File size: 13,104 Bytes
6febba8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
"""Level-of-Token DiT adapter.

The backbone still runs attention. This module replaces only the visual
token pack, the positional centers, and the velocity head:

    tokens (B, T, H, W, D) in scaled y-space
        -> per-extent input head + shape MLP          (eqs. 11–12)
        -> backbone(hidden, centers, sigma) -> hidden
        -> per-extent output head                      (eq. 10)
        -> LakonLab asymmetric-velocity recovery       (eq. 9)
        -> dense velocity on the uniform lattice

Text, audio, and the transformer's own blocks stay inside ``backbone``.
It must return one hidden vector per LoT token, in ``layout.rects`` order.
"""

from __future__ import annotations

import torch
from torch import nn

from flow import (
    apply_extent_scales,
    asymmetric_target,
    fit_procrustes,
    gather_extent,
    mean_basis,
    recover_dense_velocity,
    scatter_extent,
)
from layout import LotLayout, shape_features


def _key(extent: tuple[int, int, int]) -> str:
    et, eh, ew = extent
    return f"{et}x{eh}x{ew}"


class ExtentBank(nn.Module):
    """Fixed per-extent lifts ``A`` and data scales ``s``.

    Bases start as the mean lift and stay frozen. ``fit_`` replaces one
    extent with a Procrustes fit; training does not update these buffers.
    """

    def __init__(self, token_dim: int, extents: list[tuple[int, int, int]]):
        super().__init__()
        if not extents:
            raise ValueError("at least one extent is required")
        self.token_dim = int(token_dim)
        self.extents = [tuple(map(int, extent)) for extent in extents]
        seen = set()
        for extent in self.extents:
            if extent in seen:
                raise ValueError(f"duplicate extent {extent}")
            seen.add(extent)
            if min(extent) < 1:
                raise ValueError(f"bad extent {extent}")
            key = _key(extent)
            self.register_buffer(f"A_{key}", mean_basis(self.token_dim, extent))
            self.register_buffer(f"s_{key}", torch.ones(()))

    def basis(self, extent: tuple[int, int, int]) -> torch.Tensor:
        return getattr(self, f"A_{_key(extent)}")

    def scale(self, extent: tuple[int, int, int]) -> torch.Tensor:
        return getattr(self, f"s_{_key(extent)}")

    def scales(self) -> dict[tuple[int, int, int], torch.Tensor]:
        return {extent: self.scale(extent) for extent in self.extents}

    def dense_dim(self, extent: tuple[int, int, int]) -> int:
        et, eh, ew = extent
        return self.token_dim * et * eh * ew

    @torch.no_grad()
    def fit_(self, extent: tuple[int, int, int], dense: torch.Tensor, reference: torch.Tensor) -> None:
        """Replace ``A`` and ``s`` for one extent. Inputs are detached."""
        if tuple(extent) not in self.extents:
            raise KeyError(f"extent {extent} is not in this bank")
        if dense.shape[-1] != self.dense_dim(extent) or reference.shape[-1] != self.token_dim:
            raise ValueError(
                f"expected dense (N, {self.dense_dim(extent)}) and reference (N, {self.token_dim})"
            )
        basis, scale = fit_procrustes(dense, reference)
        slot = self.basis(extent)
        slot.copy_(basis.to(device=slot.device, dtype=slot.dtype))
        scale_slot = self.scale(extent)
        scale_slot.copy_(scale.to(device=scale_slot.device, dtype=scale_slot.dtype))

    def scale_clean(self, x0: torch.Tensor, layout: LotLayout) -> torch.Tensor:
        """Eq. 16. ``x0`` is ``(B, T, H, W, D)`` in the pretrained latent space."""
        self._check_layout(layout)
        return apply_extent_scales(x0, layout, self.scales(), invert=False)

    def unscale(self, y0: torch.Tensor, layout: LotLayout) -> torch.Tensor:
        """Map a y-space clean estimate back to the pretrained latent."""
        self._check_layout(layout)
        return apply_extent_scales(y0, layout, self.scales(), invert=True)

    def _check_layout(self, layout: LotLayout) -> None:
        missing = sorted({rect.extent for rect in layout.rects} - set(self.extents))
        if missing:
            raise KeyError(f"layout uses extents with no basis: {missing}")


class LotVisualAdapter(nn.Module):
    """Extent heads and shape MLP wrapped around an external backbone."""

