eval_reproduce_reg_write_subspace_adv.py

"""
zer0int 2026 ~ github.com/zer0int/CLIP-fine-tune
______________________________________________________________

----- VARIANT: Adversarial Dataset (included in this repo) ----- 

CLIP ViT-L/14 late-half probing + hidden-space reps + survival-aware head writes + plotting + dyslexify.

  1) Dyslexify JSON production:
       - Writes a block->disabled_heads JSON you can feed back via --dyslexify_json.
       - By default uses a survival-weighted score per head (configurable below).

  2) Dyslexify-only mode:
       - If --dyslexify_json is provided, the script runs ONLY dyslexify evaluation + related outputs (no EXP1/EXP2 full rerun).

  3) Plots (lots):
       - Per-model line plots over blocks for:
            reg_attn_mass, reg_write_norm, reg_write_survival, delta(attn_frac - write_frac)
       - Per-model heatmaps (heads x blocks) for the same.
       - Per-block (for each probed block): side-by-side pies (attn_frac vs write_frac), plus scatter attn vs write.
       - Per-model correlation curve corr(attn_frac, write_frac) over blocks.

"""

from __future__ import annotations

import os
import re
import json
import csv
import math
import argparse
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple

import numpy as np
import torch
from tqdm import tqdm

from colorama import Fore, Style, init as colorama_init
colorama_init(autoreset=True)

import warnings
warnings.filterwarnings("ignore", category=FutureWarning)
warnings.filterwarnings("ignore", category=UserWarning)
warnings.filterwarnings("ignore", category=DeprecationWarning)

import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

import attnclipdecouple as clip  # capture: last_probs/last_v

from utils_clip_loader.cliptools import fix_random_seed
fix_random_seed()

from utils_clip_loader.clip_anything_to_openai import load_openai_clip_anything


# ============================================================
# Models: OpenAI / local path .pt .safetensors / HuggingFace Hub
# ============================================================

MODELS: List[Tuple[str, str]] = [
    ("pretrained", "ViT-L/14"),
    ("gmp-clip", "zer0int/CLIP-GmP-ViT-L-14"),
    ("ko-clip", "zer0int/CLIP-KO-LITE-TypoAttack-Attn-Dropout-ViT-L-14"),
    ("regr-norm", "zer0int/CLIP-Regression-ViT-L-14"),
    ("regr-brut", "zer0int/CLIP-Regression-BRUT-ViT-L-14"),
]





@dataclass
class PairSample:
    image: Any  # PIL
    correct_label: str
    distractor_label: str
    meta: Dict[str, Any]


@dataclass
class DatasetBundle:
    """
    mode:
      - "paired": two variants with shared keys (clean vs typo) for delta-style analyses
      - "single": one split only (typo images only) -> EXP1 is skipped; EXP2 still runs
    """
    mode: str
    variant_a: str
    variant_b: str
    ordered_keys: List[str]
    map_a: Dict[str, PairSample]
    map_b: Dict[str, PairSample]
    map_single: Dict[str, PairSample]
    dataset_name: str


def _is_image_file(p: str) -> bool:
    ext = os.path.splitext(p)[1].lower()
    return ext in [".jpg", ".jpeg", ".png", ".webp", ".bmp"]


def _safe_stem(p: str) -> str:
    return os.path.splitext(os.path.basename(p))[0]


def _try_load_json(path: str) -> Optional[Any]:
    if not os.path.isfile(path):
        return None
    try:
        with open(path, "r", encoding="utf-8") as f:
            return json.load(f)
    except Exception:
        return None


def _try_load_jsonl(path: str) -> Optional[List[Dict[str, Any]]]:
    if not os.path.isfile(path):
        return None
    rows = []
    try:
        with open(path, "r", encoding="utf-8") as f:
            for line in f:
                line = line.strip()
                if not line:
                    continue
                rows.append(json.loads(line))
        return rows
    except Exception:
        return None


def _try_load_csv(path: str) -> Optional[List[Dict[str, Any]]]:
    if not os.path.isfile(path):
        return None
    try:
        with open(path, "r", encoding="utf-8") as f:
            reader = csv.DictReader(f)
            return [dict(r) for r in reader]
    except Exception:
        return None


def _normalize_meta_entry(entry: Dict[str, Any]) -> Tuple[str, str, str, Optional[str], Optional[str]]:
    """
    Returns: (img_path_or_name, correct_label, distractor_label, clean_path, typo_path)

    Supports many key spellings because your local datasets tend to evolve.
    """
    def pick(*keys, default=""):
        for k in keys:
            if k in entry and entry[k] not in (None, ""):
                return entry[k]
        return default

    img = pick("image", "img", "path", "file", "filename", "relpath", "image_path", "img_path", default="")
    correct = pick("correct_label", "object_label", "label", "gt", "class", "y", default="")
    distract = pick("distractor_label", "attack_word", "attack", "typo", "word", "target", default="")

    clean_path = pick("clean_image", "clean_path", "orig_image", "original_image", "image_clean", default=None)
    typo_path = pick("typo_image", "adv_image", "attack_image", "image_adv", "image_typo", default=None)

    return str(img), str(correct), str(distract), (str(clean_path) if clean_path else None), (str(typo_path) if typo_path else None)


def _find_first_existing(base_dir: str, candidates: List[str]) -> Optional[str]:
    for name in candidates:
        p = os.path.join(base_dir, name)
        if os.path.isfile(p):
            return p
    return None


def load_typoattack_bundle(
    base_dir: str = "image_sets/adv_dataset/typoattack",
    limit: Optional[int] = None,
) -> DatasetBundle:
    """
    Local dataset loader for typoattack.

    Priority order:
      1) Paired layout via subfolders (clean/orig vs typo/adv/attack) paired by filename stem.
      2) Metadata file describing pairs (json/jsonl/csv) with clean_path + typo_path.
      3) Single split: just images in base_dir (or in "images/" subfolder) with optional metadata for labels.

    If labels are absent, correct_label/distractor_label are set to "" (but EXP2 still runs).
    """
    base_dir = os.path.normpath(base_dir)

    # (1) folder-pair heuristic
    subdirs = {d.lower(): os.path.join(base_dir, d) for d in os.listdir(base_dir) if os.path.isdir(os.path.join(base_dir, d))}
    clean_dir = None
    typo_dir = None
    for k in ["clean", "orig", "original", "noattack", "no_attack", "base"]:
        if k in subdirs:
            clean_dir = subdirs[k]
            break
    for k in ["typo", "adv", "attack", "attacked", "adversarial", "perturbed"]:
        if k in subdirs:
            typo_dir = subdirs[k]
            break

    if clean_dir is not None and typo_dir is not None:
        clean_imgs = [os.path.join(clean_dir, f) for f in os.listdir(clean_dir) if _is_image_file(f)]
        typo_imgs = [os.path.join(typo_dir, f) for f in os.listdir(typo_dir) if _is_image_file(f)]
        clean_map = { _safe_stem(p): p for p in clean_imgs }
        typo_map = { _safe_stem(p): p for p in typo_imgs }
        keys = sorted(set(clean_map.keys()).intersection(set(typo_map.keys())))
        if limit is not None:
            keys = keys[:limit]

        from PIL import Image
        map_a, map_b = {}, {}
        for k in keys:
            map_a[k] = PairSample(
                image=Image.open(clean_map[k]).convert("RGB"),
                correct_label="",
                distractor_label="",
                meta=dict(id=k, pair_key=k, dataset="typoattack", variant="clean", path=clean_map[k]),
            )
            map_b[k] = PairSample(
                image=Image.open(typo_map[k]).convert("RGB"),
                correct_label="",
                distractor_label="",
                meta=dict(id=k, pair_key=k, dataset="typoattack", variant="typo", path=typo_map[k]),
            )
        return DatasetBundle(
            mode="paired",
            variant_a="clean",
            variant_b="typo",
            ordered_keys=keys,
            map_a=map_a,
            map_b=map_b,
            map_single={},
            dataset_name="typoattack",
        )

    # (2) metadata-driven pairing or labeling
    meta_path = _find_first_existing(base_dir, [
        "pairs.json", "annotations.json", "metadata.json", "dataset.json",
        "pairs.jsonl", "annotations.jsonl", "metadata.jsonl",
        "pairs.csv", "annotations.csv", "metadata.csv",
    ])

    meta_obj = None
    meta_rows = None
    if meta_path:
        if meta_path.endswith(".json"):
            meta_obj = _try_load_json(meta_path)
        elif meta_path.endswith(".jsonl"):
            meta_rows = _try_load_jsonl(meta_path)
        elif meta_path.endswith(".csv"):
            meta_rows = _try_load_csv(meta_path)

    # Normalize metadata rows to list[dict]
    rows: List[Dict[str, Any]] = []
    if isinstance(meta_obj, dict):
        # dict mapping -> allow { "relpath": [correct, distract] } or { "relpath": {...} }
        for k, v in meta_obj.items():
            if isinstance(v, list) and len(v) >= 2:
                rows.append({"image": k, "correct_label": v[0], "distractor_label": v[1]})
            elif isinstance(v, dict):
                vv = dict(v)
                vv.setdefault("image", k)
                rows.append(vv)
            else:
                rows.append({"image": k})
    elif isinstance(meta_obj, list):
        rows = [dict(x) for x in meta_obj]
    elif isinstance(meta_rows, list):
        rows = [dict(x) for x in meta_rows]

