olivv-cs-ComfyUI-FunPack/modules/sampling/block_influence/measure.py

"""The probe's arithmetic and its stored profile: no model, no routes."""

import contextlib

import torch

from ..._core import config, dit_hooks, registry

KIND = "block_influence"
MIN_PER_GROUP = 2           # liked AND disliked before a difference is one
SWITCH = config.ROOT / "block_influence.enabled"   # on disk: the choice survives a restart


# Why recording cannot or did not store anything, for the panel (shown only while recording is on).
problem = None


def enabled() -> bool:
    # Off until someone turns it on: with no switch file yet nothing is recorded.
    try:
        return SWITCH.read_text().strip() == "1"
    except OSError:
        return False


def set_enabled(on: bool) -> bool:
    SWITCH.parent.mkdir(parents=True, exist_ok=True)
    SWITCH.write_text("1" if on else "0")
    return bool(on)


def _aside():
    """The measurement's own tensors stay out of ComfyUI's memory recorder (aimdo), which assumes every step
    allocates the same: the running sums grow differently from step to step and broke it ("aimdo memory compile
    error") whenever no other module had switched the recorder off. The block itself still runs inside it."""
    try:
        from comfy.model_prefetch import pause_malloc_graph
    except ImportError:                          # a ComfyUI without the recorder has nothing to pause
        return contextlib.nullcontext()
    return pause_malloc_graph()


class Tally:
    """One run's per-block sums, kept on the device until the run ends.

    No `.item()` per block per step (50 blocks x steps host syncs): sums are 0-dim
    tensors and are read once, after sampling. Rows are strided down to
    ~`max_rows` before the norms so the cost does not grow with H3's sequence length.
    """

    def __init__(self, n, max_rows=512):
        self.n, self.max_rows = n, max_rows
        self._sum = {k: [None] * n for k in ("raw", "ratio", "novelty")}
        self._count = {k: [0] * n for k in ("raw", "ratio", "novelty")}
        self._prev = (None, -1)         # previous block's push, for the novelty cosine
        self._spans = {}

    def _span(self, args, src):
        seq = int(src.shape[0])
        if seq not in self._spans:
            self._spans[seq] = dit_hooks.row_span(
                dit_hooks.target_rows(args.get("mod_segments"), seq, src.device, "video"))
        return self._spans[seq]

    def _add(self, kind, block, value):
        s = self._sum[kind][block]
        self._sum[kind][block] = value if s is None else s + value
        self._count[kind][block] += 1

    def measure(self, block, args, original):
        """Run the block and record how far it moved the picture rows."""
        src = args.get("img")
        sel = before = None
        with _aside():
            if torch.is_tensor(src) and src.dim() == 2:
                span = self._span(args, src)
                if span is not None:
                    sel = slice(span[0], span[1], max(1, (span[1] - span[0]) // self.max_rows))
                    # A COPY, read before the block runs: a block that writes its residual in
                    # place leaves `src` already holding the result, and every delta is zero.
                    before = src[sel].detach().to(torch.float32, copy=True)
        out = original(args)
        with _aside():
            if before is not None and torch.is_tensor(out.get("img")) \
                    and out["img"].shape[0] == src.shape[0]:
                delta = out["img"][sel].detach().float() - before
                push = delta.norm()
                self._add("raw", block, push)
                self._add("ratio", block, push / before.norm().clamp(min=1e-8))
                last, last_block = self._prev
                # A block index that does not exceed the last one means a new step began:
                # pairing across it would compare the end of one step to the start of the next.
                if last is not None and block > last_block and last.shape == delta.shape:
                    self._add("novelty", block, torch.nn.functional.cosine_similarity(
                        delta.flatten(), last.flatten(), dim=0))
                self._prev = (delta, block)
        return out

    def blocks_seen(self) -> int:
        return sum(1 for c in self._count["raw"] if c)

    def rows(self):
        """{name: [n_blocks] tensor} (NaN where a block saw nothing), or None when
        nothing was measured -- which must be said, not stored as a flat profile."""
        if not self.blocks_seen():
            return None
        rows = {}
        for name in ("ratio", "raw", "novelty"):
            have = [i for i, c in enumerate(self._count[name]) if c]
            vec = torch.full((self.n,), float("nan"))
            if have:
                means = torch.stack([self._sum[name][i].float() / self._count[name][i] for i in have])
                vec[have] = means.cpu()                      # the one host sync
            rows[name] = vec
        return rows


def profile(rows) -> dict:
    """Depth profile from a key's rated rows.

    `overall` is where the model does its work regardless of ratings; `share` is
    each block's RAW push as a fraction of all blocks' (rank by this: the ratio
    shrinks with depth by construction, since the stream grows); `difference` is
    liked minus disliked; `flatness` is the spread of RAW push as a fraction of
    its mean -- near 0 means every block moves the stream equally and there is
    nothing to aim at. `novelty` ~1: a block amplifies its predecessor, ~0 it
    adds something new, negative it partly undoes it.
    """
    rows = [r for r in rows if {"ratio", "raw", "novelty"} <= set(r["rows"])]
    empty = {"used": 0, "overall": {}, "difference": None, "share": {}, "novelty": {}, "mean_novelty": None,
             "flatness": None, "n_liked": 0, "n_disliked": 0}
    if not rows:
        return empty
    n = len(rows[-1]["rows"]["ratio"])
    rows = [r for r in rows if len(r["rows"]["ratio"]) == n]     # a different model's rows
    reward = torch.tensor([float(r["reward"]) for r in rows])
    stack = {k: torch.stack([r["rows"][k].float() for r in rows]) for k in ("ratio", "raw", "novelty")}