    def __init__(
        self,
        token_dim: int,
        hidden_size: int,
        extents: list[tuple[int, int, int]],
        *,
        include_time: bool = False,
        shape_hidden: int = 256,
    ):
        super().__init__()
        self.token_dim = int(token_dim)
        self.hidden_size = int(hidden_size)
        self.include_time = bool(include_time)
        self.bank = ExtentBank(self.token_dim, extents)
        feat_dim = 5 if self.include_time else 4
        width = max(feat_dim, int(shape_hidden))
        self.shape_mlp = nn.Sequential(
            nn.Linear(feat_dim, width),
            nn.SiLU(),
            nn.Linear(width, self.hidden_size),
        )
        nn.init.zeros_(self.shape_mlp[-1].weight)
        nn.init.zeros_(self.shape_mlp[-1].bias)
        self.in_proj = nn.ModuleDict()
        self.out_proj = nn.ModuleDict()
        for extent in self.bank.extents:
            dense_dim = self.bank.dense_dim(extent)
            self.in_proj[_key(extent)] = nn.Linear(dense_dim, self.hidden_size)
            self.out_proj[_key(extent)] = nn.Linear(self.hidden_size, dense_dim)
        self.register_buffer("pretrained_in", torch.empty(self.hidden_size, self.token_dim))
        self.register_buffer("pretrained_out", torch.empty(self.token_dim, self.hidden_size))
        self.register_buffer("pretrained_bias_in", torch.empty(self.hidden_size))
        self.register_buffer("pretrained_bias_out", torch.empty(self.token_dim))
        self.register_buffer("pretrained_ready", torch.zeros((), dtype=torch.int32))

    @torch.no_grad()
    def init_from_pretrained(
        self,
        weight_in: torch.Tensor,
        weight_out: torch.Tensor,
        bias_in: torch.Tensor | None = None,
        bias_out: torch.Tensor | None = None,
    ) -> None:
        """Initialize heads as ``W_in,e = W_in A^T`` and ``W_e = A W_out``.

        ``weight_in`` is ``(hidden, D)`` and ``weight_out`` is ``(D, hidden)``,
        matching ``nn.Linear`` layout. Output bias is lifted by ``A`` so a
        unit extent reproduces the pretrained bias and a larger extent starts
        on the same subspace.
        """
        hidden, token_dim = self.hidden_size, self.token_dim
        if tuple(weight_in.shape) != (hidden, token_dim):
            raise ValueError(f"weight_in shape {tuple(weight_in.shape)} != {(hidden, token_dim)}")
        if tuple(weight_out.shape) != (token_dim, hidden):
            raise ValueError(f"weight_out shape {tuple(weight_out.shape)} != {(token_dim, hidden)}")
        self.pretrained_in.copy_(weight_in.detach().to(dtype=self.pretrained_in.dtype, device=self.pretrained_in.device).clone())
        self.pretrained_out.copy_(weight_out.detach().to(dtype=self.pretrained_out.dtype, device=self.pretrained_out.device).clone())
        if bias_in is None:
            self.pretrained_bias_in.zero_()
        else:
            self.pretrained_bias_in.copy_(bias_in.detach().to(dtype=self.pretrained_bias_in.dtype, device=self.pretrained_bias_in.device).clone())
        if bias_out is None:
            self.pretrained_bias_out.zero_()
        else:
            self.pretrained_bias_out.copy_(bias_out.detach().to(dtype=self.pretrained_bias_out.dtype, device=self.pretrained_bias_out.device).clone())
        self.pretrained_ready.fill_(1)
        for extent in self.bank.extents:
            key = _key(extent)
            in_layer = self.in_proj[key]
            out_layer = self.out_proj[key]
            basis = self.bank.basis(extent).to(dtype=in_layer.weight.dtype, device=in_layer.weight.device)
            weight_in_m = weight_in.to(dtype=in_layer.weight.dtype, device=in_layer.weight.device)
            weight_out_m = weight_out.to(dtype=out_layer.weight.dtype, device=out_layer.weight.device)
            in_layer.weight.copy_(weight_in_m @ basis.T)
            out_layer.weight.copy_(basis @ weight_out_m)
            if bias_in is None:
                in_layer.bias.zero_()
            else:
                in_layer.bias.copy_(bias_in.to(dtype=in_layer.bias.dtype, device=in_layer.bias.device))
            if bias_out is None:
                out_layer.bias.zero_()
            else:
                lifted = basis @ bias_out.to(dtype=basis.dtype, device=basis.device)
                out_layer.bias.copy_(lifted.to(dtype=out_layer.bias.dtype))

    @torch.no_grad()
    def fit_extent(self, extent: tuple[int, int, int], dense: torch.Tensor, reference: torch.Tensor) -> None:
        """Fit one basis, then rebuild every head from the stored pretrained maps.

