mechanistic_interp/compare_baselines.py

"""
Compare editing-method effects on the toilet readout, per layer.

Idea
────
For each model variant (base, lora, efuf, nullu) read the *toilet* attention
probe score σ(logit) ∈ [0,1] at every layer, on bathroom-only images, and compare
each method to base:

    cell[method, l] = mean_images( σ_toilet_l(method) − σ_toilet_l(base) )

NO gradient steering — the *edit itself* is the intervention. To isolate the
representational effect of the edit (not a change in what the model says), every
method reads the SAME fixed context per image: the BASE model's generated caption
(or --forced_text). The probe pools over those caption tokens; only the model
weights differ between rows. The base row is therefore exactly 0 (reference).

Images: bathroom-only (bathroom=1 & toilet=0), split into two populations by the
base model's hallucination flag (base_mentions_object on --base_prompt):
  • NON-hallucinating (base did NOT say toilet)
  • HALLUCINATING     (base DID say toilet)
One figure, two panels (non | halluc), each a (methods × layers) heatmap; base − base = 0.

Models are loaded ONE AT A TIME (base first, to fix the captions), so we never hold
four 7B models on the GPU simultaneously.

Saves: heatmap PNG + the raw (methods × layers) matrices as JSON.
"""

import argparse
import gc
import json
import os
import random

import numpy as np
import torch as t
from PIL import Image
from transformers import AutoConfig, LlavaProcessor

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

from model.llava.hooked_llava import HookedSAELlavaConditionalGeneration
from mechanistic_interp.sequence_probe import sequence_layer_probes_from_checkpoint

_HOOK_SUFFIX = {"pre": "hook_resid_pre", "mid": "hook_resid_mid", "post": "hook_resid_post"}


# ── Model-edit helpers (canonical home; gradient_ascent.py / train_probe_latent.py
#    import apply_efuf_edit / apply_nullu_edit from here). ──────────────────────────
# EFUF overwrites the multimodal projector; map each .pth key -> (submodule, attr).
_EFUF_PROJ_KEYS = {
    "model.mm_projector.0.weight": ("linear_1", "weight"),
    "model.mm_projector.0.bias":   ("linear_1", "bias"),
    "model.mm_projector.2.weight": ("linear_2", "weight"),
    "model.mm_projector.2.bias":   ("linear_2", "bias"),
}
# Nullu splices mlp.down_proj.weight; accept the common HF key prefixes.
_DOWN_PROJ_PREFIXES = ("model.layers", "language_model.model.layers",
                       "language_model.layers", "model.language_model.layers")
_DOWN_PROJ_SUFFIX = ".mlp.down_proj.weight"


def apply_efuf_edit(model, ckpt_path):
    """Overwrite the multimodal projector in-place from an EFUF .pth checkpoint."""
    raw = t.load(ckpt_path, map_location="cpu", weights_only=True)
    missing = [k for k in _EFUF_PROJ_KEYS if k not in raw]
    if missing:
        raise KeyError(f"EFUF ckpt {ckpt_path} missing {missing}; has {list(raw.keys())}")
    proj = model.multi_modal_projector
    with t.no_grad():
        for key, (lin, attr) in _EFUF_PROJ_KEYS.items():
            tgt = getattr(getattr(proj, lin), attr)
            tgt.data.copy_(raw[key].to(device=tgt.device, dtype=tgt.dtype))
    return len(_EFUF_PROJ_KEYS)


def _match_down_proj_layer(key):
    """Return the layer index if `key` is an mlp.down_proj.weight, else None."""
    if not key.endswith(_DOWN_PROJ_SUFFIX):
        return None
    for prefix in _DOWN_PROJ_PREFIXES:
        if key.startswith(prefix + "."):
            mid = key[len(prefix) + 1: -len(_DOWN_PROJ_SUFFIX)]
            if mid.isdigit():
                return int(mid)
    return None