    # If metadata provides explicit clean+typo paths -> paired
    has_pairs = False
    for r in rows[: min(50, len(rows))]:
        _, _, _, c, t = _normalize_meta_entry(r)
        if c and t:
            has_pairs = True
            break

    if has_pairs:
        from PIL import Image
        map_a, map_b = {}, {}
        keys = []
        for r in rows:
            img, corr, dist, clean_path, typo_path = _normalize_meta_entry(r)
            if not clean_path or not typo_path:
                continue

            cpath = clean_path if os.path.isabs(clean_path) else os.path.join(base_dir, clean_path)
            tpath = typo_path if os.path.isabs(typo_path) else os.path.join(base_dir, typo_path)
            if not (os.path.isfile(cpath) and os.path.isfile(tpath)):
                continue

            key = str(r.get("pair_key", "")) or _safe_stem(cpath)
            keys.append(key)

            map_a[key] = PairSample(
                image=Image.open(cpath).convert("RGB"),
                correct_label=corr,
                distractor_label=dist,
                meta=dict(id=key, pair_key=key, dataset="typoattack", variant="clean", path=cpath, meta=r),
            )
            map_b[key] = PairSample(
                image=Image.open(tpath).convert("RGB"),
                correct_label=corr,
                distractor_label=dist,
                meta=dict(id=key, pair_key=key, dataset="typoattack", variant="typo", path=tpath, meta=r),
            )

        keys = [k for k in keys if (k in map_a and k in map_b)]
        keys = sorted(dict.fromkeys(keys))  # stable unique
        if limit is not None:
            keys = keys[:limit]

        return DatasetBundle(
            mode="paired",
            variant_a="clean",
            variant_b="typo",
            ordered_keys=keys,
            map_a={k: map_a[k] for k in keys},
            map_b={k: map_b[k] for k in keys},
            map_single={},
            dataset_name="typoattack",
        )

    # Otherwise: single split
    # Find images in base_dir or images/ subdir
    img_root = base_dir
    images_sub = os.path.join(base_dir, "images")
    if os.path.isdir(images_sub):
        img_root = images_sub

    all_imgs = []
    for root, _, files in os.walk(img_root):
        for f in files:
            if _is_image_file(f):
                all_imgs.append(os.path.join(root, f))
    all_imgs = sorted(all_imgs)
    if limit is not None:
        all_imgs = all_imgs[:limit]

    # Optional labels from metadata: map by relpath or stem
    label_by_key: Dict[str, Tuple[str, str]] = {}
    if rows:
        for r in rows:
            img, corr, dist, _, _ = _normalize_meta_entry(r)
            if not img:
                continue
            key = str(r.get("id", "")) or str(r.get("pair_key", "")) or _safe_stem(img)
            label_by_key[key] = (corr, dist)

    from PIL import Image
    map_single = {}
    keys = []
    for p in all_imgs:
        key = _safe_stem(p)
        corr, dist = label_by_key.get(key, ("", ""))
        keys.append(key)
        map_single[key] = PairSample(
            image=Image.open(p).convert("RGB"),
            correct_label=corr,
            distractor_label=dist,
            meta=dict(id=key, pair_key=key, dataset="typoattack", variant="single", path=p),
        )

    return DatasetBundle(
        mode="single",
        variant_a="single",
        variant_b="",
        ordered_keys=keys,
        map_a={},
        map_b={},
        map_single=map_single,
        dataset_name="typoattack",
    )


# ============================================================
# Math helpers (PCA + subspace angles)
# ============================================================

def pca_evr_topk(X: np.ndarray, k: int = 8) -> List[float]:
    """
    X: [N, D] (centered inside)
    returns: top-k explained variance ratios
    """
    Xc = X - X.mean(axis=0, keepdims=True)
    # economy SVD
    _, S, _ = np.linalg.svd(Xc, full_matrices=False)
    var = (S ** 2)
    evr = var / (var.sum() + 1e-12)
    return evr[:k].tolist()


def principal_angles_deg(U: np.ndarray, V: np.ndarray, k: int = 8) -> List[float]:
    """
    U: [D, k] orthonormal basis
    V: [D, k] orthonormal basis
    Returns k principal angles in degrees.
    """
    M = U.T @ V
    _, s, _ = np.linalg.svd(M, full_matrices=False)
    s = np.clip(s, -1.0, 1.0)
    ang = np.arccos(s)
    return (ang * (180.0 / np.pi)).tolist()


def orthonormal_basis_from_samples(X: np.ndarray, k: int = 8) -> np.ndarray:
    """
    X: [N, D], returns basis [D, k]
    """
    Xc = X - X.mean(axis=0, keepdims=True)
    _, _, Vt = np.linalg.svd(Xc, full_matrices=False)
    B = Vt[:k].T  # [D,k]
    return B


def decompose_delta_into_subspace(
    dZ: np.ndarray,       # [N,D]
    basis: np.ndarray,    # [D,k] orthonormal
) -> Tuple[np.ndarray, np.ndarray]:
    """
    Returns (dz_reg, dz_non) per-sample norms:
      dz_reg = ||P_basis(dZ)||
      dz_non = ||dZ - P_basis(dZ)||
    """
    # projection: P = B B^T (since orthonormal)
    proj = dZ @ basis @ basis.T  # [N,D]
    dz_reg = np.linalg.norm(proj, axis=1)
    dz_non = np.linalg.norm(dZ - proj, axis=1)
    return dz_reg, dz_non


# ============================================================
# Register mask helper
# ============================================================

def make_implicit_register_mask(
    patch_token_norms: torch.Tensor,   # [B, P]
    register_threshold: float = 70.0,
    max_registers: Optional[int] = 8,
    min_registers: int = 1
) -> torch.Tensor:
    """
    Returns bool mask [B, P] where True = register.
    NOTE: this function stays "pure". The block-13 fallback is implemented in BlockCapture.
    """
    B, P = patch_token_norms.shape
    mask = patch_token_norms > register_threshold

    if max_registers is not None:
        out = torch.zeros_like(mask)
        for b in range(B):
            idx = torch.nonzero(mask[b], as_tuple=False).flatten()
            if idx.numel() == 0:
                topk = torch.topk(patch_token_norms[b], k=min_registers, largest=True).indices
                out[b, topk] = True
            else:
                k = min(max_registers, idx.numel())
                topk = idx[torch.topk(patch_token_norms[b, idx], k=k, largest=True).indices]
                out[b, topk] = True
        return out

    out = mask.clone()
    for b in range(B):
        if out[b].sum().item() < min_registers:
            topk = torch.topk(patch_token_norms[b], k=min_registers, largest=True).indices
            out[b, topk] = True
    return out


@torch.no_grad()
def tokens_to_final_embedding(visual: torch.nn.Module, x_lnd: torch.Tensor) -> torch.Tensor:
    """
    x_lnd: [T,B,C] token stream after some block (or final block),
    runs ln_post + proj on CLS token to produce image embedding (unnormalized).
    """
    x_btn = x_lnd.permute(1, 0, 2)  # [B,T,C]
    cls_h = x_btn[:, 0, :]          # [B,C] pre-ln_post
    cls = visual.ln_post(cls_h)     # [B,C]
    if getattr(visual, "proj", None) is not None:
        cls = cls @ visual.proj     # [B,D]
    return cls


# ============================================================
# EXP plumbing via forward hooks
# ============================================================

class BlockCapture:
    """
    Captures per-block hidden reps + attention caches when capture_layers includes the block index.
    Stores only what we need (mostly CPU) to avoid VRAM bloat.

    # supports a "register reference mask" inferred from a ref block (default 13).
    If a layer has *zero* tokens above threshold, we reuse the ref mask indices for that sample.
    """
    def __init__(
        self,
        blocks: List[int],
        rep_blocks: List[int],
        survival_blocks: List[int],
        register_threshold: float,
        max_registers: int,
        min_registers: int,
        n_heads: int,
        head_dim: int,
        width: int,
        keep_gpu_for_survival: bool = True,
    ):
        self.blocks = set(blocks)
        self.rep_blocks = set(rep_blocks)
        self.survival_blocks = set(survival_blocks)

        self.register_threshold = float(register_threshold)
        self.max_registers = int(max_registers)
        self.min_registers = int(min_registers)

        self.n_heads = int(n_heads)
        self.head_dim = int(head_dim)
        self.width = int(width)

        # per-batch captures (cleared each forward)
        self.batch_cls_hidden: Dict[int, torch.Tensor] = {}
        self.batch_patch_pool: Dict[int, torch.Tensor] = {}
        self.batch_reg_pool: Dict[int, torch.Tensor] = {}

        self.batch_reg_attn_mass: Dict[int, torch.Tensor] = {}
        self.batch_reg_write_norm: Dict[int, torch.Tensor] = {}

        self.keep_gpu_for_survival = bool(keep_gpu_for_survival)
        self.batch_x_out_gpu: Dict[int, torch.Tensor] = {}
        self.batch_write_vecs_gpu: Dict[int, torch.Tensor] = {}

        self.suspend: bool = False  # IMPORTANT

        self.ref_reg_mask_device: Optional[torch.Tensor] = None  # [B,P] bool on device for the *current batch*

    def set_ref_reg_mask(self, mask_device: Optional[torch.Tensor]):
        self.ref_reg_mask_device = mask_device


    def clear_batch(self):
        self.batch_cls_hidden.clear()
        self.batch_patch_pool.clear()
        self.batch_reg_pool.clear()
        self.batch_reg_attn_mass.clear()
        self.batch_reg_write_norm.clear()
        if self.keep_gpu_for_survival:
            self.batch_x_out_gpu.clear()
            self.batch_write_vecs_gpu.clear()
        self.ref_reg_mask_device = None


    def hook_for_block(self, block_idx: int):
        def _hook(module: torch.nn.Module, inputs: Tuple[torch.Tensor, ...], output: torch.Tensor):
            if self.suspend:
                return
            if not isinstance(output, torch.Tensor) or output.dim() != 3:
                return

            x_lnd = output
            T, B, C = x_lnd.shape

            x_btn = x_lnd.permute(1, 0, 2)                 # [B,T,C]
            cls_h = x_btn[:, 0, :]                         # [B,C]
            patches = x_btn[:, 1:, :]                      # [B,P,C]
            patch_norms = patches.norm(dim=-1)             # [B,P]

            reg_mask = make_implicit_register_mask(
                patch_token_norms=patch_norms,
                register_threshold=self.register_threshold,
                max_registers=self.max_registers,
                min_registers=self.min_registers
            )                                              # [B,P] bool