        ``fit_`` on the bank alone updates ``A`` and leaves the heads on the
        previous subspace. Call this instead once ``init_from_pretrained`` has run.
        """
        self.bank.fit_(extent, dense, reference)
        if int(self.pretrained_ready) != 1:
            return
        self.init_from_pretrained(
            self.pretrained_in,
            self.pretrained_out,
            self.pretrained_bias_in,
            self.pretrained_bias_out,
        )

    def embed(self, tokens: torch.Tensor, layout: LotLayout) -> tuple[torch.Tensor, torch.Tensor]:
        """Return ``(B, L, hidden)`` and ``(L, 3)`` finest-grid centers."""
        self._check_tokens(tokens, layout)
        hidden = tokens.new_zeros(tokens.shape[0], layout.count, self.hidden_size)
        for extent, indices in layout.groups():
            rects = [layout.rects[index] for index in indices]
            dense = gather_extent(tokens, rects)
            projected = self.in_proj[_key(extent)](dense)
            features = shape_features(extent, self.include_time).to(device=tokens.device, dtype=tokens.dtype)
            projected = projected + self.shape_mlp(features)
            hidden[:, indices] = projected
        centers = layout.centers().to(device=tokens.device)
        return hidden, centers

    def velocity_from_states(
        self,
        states: torch.Tensor,
        tokens: torch.Tensor,
        sigma: torch.Tensor | float,
        layout: LotLayout,
        *,
        x0_minus_eps: bool = False,
    ) -> torch.Tensor:
        """Extent heads plus eq. 9, scattered onto the uniform lattice.

        Eq. 9 assumes the head predicts ``P eps - x0`` and returns ``eps - x0``.
        H3's head predicts the negation (``x0 - eps``; H3's scheduler steps ``x0 = x_t + sigma * v``).
        ``x0_minus_eps=True`` flips the head before recovery and the velocity
        after, so the result stays in the caller's convention. A unit extent has
        no complement, so 1x1 parity cannot see a wrong sign here.
        """
        sign = -1.0 if x0_minus_eps else 1.0
        self._check_tokens(tokens, layout)
        if states.shape != (tokens.shape[0], layout.count, self.hidden_size):
            raise ValueError(
                f"backbone returned {tuple(states.shape)}, expected "
                f"{(tokens.shape[0], layout.count, self.hidden_size)}"
            )
        velocity = tokens.new_zeros(tokens.shape)
        for extent, indices in layout.groups():
            rects = [layout.rects[index] for index in indices]
            u_a = sign * self.out_proj[_key(extent)](states[:, indices])
            x_t = gather_extent(tokens, rects)
            recovered = recover_dense_velocity(u_a, x_t, self.bank.basis(extent), sigma)
            scatter_extent(velocity, sign * recovered, rects)
        return velocity

    def forward(self, tokens, sigma, layout: LotLayout, backbone, backbone_kwargs=None):
        """Dense y-space velocity. ``backbone(hidden, centers, sigma, **kwargs)``."""
        hidden, centers = self.embed(tokens, layout)
        kwargs = {} if backbone_kwargs is None else backbone_kwargs
        states = backbone(hidden, centers, sigma, **kwargs)
        return self.velocity_from_states(states, tokens, sigma, layout)

    def asymmetric_targets(
        self,
        y0: torch.Tensor,
        noise: torch.Tensor,
        layout: LotLayout,
    ) -> list[tuple[list[int], torch.Tensor]]:
        """Per-extent eq. 8 targets, for tests and for a direct head loss."""
        self.bank._check_layout(layout)
        targets = []
        for extent, indices in layout.groups():
            rects = [layout.rects[index] for index in indices]
            clean = gather_extent(y0, rects)
            eps = gather_extent(noise, rects)
            targets.append((indices, asymmetric_target(clean, eps, self.bank.basis(extent))))
        return targets

    def _check_tokens(self, tokens: torch.Tensor, layout: LotLayout) -> None:
        expected = (layout.time, layout.height, layout.width, self.token_dim)
        if tokens.ndim != 5 or tuple(tokens.shape[1:]) != expected:
            raise ValueError(f"tokens shape {tuple(tokens.shape)} != (B, {expected[0]}, {expected[1]}, {expected[2]}, {expected[3]})")
        self.bank._check_layout(layout)