def apply_nullu_edit(model, lm_layers, ckpt_dir, lowest, highest):
    """Splice Nullu-edited mlp.down_proj.weight into layers [lowest, highest) in-place."""
    import glob
    from safetensors.torch import safe_open
    wanted = set(range(lowest, highest))
    found = {}
    shards = sorted(glob.glob(os.path.join(ckpt_dir, "*.safetensors")))
    if shards:
        for sh in shards:
            with safe_open(sh, framework="pt") as f:
                for key in f.keys():
                    idx = _match_down_proj_layer(key)
                    if idx is not None and idx in wanted and idx not in found:
                        found[idx] = f.get_tensor(key)
            if len(found) == len(wanted):
                break
    else:
        for bf in sorted(glob.glob(os.path.join(ckpt_dir, "*.bin"))):
            sd = t.load(bf, map_location="cpu", weights_only=True)
            for key, tns in sd.items():
                idx = _match_down_proj_layer(key)
                if idx is not None and idx in wanted and idx not in found:
                    found[idx] = tns
            if len(found) == len(wanted):
                break
    missing = wanted - set(found)
    if missing:
        raise KeyError(f"Nullu dir {ckpt_dir} missing down_proj for layers {sorted(missing)}")
    with t.no_grad():
        for idx, w in found.items():
            tgt = lm_layers[idx].mlp.down_proj.weight
            tgt.data.copy_(w.to(device=tgt.device, dtype=tgt.dtype))
    return len(found)


# ── Model-variant loading (base / lora / nullu / efuf), per attribution_patching ──
def _decoder_layers(model):
    lm = getattr(model, "language_model", None) or getattr(getattr(model, "model", None), "language_model", None)
    inner = getattr(lm, "model", None)
    if inner is not None and hasattr(inner, "layers"):
        return inner.layers
    return lm.layers


def build_model(variant, args, dtype, device):
    """Build one model variant. PROBES are loaded separately (base-trained); only the
    MODEL weights change here. Caller is responsible for freeing the previous model."""
    if variant == "lora":
        from model.llava.hooked_lora_llava import HookedLoRALlava
        model = HookedLoRALlava.from_pretrained(
            args.model_name, torch_dtype=dtype, device_map={"": device}).eval()
        lora_dir = args.lora_path if os.path.isdir(args.lora_path) else os.path.dirname(args.lora_path)
        model.load_lora_adapter(lora_dir, merge=True)
        print(f"[compare] model=lora (merged {lora_dir})")
    else:
        model = HookedSAELlavaConditionalGeneration.from_pretrained(
            args.model_name, torch_dtype=dtype, device_map={"": device}).eval()
        if variant == "efuf":
            n = apply_efuf_edit(model, args.efuf_path)
            print(f"[compare] model=efuf ({n} proj tensors, {args.efuf_path})")
        elif variant == "nullu":
            n = apply_nullu_edit(model, _decoder_layers(model), args.nullu_path,
                                 args.nullu_lowest, args.nullu_highest)
            print(f"[compare] model=nullu ({n} layers [{args.nullu_lowest},{args.nullu_highest}), {args.nullu_path})")
        else:
            print(f"[compare] model=base")
    return model


def free_model(model):
    del model
    gc.collect()
    if t.cuda.is_available():
        t.cuda.empty_cache()


def hp_name(layer: int, hook_type: str) -> str:
    return f"model.language_model.layers.{layer}.{_HOOK_SUFFIX[hook_type]}"


def caption_slice(attn_mask, asst_text, processor, max_seq_tokens):
    """(start, end) indices of the caption tokens at the tail of the real region,
    matching training's a[:max_seq_tokens]."""
    seq_len = int(attn_mask.sum().item())
    cap_len = len(processor.tokenizer(asst_text, add_special_tokens=False)["input_ids"])
    if cap_len <= 0:
        return None
    start = seq_len - cap_len
    end = start + min(cap_len, max_seq_tokens)
    if start < 0:
        return None
    return start, end


@t.no_grad()
def generate_caption(model, processor, image, question, device, max_new_tokens):
    prompt = f"USER: <image>\n{question}\nASSISTANT:"
    inp = processor(images=[image], text=[prompt], return_tensors="pt").to(device)
    out = model.generate(**inp, do_sample=False, num_beams=1, use_cache=True,
                         max_new_tokens=max_new_tokens)
    cap = processor.batch_decode(out, skip_special_tokens=True)[0]
    return cap.split("ASSISTANT:")[-1].strip()


def probe_logit(probe_module, layer, seq_feats):
    """Scalar logit for one image from one layer's attention probe. seq_feats: (1,T,d)."""
    kpm = t.zeros(seq_feats.shape[:2], dtype=t.bool, device=seq_feats.device)  # no padding
    return probe_module.probes[probe_module._idx[layer]](seq_feats, kpm).squeeze(0)


def probe_score(probe_module, layer, seq_feats):
    """Bounded probe score σ(logit) = P(concept present) ∈ [0,1]."""
    return t.sigmoid(probe_logit(probe_module, layer, seq_feats))