            # block-13-based fallback when *none* above threshold.
            if self.ref_reg_mask_device is not None:
                # detect "none above threshold" (true none, not the min_registers fill)
                none_above = (patch_norms > self.register_threshold).sum(dim=1) == 0  # [B]
                if none_above.any():
                    ref = self.ref_reg_mask_device
                    if ref.shape == reg_mask.shape:
                        reg_mask = torch.where(none_above.unsqueeze(1), ref, reg_mask)

            non_mask = ~reg_mask
            non_cnt = non_mask.sum(dim=1).clamp_min(1).unsqueeze(-1)
            reg_cnt = reg_mask.sum(dim=1).clamp_min(1).unsqueeze(-1)
            patch_pool = (patches * non_mask.unsqueeze(-1)).sum(dim=1) / non_cnt
            reg_pool = (patches * reg_mask.unsqueeze(-1)).sum(dim=1) / reg_cnt

            if block_idx in self.rep_blocks:
                self.batch_cls_hidden[block_idx] = cls_h.detach().float().cpu()
                self.batch_patch_pool[block_idx] = patch_pool.detach().float().cpu()
                self.batch_reg_pool[block_idx] = reg_pool.detach().float().cpu()

            attn = getattr(module, "attn", None)
            if attn is None:
                return
            probs = getattr(attn, "last_probs", None)  # [B,H,T,S]
            v = getattr(attn, "last_v", None)          # [B,H,S,D]
            if probs is None or v is None:
                return
            if probs.shape[0] != B or v.shape[0] != B:
                return

            P = reg_mask.shape[1]
            S = 1 + P
            src_reg = torch.zeros((B, S), dtype=torch.bool, device=probs.device)
            src_reg[:, 1:] = reg_mask.to(device=probs.device)

            probs_cls = probs[:, :, 0, :]                                   # [B,H,S]
            reg_mass = (probs_cls * src_reg.unsqueeze(1).float()).sum(dim=-1)  # [B,H]
            reg_mass_mean = reg_mass.mean(dim=0).detach().float().cpu()     # [H]

            v_reg = v * src_reg.unsqueeze(1).unsqueeze(-1).float()          # [B,H,S,D]
            head_out = (probs_cls.unsqueeze(-1) * v_reg).sum(dim=2)         # [B,H,D]

            out_proj = getattr(attn, "out_proj", None)
            if out_proj is None:
                return
            W = out_proj.weight  # [E,E]

            writes = []
            for h in range(self.n_heads):
                w_slice = W[:, h * self.head_dim:(h + 1) * self.head_dim]   # [E,D]
                wh = head_out[:, h, :] @ w_slice.T                          # [B,E]
                writes.append(wh)
            write_vecs = torch.stack(writes, dim=1)                          # [B,H,E]
            write_norm = write_vecs.norm(dim=-1).mean(dim=0).detach().float().cpu()  # [H]

            self.batch_reg_attn_mass[block_idx] = reg_mass_mean
            self.batch_reg_write_norm[block_idx] = write_norm

            if self.keep_gpu_for_survival and (block_idx in self.survival_blocks):
                self.batch_x_out_gpu[block_idx] = x_lnd.detach()
                self.batch_write_vecs_gpu[block_idx] = write_vecs.detach()

        return _hook



# ============================================================
# Inter-model deltas  ### NEW
# ============================================================

def _align_blocks_and_heads(
    blocks_a: List[int], mat_a: np.ndarray,
    blocks_b: List[int], mat_b: np.ndarray,
) -> Tuple[List[int], np.ndarray, np.ndarray]:
    """
    Align by intersection of blocks; assumes same H.
    Returns (blocks_common, a_common, b_common)
    """
    set_a = set(blocks_a)
    set_b = set(blocks_b)
    common = sorted(set_a.intersection(set_b))
    if not common:
        return [], np.zeros((0, mat_a.shape[1]), dtype=np.float32), np.zeros((0, mat_b.shape[1]), dtype=np.float32)
    ia = [blocks_a.index(b) for b in common]
    ib = [blocks_b.index(b) for b in common]
    return common, mat_a[ia], mat_b[ib]


def _top_abs_deltas(delta_mat: np.ndarray, blocks: List[int], k: int = 12) -> List[Tuple[int, int, float]]:
    """
    Returns list of (block, head, delta) sorted by |delta| desc.
    """
    if delta_mat.size == 0:
        return []
    flat = delta_mat.reshape(-1)
    idx = np.argsort(np.abs(flat))[::-1][:k]
    out = []
    H = delta_mat.shape[1]
    for j in idx:
        bi = int(j // H)
        hi = int(j % H)
        out.append((int(blocks[bi]), hi, float(delta_mat[bi, hi])))
    return out


def write_inter_delta_summary_txt(
    out_path: str,
    model_a: str,
    model_b: str,
    blocks_common: List[int],
    d_attn: np.ndarray,
    d_write: np.ndarray,
    d_surv_blocks: List[int],
    d_surv: Optional[np.ndarray],
    d_delta_frac: np.ndarray,
):
    lines = []
    lines.append(f"Inter-model deltas: {model_a} - {model_b}")
    lines.append(f"Blocks(attn/write): {blocks_common}")
    if d_surv is not None:
        lines.append(f"Blocks(survival): {d_surv_blocks}")
    lines.append("")

    lines.append("Most stark deltas by |value|:")
    lines.append("  reg_attn_mass:")
    for b, h, v in _top_abs_deltas(d_attn, blocks_common, k=10):
        lines.append(f"    b{b:02d} h{h:02d}: {v:+.6f}")
    lines.append("  reg_write_norm:")
    for b, h, v in _top_abs_deltas(d_write, blocks_common, k=10):
        lines.append(f"    b{b:02d} h{h:02d}: {v:+.6f}")
    lines.append("  delta(attn_frac - write_frac):")
    for b, h, v in _top_abs_deltas(d_delta_frac, blocks_common, k=10):
        lines.append(f"    b{b:02d} h{h:02d}: {v:+.6f}")

    if d_surv is not None and len(d_surv_blocks) > 0:
        lines.append("  reg_write_survival:")
        for b, h, v in _top_abs_deltas(d_surv, d_surv_blocks, k=10):
            lines.append(f"    b{b:02d} h{h:02d}: {v:+.6f}")

    with open(out_path, "w", encoding="utf-8") as f:
        f.write("\n".join(lines) + "\n")


def dump_inter_delta_csvs(
    out_dir: str,
    model_a: str,
    model_b: str,
    blocks_common: List[int],
    d_attn: np.ndarray,
    d_write: np.ndarray,
    d_delta_frac: np.ndarray,
    surv_blocks_common: List[int],
    d_surv: Optional[np.ndarray],
):
    ensure_dir(out_dir)

    # head stats
    p1 = os.path.join(out_dir, "inter_delta_head_stats_blocks.csv")
    with open(p1, "w", encoding="utf-8") as f:
        f.write("model_a,model_b,block,head,d_reg_attn_mass,d_reg_write_norm,d_delta_attn_minus_write_frac\n")
        for i, b in enumerate(blocks_common):
            for h in range(d_attn.shape[1]):
                f.write(
                    f"{model_a},{model_b},{b},{h},"
                    f"{d_attn[i,h]:+.6f},{d_write[i,h]:+.6f},{d_delta_frac[i,h]:+.6f}\n"
                )

    # survival
    if d_surv is not None and len(surv_blocks_common) > 0:
        p2 = os.path.join(out_dir, "inter_delta_survival_blocks.csv")
        with open(p2, "w", encoding="utf-8") as f:
            f.write("model_a,model_b,block,head,d_reg_write_survival\n")
            for i, b in enumerate(surv_blocks_common):
                for h in range(d_surv.shape[1]):
                    f.write(f"{model_a},{model_b},{b},{h},{d_surv[i,h]:+.6f}\n")


def run_inter_model_deltas(
    out_root: str,
    results: Dict[str, Dict[str, Any]],
    base_tag: str = "pretrained",
    all_inter_deltas: bool = False,
):
    """
    results[tag] must contain:
      blocks_sorted, attn_mat, write_mat, surv_blocks_sorted, surv_mat (or None)
    """
    if base_tag not in results:
        print(Fore.YELLOW + f"[inter_deltas] base_tag '{base_tag}' not found; skipping." + Style.RESET_ALL)
        return

    inter_root = os.path.join(out_root, "_inter_deltas")
    ensure_dir(inter_root)

    tags = list(results.keys())
    base = results[base_tag]

    def do_pair(a: str, b: str):
        ra = results[a]
        rb = results[b]

        blocks_common, a_attn, b_attn = _align_blocks_and_heads(ra["blocks_sorted"], ra["attn_mat"], rb["blocks_sorted"], rb["attn_mat"])
        _, a_write, b_write = _align_blocks_and_heads(ra["blocks_sorted"], ra["write_mat"], rb["blocks_sorted"], rb["write_mat"])

        if not blocks_common:
            return

        # delta matrices
        d_attn = a_attn - b_attn
        d_write = a_write - b_write

        # delta of delta(frac): compute per-model frac then subtract
        a_attn_frac = a_attn / (a_attn.sum(axis=1, keepdims=True) + 1e-12)
        a_write_frac = a_write / (a_write.sum(axis=1, keepdims=True) + 1e-12)
        b_attn_frac = b_attn / (b_attn.sum(axis=1, keepdims=True) + 1e-12)
        b_write_frac = b_write / (b_write.sum(axis=1, keepdims=True) + 1e-12)
        d_delta_frac = (a_attn_frac - a_write_frac) - (b_attn_frac - b_write_frac)

