File size: 12,666 Bytes
0a2c21d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
import os
from typing import Any

import torch
import torch.nn.functional as F


def compute_token_change_rate(token_t: torch.Tensor, token_t_minus_1: torch.Tensor, eps: float = 1e-8) -> torch.Tensor:
    """Per-token relative change using L2 norms: (||token_t|| - ||token_{t-1}||) / ||token_t||."""
    norm_t = torch.linalg.vector_norm(token_t.float(), dim=-1)
    norm_t_minus_1 = torch.linalg.vector_norm(token_t_minus_1.float(), dim=-1)
    return (norm_t - norm_t_minus_1) / norm_t.clamp(min=eps)


def cosine_match_tokens(history_tokens: torch.Tensor, noise_tokens: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """For each noise token, find the history token with highest cosine similarity."""
    history = F.normalize(history_tokens.float(), dim=-1)
    noise = F.normalize(noise_tokens.float(), dim=-1)
    similarity = noise @ history.T
    match_indices = similarity.argmax(dim=-1)
    match_scores = similarity.gather(1, match_indices.unsqueeze(1)).squeeze(1)
    return match_indices, match_scores


def extract_frame_tokens(
    token_sequence: torch.Tensor,
    history_context_length: int,
    original_context_length: int,
    grid_tokens: int,
    history_frame_index: int,
    noise_frame_index: int,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Extract one spatial frame from short-history and current-noise token sequences."""
    short_history_frames = 2
    short_len = short_history_frames * grid_tokens
    history_seq_len = history_context_length

    short_start = history_seq_len - short_len
    history_start = short_start + history_frame_index * grid_tokens
    noise_start = history_seq_len + noise_frame_index * grid_tokens

    history_tokens = token_sequence[history_start : history_start + grid_tokens]
    noise_tokens = token_sequence[noise_start : noise_start + grid_tokens]
    return history_tokens, noise_tokens


def extract_noise_frame_tokens(
    token_sequence: torch.Tensor,
    history_context_length: int,
    grid_tokens: int,
    noise_frame_index: int,
) -> torch.Tensor:
    noise_start = history_context_length + noise_frame_index * grid_tokens
    return token_sequence[noise_start : noise_start + grid_tokens]


def resolve_short_history_frame_index(history_frame: int, short_history_frames: int = 2) -> int:
    if history_frame < 0:
        history_frame += short_history_frames
    return history_frame


def resolve_history_source_frame(
    chunk_index: int,
    history_frame: int,
    *,
    keep_first_frame: bool,
    num_latent_frames_per_chunk: int,
    short_history_frames: int = 2,
) -> tuple[int, int]:
    """Map configured short-history frame to the chunk/latent-frame where it was denoised."""
    short_index = resolve_short_history_frame_index(history_frame, short_history_frames)
    if keep_first_frame:
        if short_index == 0:
            return 0, 0
        return max(0, chunk_index - 1), num_latent_frames_per_chunk - 1
    return max(0, chunk_index - 1), num_latent_frames_per_chunk - 1


def token_yx_from_index(index: int, grid_w: int) -> tuple[int, int]:
    return int(index) // grid_w, int(index) % grid_w


def token_rect_on_image(
    token_y: int,
    token_x: int,
    image_height: int,
    image_width: int,
    grid_h: int,
    grid_w: int,
    patch_h: int = 2,
    patch_w: int = 2,
) -> tuple[float, float, float, float]:
    """Map token grid cell to pixel rectangle on decoded frame."""
    latent_h = grid_h * patch_h
    latent_w = grid_w * patch_w
    ly0 = token_y * patch_h
    lx0 = token_x * patch_w
    ly1 = ly0 + patch_h
    lx1 = lx0 + patch_w
    py0 = ly0 / latent_h * image_height
    px0 = lx0 / latent_w * image_width
    py1 = ly1 / latent_h * image_height
    px1 = lx1 / latent_w * image_width
    return px0, py0, px1 - px0, py1 - py0


class TokenDynamicsDebugTracker:
    def __init__(self, config: dict[str, Any] | None):
        self.config = config
        self.state: dict[str, Any] = {}
        self._frame_rates: dict[tuple[int, int], dict[str, list]] = {}
        self.reset_chunk_records()