@t.no_grad()
def toilet_scores_per_layer(model, processor, toilet, img, asst, question,
                            device, hps, layers, max_seq_tokens):
    """Forward `img` with the fixed ASSISTANT answer `asst`, return σ_toilet at every
    layer (np array of length n_layers), or None if the caption slice is invalid."""
    forced = f"USER: <image>\n{question}\nASSISTANT: {asst}"
    fwd = processor(images=[img], text=[forced], return_tensors="pt").to(device)
    sl = caption_slice(fwd["attention_mask"], asst, processor, max_seq_tokens)
    if sl is None:
        return None
    s0, s1 = sl
    acts = {}
    def cap(name):
        def _fn(act, hook):
            acts[name] = act
        return _fn
    model.run_with_hooks(fwd, fwd_hooks=[(hp, cap(hp)) for hp in hps])
    out = np.empty(len(layers), dtype=np.float64)
    for l in layers:
        feats = acts[hps[l]][:, s0:s1].float()
        out[l] = float(probe_score(toilet, l, feats).item())
    return out


def plot_method_layer(mats, methods, layers, out, title, cbar_label="Δ toilet score σ  (method − base)"):
    """mats: {group_name: (n_methods, n_layers) np array}. One panel per group, shared
    symmetric diverging scale + single colorbar. methods on y, layers on x."""
    groups = [g for g in mats if mats[g] is not None]
    ncol = len(groups)
    fig, axes = plt.subplots(1, ncol, figsize=(0.30 * len(layers) + 2.0, 0.55 * len(methods) + 2.0),
                             squeeze=False)
    allv = np.concatenate([mats[g][np.isfinite(mats[g])].ravel() for g in groups])
    vmax = float(np.nanmax(np.abs(allv))) if allv.size else 1.0
    vmax = max(vmax, 1e-6)
    im = None
    for k, g in enumerate(groups):
        ax = axes[0][k]
        im = ax.imshow(mats[g], origin="upper", cmap="RdBu_r", vmin=-vmax, vmax=vmax, aspect="auto")
        ax.set_title(g)
        ax.set_xlabel("layer")
        ax.set_yticks(range(len(methods)))
        ax.set_yticklabels(methods)
        ax.set_xticks(range(0, len(layers), 4))
    cbar = fig.colorbar(im, ax=axes, fraction=0.025, pad=0.02)
    cbar.set_label(cbar_label)
    fig.suptitle(title, fontsize=12)
    os.makedirs(os.path.dirname(out), exist_ok=True)
    fig.savefig(out, dpi=150, bbox_inches="tight")
    plt.close(fig)
    print(f"[compare] saved {out}")


def main():
    ap = argparse.ArgumentParser(description=__doc__,
                                 formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--model_name", default="llava-hf/llava-1.5-7b-hf")
    ap.add_argument("--device_id", type=int, default=0)
    ap.add_argument("--dtype", default="bfloat16", choices=["float32", "float16", "bfloat16"])

    ap.add_argument("--methods", nargs="+", default=["base", "lora", "efuf", "nullu"],
                    choices=["base", "lora", "efuf", "nullu"],
                    help="Variants to compare. 'base' is the reference (its row is all 0); "
                         "it is always loaded first to fix the captions even if omitted here.")

    # Edit checkpoints (only loaded if the method is requested).
    ap.add_argument("--lora_path", default="/data/caotue/multilayer-sae/adv_gen_outputs/run_bathroom_toilet_v2/lora_adapter")
    ap.add_argument("--efuf_path", default="/data/caotue/multilayer-sae/EFUF/efuf/checkpoints/llava_vicuna_7b/bathroom_toilet_paper_10ep/epoch_002.pth")
    ap.add_argument("--nullu_path", default="/data/caotue/nullu/edited_models/LLaVA-7B-top4-0-32-bathroom_toilet")
    ap.add_argument("--nullu_lowest", type=int, default=8)
    ap.add_argument("--nullu_highest", type=int, default=32)