        # survival align
        d_surv = None
        surv_blocks_common = []
        if (ra.get("surv_mat") is not None) and (rb.get("surv_mat") is not None):
            surv_blocks_common, a_surv, b_surv = _align_blocks_and_heads(
                ra["surv_blocks_sorted"], ra["surv_mat"],
                rb["surv_blocks_sorted"], rb["surv_mat"],
            )
            if surv_blocks_common:
                d_surv = a_surv - b_surv

        pair_dir = os.path.join(inter_root, f"{a}_minus_{b}")
        plots_dir = os.path.join(pair_dir, "plots")
        ensure_dir(pair_dir)
        ensure_dir(plots_dir)

        # CSV + summary
        dump_inter_delta_csvs(
            out_dir=pair_dir,
            model_a=a,
            model_b=b,
            blocks_common=blocks_common,
            d_attn=d_attn,
            d_write=d_write,
            d_delta_frac=d_delta_frac,
            surv_blocks_common=surv_blocks_common,
            d_surv=d_surv,
        )
        write_inter_delta_summary_txt(
            out_path=os.path.join(pair_dir, "summary.txt"),
            model_a=a,
            model_b=b,
            blocks_common=blocks_common,
            d_attn=d_attn,
            d_write=d_write,
            d_surv_blocks=surv_blocks_common,
            d_surv=d_surv,
            d_delta_frac=d_delta_frac,
        )

        # plots (reuse your existing helpers)
        plot_line_over_blocks(
            blocks_common, d_attn,
            out_path=os.path.join(plots_dir, f"{a}_minus_{b}__reg_attn_mass_delta_line.png"),
            title=f"{a} - {b}: Δ reg_attn_mass over blocks (per head)",
            ylabel="Δ reg_attn_mass",
            legend=True,
        )
        plot_line_over_blocks(
            blocks_common, d_write,
            out_path=os.path.join(plots_dir, f"{a}_minus_{b}__reg_write_norm_delta_line.png"),
            title=f"{a} - {b}: Δ reg_write_norm over blocks (per head)",
            ylabel="Δ reg_write_norm",
            legend=True,
        )
        plot_line_over_blocks(
            blocks_common, d_delta_frac,
            out_path=os.path.join(plots_dir, f"{a}_minus_{b}__delta_attn_minus_write_frac_delta_line.png"),
            title=f"{a} - {b}: Δ[attn_frac - write_frac] over blocks (per head)",
            ylabel="Δ(attn_frac - write_frac)",
            legend=True,
        )

        plot_heatmap(
            blocks_common, d_attn,
            out_path=os.path.join(plots_dir, f"{a}_minus_{b}__reg_attn_mass_delta_heatmap.png"),
            title=f"{a} - {b}: Δ reg_attn_mass heatmap (heads x blocks)",
        )
        plot_heatmap(
            blocks_common, d_write,
            out_path=os.path.join(plots_dir, f"{a}_minus_{b}__reg_write_norm_delta_heatmap.png"),
            title=f"{a} - {b}: Δ reg_write_norm heatmap (heads x blocks)",
        )
        plot_heatmap(
            blocks_common, d_delta_frac,
            out_path=os.path.join(plots_dir, f"{a}_minus_{b}__delta_attn_minus_write_frac_delta_heatmap.png"),
            title=f"{a} - {b}: Δ[attn_frac - write_frac] heatmap (heads x blocks)",
        )

        if d_surv is not None and len(surv_blocks_common) > 0:
            plot_line_over_blocks(
                surv_blocks_common, d_surv,
                out_path=os.path.join(plots_dir, f"{a}_minus_{b}__reg_write_survival_delta_line.png"),
                title=f"{a} - {b}: Δ reg_write_survival over blocks (per head)",
                ylabel="Δ reg_write_survival",
                legend=True,
            )
            plot_heatmap(
                surv_blocks_common, d_surv,
                out_path=os.path.join(plots_dir, f"{a}_minus_{b}__reg_write_survival_delta_heatmap.png"),
                title=f"{a} - {b}: Δ reg_write_survival heatmap (heads x blocks)",
            )

        print(Fore.CYAN + f"[inter_deltas] wrote: {pair_dir}" + Style.RESET_ALL)

    # (A) everyone vs base
    for tag in tags:
        if tag == base_tag:
            continue
        do_pair(tag, base_tag)

    # (B) optionally compare all non-base models to each other
    if all_inter_deltas:
        non_base = [t for t in tags if t != base_tag]
        for i in range(len(non_base)):
            for j in range(i + 1, len(non_base)):
                do_pair(non_base[i], non_base[j])



# ============================================================
# EXP1: hidden-space metrics
# ============================================================

def compute_exp1_hidden_metrics(
    reps_noscam: Dict[str, Dict[str, np.ndarray]],
    reps_synth: Dict[str, Dict[str, np.ndarray]],
    rep_blocks: List[int],
    k: int = 8
) -> Dict[str, Any]:
    """
    reps_*: pair_key -> {"cls_b22":..., "patch_b22":..., "reg_b22":..., ...}

    Returns:
      metrics dict with PCA EVR and principal angles for each rep choice.
    """
    keys = sorted(set(reps_noscam.keys()).intersection(set(reps_synth.keys())))
    if len(keys) < 32:
        raise RuntimeError(f"Too few paired samples for EXP1: {len(keys)}")

    out: Dict[str, Any] = {}

    # We compute for b22/b23 only (or whatever is in rep_blocks).
    # We also compute patch_delta_b23m22 if both present.
    def has(block: int, name: str) -> bool:
        k0 = keys[0]
        return f"{name}_b{block}" in reps_noscam[k0]

    # build patch_delta if possible
    if (22 in rep_blocks) and (23 in rep_blocks) and has(22, "patch") and has(23, "patch"):
        for d in (reps_noscam, reps_synth):
            for kk in keys:
                d[kk]["patch_delta_b23m22"] = d[kk]["patch_b23"] - d[kk]["patch_b22"]

    rep_specs: List[Tuple[str, str]] = []
    for b in rep_blocks:
        rep_specs.extend([
            (f"cls_b{b}", f"reg_b{b}"),
            (f"patch_b{b}", f"reg_b{b}"),
        ])
    if "patch_delta_b23m22" in reps_noscam[keys[0]]:
        rep_specs.append(("patch_delta_b23m22", "reg_b23"))

    for rep_name, reg_name in rep_specs:
        dZ = np.stack([reps_synth[kk][rep_name] - reps_noscam[kk][rep_name] for kk in keys], axis=0)  # [N,D]
        R = np.stack([reps_noscam[kk][reg_name] for kk in keys], axis=0)                               # [N,D]

        evr = pca_evr_topk(dZ, k=k)
        U = orthonormal_basis_from_samples(dZ, k=k)
        V = orthonormal_basis_from_samples(R, k=k)
        ang = principal_angles_deg(U, V, k=k)

        out[rep_name] = dict(
            n_pairs=len(keys),
            pca_evr_topk=evr,
            principal_angles_deg=ang,
        )

    return out


# ============================================================
# EXP2 survival-aware: inject per-head write vectors at block output, run remaining blocks once
# ============================================================

@torch.no_grad()
def compute_survival_scores_for_block(
    visual: torch.nn.Module,
    transformer: torch.nn.Module,
    block_idx: int,
    x_out_lnd: torch.Tensor,            # [T,B,C] after block block_idx
    write_vecs: torch.Tensor,           # [B,H,C] write vectors in width space
) -> torch.Tensor:
    """
    Returns survival score per head: mean ||delta(final_embedding)|| over batch. Shape [H].
    Vectorized across heads: remainder forward sees batch size B*H.
    """
    T, B, C = x_out_lnd.shape
    H = write_vecs.shape[1]

    # baseline remainder
    x_rem = x_out_lnd
    for j in range(block_idx + 1, transformer.layers):
        x_rem = transformer.resblocks[j](x_rem)
    emb_base = tokens_to_final_embedding(visual, x_rem)     # [B,D]

    # expanded stream for all heads
    x_rep = x_out_lnd.repeat_interleave(H, dim=1)           # [T, B*H, C]
    add = write_vecs.reshape(B * H, C)                      # [B*H,C]
    x_rep = x_rep.clone()
    x_rep[0, :, :] = x_rep[0, :, :] + add                  # inject into CLS token

    x_rem2 = x_rep
    for j in range(block_idx + 1, transformer.layers):
        x_rem2 = transformer.resblocks[j](x_rem2)
    emb_rep = tokens_to_final_embedding(visual, x_rem2)     # [B*H,D]
    emb_rep = emb_rep.view(B, H, emb_rep.shape[-1])         # [B,H,D]

    delta = emb_rep - emb_base.unsqueeze(1)                 # [B,H,D]
    scores = delta.norm(dim=-1).mean(dim=0)                 # [H]
    return scores.detach().float().cpu()