    def mean(name, mask=None):
        m = stack[name] if mask is None else stack[name][mask]
        return torch.nanmean(m, dim=0) if len(m) else torch.full((n,), float("nan"))

    def table(vec):
        return {str(i): float(v) for i, v in enumerate(vec) if torch.isfinite(v)}

    overall, raw, novelty = mean("ratio"), mean("raw"), mean("novelty")
    seen = torch.isfinite(raw)
    share = table(raw / raw[seen].sum()) if seen.any() and raw[seen].sum() > 1e-12 else {}
    liked, disliked = reward > 0, reward < 0
    difference = None
    if int(liked.sum()) >= MIN_PER_GROUP and int(disliked.sum()) >= MIN_PER_GROUP:
        difference = table(mean("ratio", liked) - mean("ratio", disliked))
    # Flatness is read on RAW push: the ratio falls with depth by construction (the stream grows), so
    # a flat model would read as uneven on it.
    vals = raw[torch.isfinite(raw)]
    flat = float(vals.std(unbiased=False) / vals.mean()) if len(vals) and abs(float(vals.mean())) > 1e-12 else None
    nov = novelty[torch.isfinite(novelty)]
    return {"used": len(rows), "overall": table(overall), "difference": difference, "share": share,
            "novelty": table(novelty), "mean_novelty": float(nov.mean()) if len(nov) else None,
            "flatness": flat, "n_liked": int(liked.sum()), "n_disliked": int(disliked.sum())}


GROUPS = ("liked", "image", "composition", "both")


def _group(row):
    """liked; or a dislike sorted by its axis: image, composition, or both (plain dislike)."""
    reward = float(row["reward"])
    if reward > 0:
        return "liked"
    if reward == 0:
        return None
    return row.get("axis") if row.get("axis") in ("image", "composition") else "both"


def groups(rows, shuffles=200, seed=0) -> dict:
    """Do the four groups point in different directions over the block profile?

    For each dislike group, `d = mean(group) - mean(liked)`; the answer is the cosine between
    those directions. The control reshuffles the group labels over the same clips (same group
    sizes) and asks how often that cosine comes out as high. A real difference should sit
    above the shuffled range; inside it, the ratings are not telling the groups apart yet.
    """
    rows = [r for r in rows if "ratio" in r["rows"]]
    if not rows:
        return {"used": 0, "counts": {}, "cos": {}, "note": "nothing recorded yet"}
    n = len(rows[-1]["rows"]["ratio"])
    rows = [r for r in rows if len(r["rows"]["ratio"]) == n]
    stack = torch.stack([r["rows"]["ratio"].float() for r in rows])
    # A block some clip never measured is left out of every group, not filled in with a guess.
    stack = stack[:, torch.isfinite(stack).all(dim=0)]
    if stack.shape[1] == 0:
        return {"used": len(rows), "counts": {g: sum(1 for r in rows if _group(r) == g) for g in GROUPS},
                "cos": {}, "note": "no block was measured on every clip"}
    labels = [_group(r) for r in rows]
    counts = {g: labels.count(g) for g in GROUPS}

    def directions(labs):
        means = {}
        for g in GROUPS:
            idx = [i for i, x in enumerate(labs) if x == g]
            if len(idx) >= MIN_PER_GROUP:
                means[g] = stack[idx].mean(dim=0)
        if "liked" not in means:
            return {}
        return {g: means[g] - means["liked"] for g in ("image", "composition", "both") if g in means}

    def cosines(d):
        pairs = {}
        names = sorted(d)
        for i, a in enumerate(names):
            for b in names[i + 1:]:
                pairs[f"{a} vs {b}"] = float(torch.nn.functional.cosine_similarity(d[a], d[b], dim=0))
        return pairs

    observed = cosines(directions(labels))
    if not observed:
        return {"used": len(rows), "counts": counts, "cos": {}, "note": "not enough clips in each group yet"}
    gen = torch.Generator().manual_seed(seed)
    null = {k: [] for k in observed}
    for _ in range(shuffles):
        perm = torch.randperm(len(labels), generator=gen).tolist()
        for k, v in cosines(directions([labels[i] for i in perm])).items():
            null[k].append(v)
    chance = {k: {"mean": sum(v) / len(v), "p95": sorted(v)[int(0.95 * (len(v) - 1))]} for k, v in null.items() if v}
    return {"used": len(rows), "counts": counts, "cos": observed, "chance": chance}


def _kind(key):
    kind = registry.current().ask("taste_kind", key, KIND)
    if kind is None:
        raise ValueError("the Taste key module is off, so there is nowhere to read profiles from")
    return kind


def resolve(key):
    """The key to show. The panel asks for a placeholder ("default") when it does not know
    which Taste key the runs write to, so a key with nothing recorded falls back to the key
    the latest capture went to -- and the answer says which one it is."""
    if _kind(key).path().exists():
        return key
    latest = registry.current().ask("taste_latest_key")
    return latest if latest and _kind(latest).path().exists() else key


def state(key, fallback=True) -> dict:
    """What Settings > Refinement & Taste shows for one key. `fallback` False answers for
    exactly that key (after a Clear, which must not show another key's data)."""
    key = resolve(key) if fallback else key
    rows = _kind(key).rows()
    prof = profile(rows)
    shown = problem if enabled() else None
    return {"key": key, "enabled": enabled(), "runs": len(rows), "problem": shown,
            "skipped": len(rows) - prof["used"], "min_per_group": MIN_PER_GROUP,
            **prof}


def four_groups(key, fallback=True) -> dict:
    """The four-group report for one key, for the panel (same key rules as state())."""
    key = resolve(key) if fallback else key
    return {"key": key, "enabled": enabled(), **groups(_kind(key).rows())}


def clear(key):
    _kind(key).clear()


def path_of(key):
    return _kind(key).path()