    # Probe — base-trained toilet readout.
    ap.add_argument("--toilet_probe", default="/data/caotue/latent_probes/seqprobes_toilet/post/seqprobe.pth")

    ap.add_argument("--image_folder", default="/data/caotue/CC3M-Dataset/cc3m_images")
    ap.add_argument("--samples_json", default="mechanistic_interp/toilet_bathroom/samples.json")
    ap.add_argument("--category", default="bathroom_only",
                    choices=["bathroom_only", "bathroom_toilet", "toilet_only", "others"],
                    help="Dataset to evaluate. bathroom_only = toilet ABSENT (base mention = "
                         "hallucination); bathroom_toilet / toilet_only = toilet PRESENT (Δσ<0 = "
                         "edit suppresses a real toilet); others = unrelated negatives from "
                         "--others_jsonl (single 'unrelated' panel; Δσ≠0 = edit spuriously moves toilet).")
    ap.add_argument("--others_jsonl", default="mechanistic_interp/neg_cc3m_5k.json",
                    help="For --category others: {train:[ids], validation:[ids]} of CC3M stems.")
    ap.add_argument("--others_split", default="validation", choices=["validation", "train"],
                    help="Which split of --others_jsonl to use (default validation = held out "
                         "from the 4-variant probe's negatives).")
    ap.add_argument("--base_prompt", default="Describe this image.",
                    help="prompt_results key whose base_mentions_object flag splits the two populations.")
    ap.add_argument("--population", default="both",
                    choices=["all", "both", "said", "unsaid", "halluc", "non"],
                    help="How to group the category: 'all' = ONE set, no base-mention split; "
                         "said/halluc = base DID mention toilet; unsaid/non = did not; both = both panels.")
    ap.add_argument("--id_col", default="image_id")
    ap.add_argument("--question", default="Describe this image.",
                    help="The text fed to every model (USER turn).")
    ap.add_argument("--forced_text", default="This image features a bathroom with a",
                    help="Constant ASSISTANT answer used as the fixed context for ALL "
                         "images/methods (the probe reads these tokens). Set to '' to instead "
                         "use the base model's per-image generated caption.")
    ap.add_argument("--num_images", type=int, default=100,
                    help="Random-subsample cap PER population (0 = ALL). Shuffled by --seed.")
    ap.add_argument("--seed", type=int, default=0)

    ap.add_argument("--hook_type", default="post", choices=["pre", "mid", "post"])
    ap.add_argument("--max_new_tokens", type=int, default=64)
    ap.add_argument("--max_seq_tokens", type=int, default=64)

    ap.add_argument("--out", default="mechanistic_interp/graph/compare_baselines_toilet.png")
    ap.add_argument("--out_json", default="mechanistic_interp/graph/compare_baselines_toilet.json")
    args = ap.parse_args()

    dtype = {"float32": t.float32, "float16": t.float16, "bfloat16": t.bfloat16}[args.dtype]
    device = f"cuda:{args.device_id}" if t.cuda.is_available() else "cpu"

    n_layers = AutoConfig.from_pretrained(args.model_name).text_config.num_hidden_layers
    layers = list(range(n_layers))
    hps = [hp_name(l, args.hook_type) for l in layers]

    # base must run first (it fixes the per-image captions); keep requested order otherwise.
    methods = ["base"] + [m for m in args.methods if m != "base"]
    print(f"[compare] device={device} dtype={dtype} layers={n_layers}")
    print(f"[compare] methods={methods}  readout=toilet ({args.toilet_probe})")
    print(f"[compare] context={'forced: ' + repr(args.forced_text) if args.forced_text else 'base caption per image'}")

    # ── Select images for the chosen dataset ──────────────────────────────────────
    stem_to_file = {}
    for root, _, files in os.walk(args.image_folder):
        for fn in files:
            if fn.lower().endswith((".jpg", ".jpeg", ".png", ".webp")):
                stem_to_file[os.path.splitext(fn)[0]] = os.path.join(root, fn)