# ============================================================
# Dyslexify
# ============================================================

def load_head_mask_config(path: str) -> Dict[int, List[int]]:
    """
    Expected JSON format:
      { "18": [10,12], "22": [10], "23": [10,2] }
    Means: for each block, these heads are DISABLED (masked to 0), others kept.
    """
    with open(path, "r", encoding="utf-8") as f:
        cfg = json.load(f)
    out: Dict[int, List[int]] = {}
    for k, v in cfg.items():
        out[int(k)] = [int(x) for x in v]
    return out


def apply_head_mask(transformer: torch.nn.Module, block_to_disable_heads: Dict[int, List[int]], n_heads: int):
    """
    Sets blk.attn.head_mask for specified blocks; None otherwise.
    head_mask is 1 for keep, 0 for disable.
    """
    for i, blk in enumerate(transformer.resblocks):
        if i in block_to_disable_heads:
            disable = set(block_to_disable_heads[i])
            mask = torch.ones((n_heads,), dtype=torch.float32)
            for h in disable:
                if 0 <= h < n_heads:
                    mask[h] = 0.0
            blk.attn.head_mask = mask
        else:
            blk.attn.head_mask = None


def build_dyslexify_json_from_stats(
    blocks: List[int],
    attn: np.ndarray,      # [n_blocks, H] raw mean reg_attn_mass
    write: np.ndarray,     # [n_blocks, H] raw mean reg_write_norm
    surv: Optional[np.ndarray],  # [n_blocks, H] mean reg_write_survival, may be None
    topk: int,
    score_mode: str = "survival",
) -> Dict[int, List[int]]:
    """
    Creates mapping block->list[heads_to_disable].

    score_mode:
      - "survival": use surv directly (requires surv)
      - "attn": use attn
      - "write": use write
      - "attn_write": use attn_frac * write_frac
      - "survival_write": use surv * write_frac (requires surv)
      - "survival_attn": use surv * attn_frac (requires surv)
    """
    b2i = {b: i for i, b in enumerate(blocks)}

    # normalize to fractions for combination modes
    attn_frac = attn / (attn.sum(axis=1, keepdims=True) + 1e-12)
    write_frac = write / (write.sum(axis=1, keepdims=True) + 1e-12)

    cfg: Dict[int, List[int]] = {}
    for b in blocks:
        i = b2i[b]
        if score_mode == "survival":
            if surv is None:
                raise ValueError("score_mode=survival requires survival stats.")
            s = surv[i]
        elif score_mode == "attn":
            s = attn[i]
        elif score_mode == "write":
            s = write[i]
        elif score_mode == "attn_write":
            s = attn_frac[i] * write_frac[i]
        elif score_mode == "survival_write":
            if surv is None:
                raise ValueError("score_mode=survival_write requires survival stats.")
            s = surv[i] * write_frac[i]
        elif score_mode == "survival_attn":
            if surv is None:
                raise ValueError("score_mode=survival_attn requires survival stats.")
            s = surv[i] * attn_frac[i]
        else:
            raise ValueError(f"Unknown score_mode: {score_mode}")

        idx = np.argsort(s)[::-1][:topk]
        cfg[int(b)] = [int(x) for x in idx.tolist()]
    return cfg


# ============================================================
# Plotting helpers
# ============================================================

def _head_styles(n_heads: int = 16):
    """
    Returns list of dicts: [{'color':..., 'linestyle':..., 'marker':...}, ...]
    Deterministic:
      - base colors from tab20 (first 16)
      - head10 forced black
      - even heads solid, odd heads dashed + marker 'x'
    """
    cmap = plt.get_cmap("tab20")
    styles = []
    for h in range(n_heads):
        color = cmap(h % 20)
        if h == 10:
            color = "black"
        linestyle = "-" if (h % 2 == 0) else "--"
        marker = None if (h % 2 == 0) else "x"
        styles.append(dict(color=color, linestyle=linestyle, marker=marker))
    return styles


def plot_line_over_blocks(
    blocks: List[int],
    values: np.ndarray,  # [n_blocks, H]
    out_path: str,
    title: str,
    ylabel: str,
    legend: bool = True,
):
    H = values.shape[1]
    styles = _head_styles(H)

    plt.figure(figsize=(12, 6))
    x = np.array(blocks, dtype=int)
    for h in range(H):
        st = styles[h]
        plt.plot(
            x, values[:, h],
            label=f"h{h}",
            linewidth=1.8,
            **st
        )
    plt.title(title)
    plt.xlabel("block")
    plt.ylabel(ylabel)
    plt.grid(True, alpha=0.3)
    if legend:
        plt.legend(ncol=4, fontsize=8, frameon=False)
    plt.tight_layout()
    plt.savefig(out_path, dpi=200)
    plt.close()


def plot_heatmap(
    blocks: List[int],
    values: np.ndarray,  # [n_blocks, H]
    out_path: str,
    title: str,
    xlabel: str = "block",
    ylabel: str = "head",
):
    plt.figure(figsize=(12, 5))
    # imshow wants [H, n_blocks]
    M = values.T
    plt.imshow(M, aspect="auto", interpolation="nearest")
    plt.title(title)
    plt.xlabel(xlabel)
    plt.ylabel(ylabel)
    plt.xticks(ticks=np.arange(len(blocks)), labels=[str(b) for b in blocks], rotation=0)
    plt.yticks(ticks=np.arange(M.shape[0]), labels=[f"h{i}" for i in range(M.shape[0])])
    plt.colorbar()
    plt.tight_layout()
    plt.savefig(out_path, dpi=200)
    plt.close()


def plot_pies_for_block(
    block: int,
    attn_row: np.ndarray,   # [H]
    write_row: np.ndarray,  # [H]
    out_path: str,
    title: str,
):
    H = attn_row.shape[0]
    styles = _head_styles(H)
    colors = [styles[h]["color"] for h in range(H)]

    attn_frac = attn_row / (attn_row.sum() + 1e-12)
    write_frac = write_row / (write_row.sum() + 1e-12)

    plt.figure(figsize=(12, 6))
    plt.suptitle(title)

    ax1 = plt.subplot(1, 2, 1)
    ax1.pie(attn_frac, labels=[f"h{h}" for h in range(H)], colors=colors, textprops={"fontsize": 8})
    ax1.set_title("reg_attn_mass (fraction)")

    ax2 = plt.subplot(1, 2, 2)
    ax2.pie(write_frac, labels=[f"h{h}" for h in range(H)], colors=colors, textprops={"fontsize": 8})
    ax2.set_title("reg_write_norm (fraction)")

    plt.tight_layout(rect=[0, 0.0, 1, 0.92])
    plt.savefig(out_path, dpi=200)
    plt.close()


def plot_scatter_attn_vs_write_for_block(
    block: int,
    attn_row: np.ndarray,   # [H]
    write_row: np.ndarray,  # [H]
    out_path: str,
    title: str,
):
    H = attn_row.shape[0]
    styles = _head_styles(H)
    colors = [styles[h]["color"] for h in range(H)]

    attn_frac = attn_row / (attn_row.sum() + 1e-12)
    write_frac = write_row / (write_row.sum() + 1e-12)

    plt.figure(figsize=(7, 6))
    for h in range(H):
        plt.scatter(attn_frac[h], write_frac[h], s=60, color=colors[h])
        plt.text(attn_frac[h], write_frac[h], f"h{h}", fontsize=9)

    plt.title(title)
    plt.xlabel("attn_frac (REG mass)")
    plt.ylabel("write_frac (REG write norm)")
    plt.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(out_path, dpi=200)
    plt.close()


def plot_corr_over_blocks(
    blocks: List[int],
    attn: np.ndarray,   # [n_blocks,H]
    write: np.ndarray,  # [n_blocks,H]
    out_path: str,
    title: str,
):
    attn_frac = attn / (attn.sum(axis=1, keepdims=True) + 1e-12)
    write_frac = write / (write.sum(axis=1, keepdims=True) + 1e-12)

    cors = []
    for i in range(attn.shape[0]):
        a = attn_frac[i]
        w = write_frac[i]
        if np.std(a) < 1e-12 or np.std(w) < 1e-12:
            cors.append(0.0)
        else:
            cors.append(float(np.corrcoef(a, w)[0, 1]))

    plt.figure(figsize=(10, 4))
    plt.plot(blocks, cors, linewidth=2.0)
    plt.title(title)
    plt.xlabel("block")
    plt.ylabel("corr(attn_frac, write_frac)")
    plt.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(out_path, dpi=200)
    plt.close()


def ensure_dir(path: str):
    os.makedirs(path, exist_ok=True)