    @property
    def enabled(self) -> bool:
        return self.config is not None and self.config.get("enabled", True)

    def reset_chunk_records(self):
        self.match_indices = None
        self.match_scores = None
        self.step_indices: list[int] = []
        self.timesteps: list[float] = []
        self._prev_noise_frame_tokens: dict[int, torch.Tensor] = {}
        self._matching_done = False
        self.noise_latent_frame = None
        self.history_latent_frame = None
        self.capture_pass_name = None

    def set_state(self, **state):
        prev_chunk = self.state.get("chunk_index")
        next_chunk = state.get("chunk_index", prev_chunk)
        if prev_chunk is not None and next_chunk is not None and prev_chunk != next_chunk:
            self.reset_chunk_records()
        self.state.update(state)

    def should_record(self) -> bool:
        if not self.enabled:
            return False
        pass_names = self.config.get("pass_names", ["cond"])
        if self.state.get("pass_name", "cond") not in pass_names:
            return False
        return True

    def should_capture(self) -> bool:
        if not self.should_record():
            return False
        return self.is_analysis_chunk()

    def is_analysis_chunk(self) -> bool:
        chunks = self.config.get("chunks")
        if chunks is not None and self.state.get("chunk_index", 0) not in chunks:
            return False
        return True

    def get_recording_targets(self) -> set[tuple[int, int]]:
        """(chunk, latent_frame) pairs to record: history source frame + noise frame only."""
        targets: set[tuple[int, int]] = set()
        analysis_chunks = self.config.get("chunks")
        if not analysis_chunks:
            return targets

        noise_frame = int(self.config.get("noise_frame", 0))
        for chunk_index in analysis_chunks:
            targets.add((int(chunk_index), noise_frame))
            hist_chunk, hist_frame = resolve_history_source_frame(
                int(chunk_index),
                int(self.config.get("history_frame", -1)),
                keep_first_frame=bool(self.config.get("keep_first_frame", True)),
                num_latent_frames_per_chunk=int(self.config.get("num_latent_frames_per_chunk", 9)),
            )
            targets.add((hist_chunk, hist_frame))
        return targets

    def _record_frame_rates(self, chunk_index: int, frame_index: int, rate: torch.Tensor):
        key = (chunk_index, frame_index)
        bucket = self._frame_rates.setdefault(key, {"rates": [], "step_indices": [], "timesteps": []})
        bucket["rates"].append(rate.detach().cpu())
        bucket["step_indices"].append(int(self.state.get("step_index", 0)))
        bucket["timesteps"].append(float(self.state.get("timestep", self.state.get("step_index", 0))))

    def observe(
        self,
        latent_tokens: torch.Tensor,
        history_context_length: int,
        original_context_length: int,
    ):
        if not self.should_record():
            return

        if self.capture_pass_name is None:
            self.capture_pass_name = self.state.get("pass_name", "cond")

        grid_h, grid_w = self.config.get("grid", (24, 40))
        grid_tokens = int(grid_h) * int(grid_w)
        history_frame = int(self.config.get("history_frame", -1))
        noise_frame = int(self.config.get("noise_frame", 0))
        short_history_frame = resolve_short_history_frame_index(history_frame)

        chunk_index = int(self.state.get("chunk_index", 0))
        recording_targets = self.get_recording_targets()
        frames_to_record = sorted(frame_index for c, frame_index in recording_targets if c == chunk_index)

        for frame_index in frames_to_record:
            frame_tokens = extract_noise_frame_tokens(
                latent_tokens[0],
                history_context_length,
                grid_tokens,
                frame_index,
            )
            prev_tokens = self._prev_noise_frame_tokens.get(frame_index)
            if prev_tokens is not None:
                self._record_frame_rates(
                    chunk_index,
                    frame_index,
                    compute_token_change_rate(frame_tokens, prev_tokens),
                )
            self._prev_noise_frame_tokens[frame_index] = frame_tokens.detach()