    if args.category == "others":
        # Unrelated negatives from neg_cc3m_5k.json — no base-mention split, one panel.
        neg = json.load(open(args.others_jsonl))
        raw = neg.get(args.others_split, [])
        ids = [s for s in (os.path.splitext(os.path.basename(x))[0] for x in raw) if s in stem_to_file]
        groups = {"unrelated": ids}
    else:
        data = json.load(open(args.samples_json))
        cat_items = [it for it in data if it.get("category") == args.category]
        if args.population == "all":
            # ONE combined set for the category — no base-mention split.
            ids = [s for it in cat_items
                   if (s := os.path.splitext(os.path.basename(it[args.id_col]))[0]) in stem_to_file]
            groups = {args.category: ids}
        else:
            # Population labels depend on whether the category actually contains a toilet:
            #   toilet ABSENT  -> base mention = hallucination
            #   toilet PRESENT -> base mention = correct; missing it = a miss
            has_toilet = args.category != "bathroom_only"
            lbl = ({True: "toilet-said (correct)", False: "toilet-missed"} if has_toilet
                   else {True: "hallucinating", False: "non-hallucinating"})
            groups = {lbl[True]: [], lbl[False]: []}
            for it in cat_items:
                flag = it["prompt_results"].get(args.base_prompt, {}).get("base_mentions_object")
                if flag not in (True, False):
                    continue
                stem = os.path.splitext(os.path.basename(it[args.id_col]))[0]
                if stem in stem_to_file:
                    groups[lbl[flag]].append(stem)
            sel = {"both": [True, False], "said": [True], "halluc": [True],
                   "unsaid": [False], "non": [False]}[args.population]
            groups = {lbl[f]: groups[lbl[f]] for f in sel}
    for g in groups:
        if args.num_images and args.num_images > 0:
            random.Random(args.seed).shuffle(groups[g])
            groups[g] = groups[g][:args.num_images]
    print(f"[compare] {args.category} images: " + ", ".join(f"{g}={len(ids)}" for g, ids in groups.items()))

    toilet = sequence_layer_probes_from_checkpoint(args.toilet_probe, device)
    toilet.eval()
    processor = LlavaProcessor.from_pretrained(args.model_name)

    # caption_by_stem fixed by the base pass; per_img[method][stem] = σ_toilet per layer.
    caption_by_stem = {}
    img_cache = {}
    per_img = {m: {} for m in methods}

    def load_img(stem):
        if stem not in img_cache:
            try:
                img_cache[stem] = Image.open(stem_to_file[stem]).convert("RGB")
            except Exception:
                img_cache[stem] = None
        return img_cache[stem]

    for m in methods:
        model = build_model(m, args, dtype, device)
        for g, ids in groups.items():
            for i, stem in enumerate(ids):
                img = load_img(stem)
                if img is None:
                    continue
                if m == "base":
                    asst = args.forced_text or generate_caption(
                        model, processor, img, args.question, device, args.max_new_tokens)
                    caption_by_stem[stem] = asst
                else:
                    asst = caption_by_stem.get(stem)
                    if asst is None:        # base failed on this image → skip everywhere
                        continue
                scores = toilet_scores_per_layer(model, processor, toilet, img, asst,
                                                 args.question, device, hps, layers, args.max_seq_tokens)
                if scores is not None:
                    per_img[m][stem] = scores
            print(f"[compare] {m}: scored {len([s for s in ids if s in per_img[m]])}/{len(ids)} ({g})")
        free_model(model)

    # ── Build (methods × layers) Δ matrices per group: method − base, paired per image ──
    mats = {}
    for g, ids in groups.items():
        if not ids:
            mats[g] = None
            continue
        M = np.full((len(methods), n_layers), np.nan)
        for mi, m in enumerate(methods):
            paired = [per_img[m][s] - per_img["base"][s]
                      for s in ids if s in per_img[m] and s in per_img["base"]]
            if paired:
                M[mi] = np.mean(paired, axis=0)
        mats[g] = M

    # Base model's ABSOLUTE toilet σ per layer (so the score can be compared across prompts).
    base_abs = {}
    for g, ids in groups.items():
        arr = [per_img["base"][s] for s in ids if s in per_img["base"]]
        base_abs[g] = (np.mean(arr, axis=0).tolist() if arr else None)

    # ── Save JSON ────────────────────────────────────────────────────────────────
    os.makedirs(os.path.dirname(args.out_json), exist_ok=True)
    with open(args.out_json, "w") as f:
        json.dump({"methods": methods, "n_layers": n_layers, "hook_type": args.hook_type,
                   "context": args.forced_text if args.forced_text else "base_caption",
                   "n_images": {g: len(ids) for g, ids in groups.items()},
                   "base_abs_toilet": base_abs,
                   "delta": {g: (None if mats[g] is None else
                                 np.where(np.isnan(mats[g]), None, mats[g]).tolist())
                             for g in groups}}, f)
    print(f"[compare] saved {args.out_json}")

    # ── Heatmap ──────────────────────────────────────────────────────────────────
    ctx = "forced" if args.forced_text else "base caption"
    plot_method_layer(
        mats, methods, layers, args.out,
        title=f"toilet readout: method − base, per layer — {args.category} ({ctx} context, {args.hook_type})")


if __name__ == "__main__":
    main()