# ============================================================
# Utilities
# ============================================================

def parse_int_list(spec: str) -> List[int]:
    """
    Accepts:
      "12,13,18,22,23"
      "12-23"
      "12-23,5,7"
    """
    spec = spec.strip()
    if not spec:
        return []
    out: List[int] = []
    parts = spec.split(",")
    for p in parts:
        p = p.strip()
        if not p:
            continue
        if "-" in p:
            a, b = p.split("-")
            a = int(a.strip())
            b = int(b.strip())
            step = 1 if b >= a else -1
            out.extend(list(range(a, b + step, step)))
        else:
            out.append(int(p))
    seen = set()
    uniq = []
    for x in out:
        if x not in seen:
            seen.add(x)
            uniq.append(x)
    return uniq


def topk_heads(vals: List[float], k: int = 5) -> str:
    idx = np.argsort(np.array(vals))[::-1][:k]
    return ", ".join([f"h{int(i)}={vals[int(i)]:.4f}" for i in idx])


def dump_json(path: str, obj: Any):
    with open(path, "w", encoding="utf-8") as f:
        json.dump(obj, f, indent=2, sort_keys=True)


# ============================================================
# Dyslexify-only evaluation
# ============================================================

@torch.no_grad()
def run_dyslexify_only(
    model_tag: str,
    model_name: str,
    model: torch.nn.Module,
    preprocess,
    device: str,
    buckets: Dict[str, List[PairSample]],
    ordered_keys: List[str],
    map_n: Dict[str, PairSample],
    map_s: Dict[str, PairSample],
    blocks: List[int],
    rep_blocks: List[int],
    register_threshold: float,
    max_registers: int,
    min_registers: int,
    out_dir: str,
    dyslexify_json_path: str,
    batch_size: int,
    dz_blocks: List[int],
    dz_k: int = 8,
) -> None:
    """
    Runs ONLY dyslexify evaluation:
      - Applies head mask from JSON
      - Captures hidden reps (cls_b*, reg_b*) for dz decomposition
      - Computes dz_reg, dz_non and ratio per block in dz_blocks
      - Writes results + minimal plots

    IMPORTANT: Does not rerun full EXP1/EXP2 dumps.
    """
    ensure_dir(out_dir)
    print(Fore.YELLOW + f"[DYSLEXIFY ONLY] model={model_tag} json={dyslexify_json_path}" + Style.RESET_ALL)

    dys_cfg = load_head_mask_config(dyslexify_json_path)
    n_heads = model.visual.transformer.resblocks[0].attn.num_heads
    apply_head_mask(model.visual.transformer, dys_cfg, n_heads=n_heads)

    head_dim = model.visual.transformer.resblocks[0].attn.head_dim
    width = model.visual.transformer.width

    cap = BlockCapture(
        blocks=blocks,
        rep_blocks=rep_blocks,
        survival_blocks=[],
        register_threshold=register_threshold,
        max_registers=max_registers,
        min_registers=min_registers,
        n_heads=n_heads,
        head_dim=head_dim,
        width=width,
        keep_gpu_for_survival=False,
    )

    hooks = []
    for b in blocks:
        if 0 <= b < model.visual.transformer.layers:
            h = model.visual.transformer.resblocks[b].register_forward_hook(cap.hook_for_block(b))
            hooks.append(h)

    reps_noscam: Dict[str, Dict[str, np.ndarray]] = {}
    reps_synth: Dict[str, Dict[str, np.ndarray]] = {}

    bs = batch_size
    n_total = len(ordered_keys)

    def run_variant_batch(variant: str, key_batch: List[str]):
        cap.clear_batch()
        imgs = [map_n[k].image if variant == "NoSCAM" else map_s[k].image for k in key_batch]
        img_t = torch.stack([preprocess(im) for im in imgs], dim=0).to(device)

        _ = model.visual(
            img_t,
            return_trace=False,
            capture_layers=set(blocks),
            return_tokens=False
        )

        for b in rep_blocks:
            if b not in cap.batch_cls_hidden:
                continue
            cls_h = cap.batch_cls_hidden[b].numpy()
            reg_h = cap.batch_reg_pool[b].numpy()

            for i, k in enumerate(key_batch):
                d = reps_noscam if variant == "NoSCAM" else reps_synth
                if k not in d:
                    d[k] = {}
                d[k][f"cls_b{b}"] = cls_h[i]
                d[k][f"reg_b{b}"] = reg_h[i]

    for start in tqdm(range(0, n_total, bs), desc=f"{model_tag} dyslexify batches", leave=False):
        key_batch = ordered_keys[start:start + bs]
        run_variant_batch("NoSCAM", key_batch)
        run_variant_batch("SynthSCAM", key_batch)

    for h in hooks:
        h.remove()

    # dz decomposition per block
    keys = sorted(set(reps_noscam.keys()).intersection(set(reps_synth.keys())))
    rows = []
    for b in dz_blocks:
        rep_name = f"cls_b{b}"
        reg_name = f"reg_b{b}"
        if (keys and (rep_name not in reps_noscam[keys[0]])):
            continue

        dZ = np.stack([reps_synth[kk][rep_name] - reps_noscam[kk][rep_name] for kk in keys], axis=0)  # [N,D]
        R = np.stack([reps_noscam[kk][reg_name] for kk in keys], axis=0)                               # [N,D]
        basis = orthonormal_basis_from_samples(R, k=min(dz_k, R.shape[1], R.shape[0]))
        dz_reg, dz_non = decompose_delta_into_subspace(dZ, basis=basis)
        ratio = dz_reg / (dz_non + 1e-12)

        rows.append(dict(
            model=model_tag,
            block=b,
            n_pairs=len(keys),
            dz_reg_mean=float(np.mean(dz_reg)),
            dz_non_mean=float(np.mean(dz_non)),
            ratio_mean=float(np.mean(ratio)),
            dz_reg_median=float(np.median(dz_reg)),
            dz_non_median=float(np.median(dz_non)),
            ratio_median=float(np.median(ratio)),
        ))

    # write
    out_csv = os.path.join(out_dir, "dyslexify_dz_stats.csv")
    with open(out_csv, "w", encoding="utf-8") as f:
        f.write("model,block,n_pairs,dz_reg_mean,dz_non_mean,ratio_mean,dz_reg_median,dz_non_median,ratio_median\n")
        for r in rows:
            f.write(
                f"{r['model']},{r['block']},{r['n_pairs']},"
                f"{r['dz_reg_mean']:.6f},{r['dz_non_mean']:.6f},{r['ratio_mean']:.6f},"
                f"{r['dz_reg_median']:.6f},{r['dz_non_median']:.6f},{r['ratio_median']:.6f}\n"
            )

    # minimal plot: ratio_mean vs block
    if rows:
        plt.figure(figsize=(8, 4))
        xs = [r["block"] for r in rows]
        ys = [r["ratio_mean"] for r in rows]
        plt.plot(xs, ys, linewidth=2.0)
        plt.title("Dyslexify dz_reg/dz_non (mean) over blocks")
        plt.xlabel("block")
        plt.ylabel("ratio_mean")
        plt.grid(True, alpha=0.3)
        plt.tight_layout()
        plt.savefig(os.path.join(out_dir, "dyslexify_ratio_over_blocks.png"), dpi=200)
        plt.close()

    # summary txt
    out_txt = os.path.join(out_dir, "dyslexify_summary.txt")
    with open(out_txt, "w", encoding="utf-8") as f:
        f.write(f"Dyslexify-only evaluation\nmodel={model_tag} ({model_name})\njson={dyslexify_json_path}\n\n")
        for r in rows:
            f.write(
                f"block {r['block']}: dz_reg_mean={r['dz_reg_mean']:.6f} dz_non_mean={r['dz_non_mean']:.6f} "
                f"ratio_mean={r['ratio_mean']:.6f}\n"
            )

    print(Fore.CYAN + f"[DYSLEXIFY ONLY] wrote: {out_dir}" + Style.RESET_ALL)



@torch.no_grad()
def infer_ref_reg_mask_for_batch(
    model: torch.nn.Module,
    preprocess,
    device: str,
    samples: List[PairSample],
    ref_block: int,
    register_threshold: float,
    max_registers: int,
    min_registers: int,
) -> Optional[torch.Tensor]:
    """
    Runs a cheap forward that hooks ONLY ref_block and returns ref_reg_mask on device [B,P] bool.
    If ref_block is invalid, returns None.
    """
    if ref_block < 0 or ref_block >= model.visual.transformer.layers:
        return None

    tmp = {"mask": None}

    def _hook(_module, _inputs, output):
        if not isinstance(output, torch.Tensor) or output.dim() != 3:
            return
        x_btn = output.permute(1, 0, 2)  # [B,T,C]
        patches = x_btn[:, 1:, :]
        norms = patches.norm(dim=-1)     # [B,P]
        m = make_implicit_register_mask(
            patch_token_norms=norms,
            register_threshold=register_threshold,
            max_registers=max_registers,
            min_registers=min_registers
        )  # [B,P]
        tmp["mask"] = m.detach()

    h = model.visual.transformer.resblocks[ref_block].register_forward_hook(_hook)
    try:
        img_t = torch.stack([preprocess(s.image) for s in samples], dim=0).to(device)
        _ = model.visual(img_t, return_trace=False, capture_layers={ref_block}, return_tokens=False)
    finally:
        h.remove()

    if tmp["mask"] is None:
        return None
    return tmp["mask"].to(device=device)

# ============================================================
# Main
# ============================================================
def main():
    ap = argparse.ArgumentParser()

    ap.add_argument("--out_dir", type=str, default="out_eval_reproduce/mechaninterp_heads")
    ap.add_argument("--batch_size", type=int, default=8)
    ap.add_argument("--limit", type=int, default=None)

    ap.add_argument("--data_dir", type=str, default="image_sets/adv_dataset/typoattack")

    ap.add_argument("--register_threshold", type=float, default=70.0)
    ap.add_argument("--max_registers", type=int, default=32)
    ap.add_argument("--min_registers", type=int, default=1)

    ap.add_argument("--blocks", type=str, default="10-23")

    ap.add_argument("--ref_block", type=int, default=13)
    ap.add_argument("--no_ref_mask_fallback", action="store_true")

    # Dyslexify-only: if set, run dyslexify eval + outputs and exit
    ap.add_argument("--dyslexify_json", type=str, default=None) # Set to None, run once, then put "path/to/file.json" here

    # Dyslexify JSON production (from stats)
    ap.add_argument("--dyslexify_topk", type=int, default=3)
    ap.add_argument(
        "--dyslexify_score",
        type=str,
        default="survival",
        choices=["survival", "attn", "write", "attn_write", "survival_write", "survival_attn"],
    )
    ap.add_argument("--dyslexify_json_name", type=str, default="dyslexify_config.json")

    # Inter-model deltas
    ap.add_argument("--all_inter_deltas", action="store_true")

    args = ap.parse_args()

    os.makedirs(args.out_dir, exist_ok=True)
    device = "cuda" if torch.cuda.is_available() else "cpu"

    blocks = parse_int_list(args.blocks)
    blocks = [b for b in blocks if 0 <= b <= 10_000]  # light sanity
    if len(blocks) == 0:
        raise RuntimeError("Empty --blocks after parsing.")

