args.perturbation][args.level - 1],
        "n_samples": len(scores),
    }
    o = _metrics_block(labels, scores)
    o["Accuracy"] = o["Accuracy@0.50"]
    result["overall"] = o

    # per-fake-vs-real (clip-level), only if generator info exists
    real_idx  = [i for i, y in enumerate(labels) if y == 0]
    real_sc   = [scores[i] for i in real_idx]
    real_lab  = [labels[i] for i in real_idx]
    fake_gens = sorted({generators[i] for i, y in enumerate(labels) if y == 1 and generators[i]})
    per_fake = {}
    for fm in fake_gens:
        idxs = [i for i, (g, y) in enumerate(zip(generators, labels)) if g == fm and y == 1]
        joint_sc  = [scores[i] for i in idxs] + real_sc
        joint_lab = [labels[i] for i in idxs] + real_lab
        block = _metrics_block(joint_lab, joint_sc)
        block["Accuracy"] = block["Accuracy@0.50"]
        per_fake[fm] = block
    result["per_fake_vs_real"] = per_fake

    # fairness (overall, by demographic dim) ----------------------------------
    fairness_overall = {}
    if demographics:
        for d in DEMO_DIMS:
            groups = [demographics.get(b, {}).get(d, "") for b in basenames]
            valid = [i for i, g in enumerate(groups) if g]
            if not valid:
                continue
            fb = compute_fairness(
                [labels[i]  for i in valid],
                [scores[i]  for i in valid],
                [groups[i]  for i in valid],
            )
            fairness_overall[d] = fb
    result["fairness_overall"] = fairness_overall

    # ---- console summary ---------------------------------------------------
    print(f"\n[{args.perturbation} L{args.level}]  AUROC={_fmt4(o['AUROC'])} "
          f"AP={_fmt4(o['AP'])} Acc={_fmt4(o['Accuracy'])} "
          f"Acc@EER={_fmt4(o['Acc@EER'])} TPR@1%FPR={_fmt4(o['TPR@FPR=1%'])} "
          f"TPR@0.1%FPR={_fmt4(o['TPR@FPR=0.1%'])}  n={result['n_samples']}")
    if args.verbose:
        for fm, blk in per_fake.items():
            print(f"  [{fm}+Real] AUROC={_fmt4(blk['AUROC'])} AP={_fmt4(blk['AP'])} "
                  f"Acc={_fmt4(blk['Accuracy'])} Acc@EER={_fmt4(blk['Acc@EER'])}")
        for d in DEMO_DIMS:
            fb = fairness_overall.get(d)
            if fb:
                print(f"  fairness[{d}] F_FPR={fb['F_FPR']:.2f} F_MEO={fb['F_MEO']:.2f} "
                      f"F_DP={fb['F_DP']:.2f} F_OAE={fb['F_OAE']:.2f}")

    # ---- save / append json ------------------------------------------------
    if args.save_json:
        os.makedirs(os.path.dirname(args.save_json) or ".", exist_ok=True)
        existing = []
        if os.path.exists(args.save_json):
            try:
                with open(args.save_json) as f:
                    blob = json.load(f)
                existing = blob.get("runs", []) if isinstance(blob, dict) else []
            except Exception:
                existing = []
        run = {
            "ckpt": args.ckpt,
            "saved_at": time.strftime("%Y-%m-%d %H:%M:%S"),
            "data_config": args.data,
            "demographics_csv": args.demographics_csv,
            "seed": args.seed,
            **result,
        }
        existing.append(run)
        with open(args.save_json, "w") as f:
            json.dump({"runs": existing}, f, indent=2, default=float)
        print(f"\n>>> Appended run to {args.save_json}")


if __name__ == "__main__":
    main()