        step_index = int(self.state.get("step_index", 0))
        total_steps = int(self.state.get("total_steps", 1))
        is_last_step = step_index == total_steps - 1

        if self.should_capture() and is_last_step and not self._matching_done:
            latent_history, latent_noise = extract_frame_tokens(
                latent_tokens[0],
                history_context_length,
                original_context_length,
                grid_tokens,
                short_history_frame,
                noise_frame,
            )
            self.match_indices, self.match_scores = cosine_match_tokens(latent_history, latent_noise)
            self._matching_done = True

    def _get_frame_rate_series(self, chunk_index: int, frame_index: int) -> dict[str, Any]:
        key = (chunk_index, frame_index)
        if key not in self._frame_rates or not self._frame_rates[key]["rates"]:
            raise KeyError(f"No patch-latent change rates recorded for chunk={chunk_index}, frame={frame_index}")
        bucket = self._frame_rates[key]
        return {
            "rates": torch.stack(bucket["rates"], dim=0),
            "step_indices": bucket["step_indices"].copy(),
            "timesteps": bucket["timesteps"].copy(),
        }

    def set_visualization_latent_frames(
        self,
        denoised_latents: torch.Tensor,
        history_short_latents: torch.Tensor,
    ):
        """Store fully denoised latent frames for VAE visualization (after chunk sampling)."""
        if not self.enabled or not self._matching_done or not self.is_analysis_chunk():
            return

        history_frame = int(self.config.get("history_frame", -1))
        noise_frame = int(self.config.get("noise_frame", 0))
        short_frames = history_short_latents.shape[2]
        short_index = resolve_short_history_frame_index(history_frame, short_frames)

        self.history_latent_frame = history_short_latents[0, :, short_index].detach().float().cpu()
        self.noise_latent_frame = denoised_latents[0, :, noise_frame].detach().float().cpu()

    def save(self):
        if not self.enabled or self.match_indices is None:
            return None

        chunk_index = int(self.state.get("chunk_index", 0))
        noise_frame = int(self.config.get("noise_frame", 0))
        history_source_chunk, history_source_frame = resolve_history_source_frame(
            chunk_index,
            int(self.config.get("history_frame", -1)),
            keep_first_frame=bool(self.config.get("keep_first_frame", True)),
            num_latent_frames_per_chunk=int(self.config.get("num_latent_frames_per_chunk", 9)),
        )

        history_series = self._get_frame_rate_series(history_source_chunk, history_source_frame)
        noise_series = self._get_frame_rate_series(chunk_index, noise_frame)
        self.step_indices = noise_series["step_indices"]
        self.timesteps = noise_series["timesteps"]

        output_dir = self.config.get("output_dir", "token_dynamics_debug")
        os.makedirs(output_dir, exist_ok=True)

        filename = (
            f"token_dynamics_chunk{chunk_index}_"
            f"hist{self.config.get('history_frame', -1)}_"
            f"noise{self.config.get('noise_frame', 0)}_"
            f"{self.capture_pass_name or self.state.get('pass_name', 'cond')}.pt"
        )
        path = os.path.join(output_dir, filename)

        artifact = {
            "chunk_index": chunk_index,
            "history_frame": int(self.config.get("history_frame", -1)),
            "noise_frame": noise_frame,
            "history_source_chunk": history_source_chunk,
            "history_source_frame": history_source_frame,
            "grid": tuple(self.config.get("grid", (24, 40))),
            "match_indices": self.match_indices.cpu(),
            "match_scores": self.match_scores.cpu(),
            "history_change_rates": history_series["rates"],
            "noise_change_rates": noise_series["rates"],
            "step_indices": self.step_indices.copy(),
            "timesteps": self.timesteps.copy(),
            "total_steps": int(self.state.get("total_steps", 0)),
            "noise_latent_frame": self.noise_latent_frame,
            "history_latent_frame": self.history_latent_frame,
        }
        torch.save(artifact, path)
        return path