    # single range drives everything
    rep_blocks = list(blocks)
    survival_blocks = list(blocks)
    dyslexify_blocks = list(blocks)
    dz_blocks = list(blocks)

    use_ref_mask_fallback = not args.no_ref_mask_fallback

    # ============================================================
    # Load dataset (auto: paired or single)
    # ============================================================
    bundle = load_typoattack_bundle(base_dir=args.data_dir, limit=args.limit)

    if bundle.mode == "single":
        ordered_keys = list(bundle.ordered_keys)
        eval_map = dict(bundle.map_single)
        eval_variant = bundle.variant_a  # "single"
    else:
        # if paired exists, default to the attacked side (variant_b)
        ordered_keys = list(bundle.ordered_keys)
        eval_map = dict(bundle.map_b)
        eval_variant = bundle.variant_b  # "typo"/"adv"/...

    if len(ordered_keys) == 0:
        raise RuntimeError("Dataset produced 0 keys/images.")

    print(Fore.CYAN + f"[data] dataset={bundle.dataset_name} mode={bundle.mode} eval_variant={eval_variant} n={len(ordered_keys)}" + Style.RESET_ALL)
    if bundle.mode != "single":
        print(Fore.CYAN + f"[data] paired variants: {bundle.variant_a} vs {bundle.variant_b} (EXP1 removed; evaluating '{eval_variant}' only)" + Style.RESET_ALL)

    # ============================================================
    # Helper: run EXP2 (+plots/+optional dyslexify cfg) for a model
    # ============================================================
    @torch.no_grad()
    def run_exp2_for_model(
        tag: str,
        name: str,
        model: torch.nn.Module,
        preprocess,
        model_out: str,
        *,
        head_mask_json: Optional[str] = None,
        make_dyslexify_cfg: bool = True,
    ) -> Dict[str, Any]:
        """
        Returns dict with:
          blocks_sorted, attn_mat, write_mat, surv_blocks_sorted, surv_mat
        """
        ensure_dir(model_out)
        plots_dir = os.path.join(model_out, "plots")
        ensure_dir(plots_dir)

        attn0 = model.visual.transformer.resblocks[0].attn
        if not hasattr(attn0, "last_v") or not hasattr(attn0, "last_probs"):
            raise RuntimeError("This script requires attnclipdecouple (needs last_v/last_probs capture).")

        n_heads = model.visual.transformer.resblocks[0].attn.num_heads
        head_dim = model.visual.transformer.resblocks[0].attn.head_dim
        width = model.visual.transformer.width

        # optional dyslexify head mask apply
        if head_mask_json is not None:
            dys_cfg = load_head_mask_config(head_mask_json)
            apply_head_mask(model.visual.transformer, dys_cfg, n_heads=n_heads)
        else:
            # ensure clean state
            apply_head_mask(model.visual.transformer, {}, n_heads=n_heads)

        cap = BlockCapture(
            blocks=blocks,
            rep_blocks=rep_blocks,                 # kept for compatibility; EXP1 is removed -> only in SCAM variant
            survival_blocks=survival_blocks,
            register_threshold=args.register_threshold,
            max_registers=args.max_registers,
            min_registers=args.min_registers,
            n_heads=n_heads,
            head_dim=head_dim,
            width=width,
            keep_gpu_for_survival=True,
        )

        hooks = []
        for b in blocks:
            if 0 <= b < model.visual.transformer.layers:
                h = model.visual.transformer.resblocks[b].register_forward_hook(cap.hook_for_block(b))
                hooks.append(h)

        sum_reg_attn_mass: Dict[int, torch.Tensor] = {}
        sum_reg_write_norm: Dict[int, torch.Tensor] = {}
        count_batches: Dict[int, int] = {}

        surv_sum: Dict[int, torch.Tensor] = {}
        surv_count: Dict[int, int] = {}

        bs = args.batch_size
        n_total = len(ordered_keys)

        for start in tqdm(range(0, n_total, bs), desc=f"{tag} batches", leave=False):
            key_batch = ordered_keys[start : start + bs]
            samples = [eval_map[k] for k in key_batch]

            cap.clear_batch()

            # block-13 ref-mask fallback (one cheap forward for ref block)
            if use_ref_mask_fallback:
                ref_mask = infer_ref_reg_mask_for_batch(
                    model=model,
                    preprocess=preprocess,
                    device=device,
                    samples=samples,
                    ref_block=args.ref_block,
                    register_threshold=args.register_threshold,
                    max_registers=args.max_registers,
                    min_registers=args.min_registers,
                )
                cap.set_ref_reg_mask(ref_mask)
            else:
                cap.set_ref_reg_mask(None)

            img_t = torch.stack([preprocess(s.image) for s in samples], dim=0).to(device)

            _ = model.visual(
                img_t,
                return_trace=False,
                capture_layers=set(blocks),
                return_tokens=False,
            )

            # aggregate head stats
            for b in cap.batch_reg_attn_mass.keys():
                if b not in sum_reg_attn_mass:
                    sum_reg_attn_mass[b] = torch.zeros((n_heads,), dtype=torch.float32)
                    sum_reg_write_norm[b] = torch.zeros((n_heads,), dtype=torch.float32)
                    count_batches[b] = 0
                sum_reg_attn_mass[b] += cap.batch_reg_attn_mass[b]
                sum_reg_write_norm[b] += cap.batch_reg_write_norm[b]
                count_batches[b] += 1

            # survival-aware
            for b in survival_blocks:
                if b not in cap.batch_x_out_gpu or b not in cap.batch_write_vecs_gpu:
                    continue
                cap.suspend = True  # IMPORTANT: prevent hooks firing during expanded-batch remainder
                try:
                    scores = compute_survival_scores_for_block(
                        visual=model.visual,
                        transformer=model.visual.transformer,
                        block_idx=b,
                        x_out_lnd=cap.batch_x_out_gpu[b],
                        write_vecs=cap.batch_write_vecs_gpu[b],
                    )  # [H] on CPU
                finally:
                    cap.suspend = False

                if b not in surv_sum:
                    surv_sum[b] = torch.zeros((n_heads,), dtype=torch.float32)
                    surv_count[b] = 0
                surv_sum[b] += scores
                surv_count[b] += 1

        for h in hooks:
            h.remove()

        # means arrays aligned to blocks list order
        blocks_sorted = sorted([b for b in blocks if b in sum_reg_attn_mass])
        attn_mat = np.zeros((len(blocks_sorted), n_heads), dtype=np.float32)
        write_mat = np.zeros((len(blocks_sorted), n_heads), dtype=np.float32)
        for i, b in enumerate(blocks_sorted):
            denom = max(1, count_batches.get(b, 1))
            attn_mat[i] = (sum_reg_attn_mass[b] / denom).numpy()
            write_mat[i] = (sum_reg_write_norm[b] / denom).numpy()

        surv_blocks_sorted = sorted([b for b in survival_blocks if b in surv_sum])
        surv_mat = None
        if len(surv_blocks_sorted) > 0:
            surv_mat = np.zeros((len(surv_blocks_sorted), n_heads), dtype=np.float32)
            for i, b in enumerate(surv_blocks_sorted):
                denom = max(1, surv_count.get(b, 1))
                surv_mat[i] = (surv_sum[b] / denom).numpy()

        # CSV dumps
        with open(os.path.join(model_out, "exp2_head_stats_blocks.csv"), "w", encoding="utf-8") as f:
            f.write("model,block,head,reg_attn_mass,reg_write_norm\n")
            for i, b in enumerate(blocks_sorted):
                for hh in range(n_heads):
                    f.write(f"{tag},{b},{hh},{attn_mat[i,hh]:.6f},{write_mat[i,hh]:.6f}\n")

        with open(os.path.join(model_out, "exp2_survival_blocks.csv"), "w", encoding="utf-8") as f:
            f.write("model,block,head,reg_write_survival\n")
            if surv_mat is not None:
                for i, b in enumerate(surv_blocks_sorted):
                    for hh in range(n_heads):
                        f.write(f"{tag},{b},{hh},{surv_mat[i,hh]:.6f}\n")

        # summary.txt (EXP1 removed)
        lines = []
        lines.append(f"=== Model: {tag} | {name} ===")
        lines.append(f"dataset={bundle.dataset_name} mode={bundle.mode} eval_variant={eval_variant} n={len(ordered_keys)}")
        if head_mask_json is not None:
            lines.append(f"dyslexify_json={head_mask_json}")
        lines.append(f"blocks={blocks_sorted}")
        lines.append(f"survival_blocks={surv_blocks_sorted}")
        lines.append("")

        # top heads for a few blocks if present
        for b in [18, 22, 23]:
            if b in blocks_sorted:
                i = blocks_sorted.index(b)
                lines.append(f"EXP2 b{b} top reg_write_norm heads: " + topk_heads(write_mat[i].tolist(), 5))
                lines.append(f"EXP2 b{b} top reg_attn_mass heads: " + topk_heads(attn_mat[i].tolist(), 5))

        if surv_mat is not None:
            for b in [18, 19, 20, 21, 22, 23]:
                if b in surv_blocks_sorted:
                    i = surv_blocks_sorted.index(b)
                    lines.append(f"EXP2(survival) b{b} top reg_write_survival heads: " + topk_heads(surv_mat[i].tolist(), 5))

        with open(os.path.join(model_out, "summary.txt"), "w", encoding="utf-8") as f:
            f.write("\n".join(lines) + "\n")

        # Plots
        if len(blocks_sorted) > 0:
            plot_line_over_blocks(
                blocks_sorted, attn_mat,
                out_path=os.path.join(plots_dir, f"{tag}_reg_attn_mass_line.png"),
                title=f"{tag}: reg_attn_mass over blocks (per head)",
                ylabel="reg_attn_mass",
                legend=True,
            )
            plot_line_over_blocks(
                blocks_sorted, write_mat,
                out_path=os.path.join(plots_dir, f"{tag}_reg_write_norm_line.png"),
                title=f"{tag}: reg_write_norm over blocks (per head)",
                ylabel="reg_write_norm",
                legend=True,
            )

            attn_frac = attn_mat / (attn_mat.sum(axis=1, keepdims=True) + 1e-12)
            write_frac = write_mat / (write_mat.sum(axis=1, keepdims=True) + 1e-12)
            delta_frac = attn_frac - write_frac

            plot_line_over_blocks(
                blocks_sorted, delta_frac,
                out_path=os.path.join(plots_dir, f"{tag}_delta_attn_minus_write_frac_line.png"),
                title=f"{tag}: delta(attn_frac - write_frac) over blocks (per head)",
                ylabel="attn_frac - write_frac",
                legend=True,
            )

            plot_heatmap(
                blocks_sorted, attn_mat,
                out_path=os.path.join(plots_dir, f"{tag}_reg_attn_mass_heatmap.png"),
                title=f"{tag}: reg_attn_mass heatmap (heads x blocks)",
            )
            plot_heatmap(
                blocks_sorted, write_mat,
                out_path=os.path.join(plots_dir, f"{tag}_reg_write_norm_heatmap.png"),
                title=f"{tag}: reg_write_norm heatmap (heads x blocks)",
            )
            plot_heatmap(
                blocks_sorted, delta_frac,
                out_path=os.path.join(plots_dir, f"{tag}_delta_attn_minus_write_frac_heatmap.png"),
                title=f"{tag}: delta(attn_frac - write_frac) heatmap (heads x blocks)",
            )

            for i, b in enumerate(blocks_sorted):
                plot_pies_for_block(
                    block=b,
                    attn_row=attn_mat[i],
                    write_row=write_mat[i],
                    out_path=os.path.join(plots_dir, f"{tag}_b{b:02d}_pies_attn_vs_write.png"),
                    title=f"{tag} block {b}: REG attention vs REG write (fractions)",
                )
                plot_scatter_attn_vs_write_for_block(
                    block=b,
                    attn_row=attn_mat[i],
                    write_row=write_mat[i],
                    out_path=os.path.join(plots_dir, f"{tag}_b{b:02d}_scatter_attn_vs_write.png"),
                    title=f"{tag} block {b}: attn_frac vs write_frac (per head)",
                )

            plot_corr_over_blocks(
                blocks_sorted, attn_mat, write_mat,
                out_path=os.path.join(plots_dir, f"{tag}_corr_attn_write_over_blocks.png"),
                title=f"{tag}: corr(attn_frac, write_frac) over blocks",
            )

        if surv_mat is not None and len(surv_blocks_sorted) > 0:
            plot_line_over_blocks(
                surv_blocks_sorted, surv_mat,
                out_path=os.path.join(plots_dir, f"{tag}_reg_write_survival_line.png"),
                title=f"{tag}: reg_write_survival over blocks (per head)",
                ylabel="reg_write_survival",
                legend=True,
            )
            plot_heatmap(
                surv_blocks_sorted, surv_mat,
                out_path=os.path.join(plots_dir, f"{tag}_reg_write_survival_heatmap.png"),
                title=f"{tag}: reg_write_survival heatmap (heads x blocks)",
            )

        # Dyslexify JSON production (optional; uses the same single range)
        if make_dyslexify_cfg and len(blocks_sorted) > 0:
            dblocks = [b for b in dyslexify_blocks if b in blocks_sorted]
            if len(dblocks) == 0:
                print(Fore.YELLOW + f"[dyslexify_json] no overlapping blocks for {tag}" + Style.RESET_ALL)
            else:
                idx = [blocks_sorted.index(b) for b in dblocks]
                attn_sub = attn_mat[idx]
                write_sub = write_mat[idx]

                surv_sub = None
                if args.dyslexify_score.startswith("survival"):
                    if surv_mat is None or (not all(b in surv_blocks_sorted for b in dblocks)):
                        raise RuntimeError(
                            f"--dyslexify_score={args.dyslexify_score} requires survival for all blocks in range. "
                            f"Have survival_blocks={surv_blocks_sorted}, want={dblocks}"
                        )
                    sidx = [surv_blocks_sorted.index(b) for b in dblocks]
                    surv_sub = surv_mat[sidx]

                cfg = build_dyslexify_json_from_stats(
                    blocks=dblocks,
                    attn=attn_sub,
                    write=write_sub,
                    surv=surv_sub,
                    topk=args.dyslexify_topk,
                    score_mode=args.dyslexify_score,
                )

                cfg_path = os.path.join(model_out, args.dyslexify_json_name)
                dump_json(cfg_path, {str(k): v for k, v in cfg.items()})

                txt_path = os.path.join(model_out, "dyslexify_config_rationale.txt")
                with open(txt_path, "w", encoding="utf-8") as f:
                    f.write(
                        f"Dyslexify config generation\n"
                        f"model={tag}\nscore_mode={args.dyslexify_score}\n"
                        f"topk={args.dyslexify_topk}\nblocks={dblocks}\n\n"
                    )
                    for bi, b in enumerate(dblocks):
                        f.write(f"block {b} disable: {cfg[b]}\n")
                        f.write(f"  top attn:  {topk_heads((attn_sub[bi]).tolist(), 8)}\n")
                        f.write(f"  top write: {topk_heads((write_sub[bi]).tolist(), 8)}\n")
                        if surv_sub is not None:
                            f.write(f"  top surv:  {topk_heads((surv_sub[bi]).tolist(), 8)}\n")
                        f.write("\n")

                print(Fore.CYAN + f"[dyslexify_json] wrote: {cfg_path}" + Style.RESET_ALL)

        # console tail
        print(Fore.GREEN + "\n".join(lines[-12:]) + Style.RESET_ALL)
        print(Fore.CYAN + f"Wrote outputs to: {model_out}" + Style.RESET_ALL)

        return dict(
            blocks_sorted=blocks_sorted,
            attn_mat=attn_mat,
            write_mat=write_mat,
            surv_blocks_sorted=surv_blocks_sorted,
            surv_mat=surv_mat,
        )

    # ============================================================
    # Dyslexify-only mode: apply provided JSON, run EXP2(+plots), exit
    # ============================================================
    if args.dyslexify_json is not None:
        for tag, name in MODELS:
            print(Fore.MAGENTA + f"\n=== Dyslexify-only Model: {tag} | {name} ===" + Style.RESET_ALL)

            model, preprocess, _ = load_openai_clip_anything(clip, name, device=device, jit=False, strict=True)
            model = model.to(device).eval().float()

            model_out = os.path.join(args.out_dir, tag, "dyslexify_only")
            _ = run_exp2_for_model(
                tag=tag,
                name=name,
                model=model,
                preprocess=preprocess,
                model_out=model_out,
                head_mask_json=args.dyslexify_json,
                make_dyslexify_cfg=False,
            )
        return

    # ============================================================
    # Full run (EXP2 only; EXP1 removed) + inter-model deltas
    # ============================================================
    results: Dict[str, Dict[str, Any]] = {}

    for tag, name in MODELS:
        print(Fore.MAGENTA + f"\n=== Model: {tag} | {name} ===" + Style.RESET_ALL)

        model, preprocess, _ = load_openai_clip_anything(clip, name, device=device, jit=False, strict=True)
        model = model.to(device).eval().float()

        model_out = os.path.join(args.out_dir, tag)
        results[tag] = run_exp2_for_model(
            tag=tag,
            name=name,
            model=model,
            preprocess=preprocess,
            model_out=model_out,
            head_mask_json=None,
            make_dyslexify_cfg=True,
        )

    # inter-model deltas (pretrained baseline by default)
    run_inter_model_deltas(
        out_root=args.out_dir,
        results=results,
        base_tag="pretrained",
        all_inter_deltas=args.all_inter_deltas,
    )


if __name__ == "__main__":
    main()