File size: 25,909 Bytes
23a59ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
# wm_dataset.py
import os
import glob
import json
import bisect
import random as _random
from collections import OrderedDict
from typing import Dict, Mapping, Optional, Sequence, Union

import torch
from torch.utils.data import Dataset


class WMDataset(Dataset):
    """
    RGB-native world model dataset.

    data_dir and frames_dir may each be a str or list[str]. Lists are treated as
    paired roots; a length-1 root is broadcast to match a longer one.
    """
    def __init__(
        self,
        data_dir: Union[str, Sequence[str]],
        frames_dir: Union[str, Sequence[str]],
        seq_len: int,
        img_size: int = 128,
        action_dim: int = 16,
        lang_dim: int = 512,
        cache_mb: int = 2048,
        verbose: bool = True,
        tasks_json: str = "../tasks.json",
        tasks: Optional[list[str]] = None,
        strict_tasks: bool = True,
        ddp_partition: bool = False,
        iid_sampling: bool = False,
        samples_per_shard: int = 1,
        task_weights: Optional[Mapping[str, float]] = None,
    ):
        super().__init__()

        # normalize data_dir / frames_dir to lists, and pair them
        if isinstance(data_dir, (str, os.PathLike)):
            data_dirs = [str(data_dir)]
        else:
            data_dirs = [str(x) for x in data_dir]

        if isinstance(frames_dir, (str, os.PathLike)):
            frames_dirs = [str(frames_dir)]
        else:
            frames_dirs = [str(x) for x in frames_dir]

        if len(data_dirs) != len(frames_dirs):
            if len(data_dirs) == 1:
                data_dirs = data_dirs * len(frames_dirs)
            elif len(frames_dirs) == 1:
                frames_dirs = frames_dirs * len(data_dirs)
            else:
                raise ValueError(f"data_dir and frames_dir must have same length (or one must be length-1). "
                                 f"Got {len(data_dirs)} and {len(frames_dirs)}")

        self.data_dirs = data_dirs
        self.frames_dirs = frames_dirs
        self.sources = list(zip(self.data_dirs, self.frames_dirs))

        self.T = int(seq_len)
        self.H = int(img_size)
        self.W = int(img_size)
        self.A = int(action_dim)
        self.lang_dim = int(lang_dim)
        self.cache_bytes = int(cache_mb) * 1024 * 1024
        self.verbose = bool(verbose)
        self.tasks_filter = None if tasks is None else set(tasks)
        self.strict_tasks = bool(strict_tasks)

        # --- Task metadata (action_dim + text_embedding) ---
        self.task_meta: Optional[dict] = None
        if tasks_json and os.path.exists(tasks_json):
            try:
                with open(tasks_json, "r") as f:
                    self.task_meta = json.load(f)
            except Exception as e:
                if self.verbose:
                    print(f"[WMDataset] Warning: failed to load tasks_json={tasks_json}: {e}")
                self.task_meta = None
        elif tasks_json and self.verbose:
            print(f"[WMDataset] Warning: tasks_json not found at {tasks_json} (continuing with zeros lang_emb + default action masks).")

        self._zero_lang = torch.zeros(self.lang_dim, dtype=torch.float32)

        # LRU cache for shards: key=(task_idx, seg_idx, shard_idx) -> frames tensor
        self._cache = OrderedDict()
        self._cache_nbytes = 0

        # Discover tasks from ALL data_dirs/*.pt (dedup, preserve first-seen order)
        found_tasks = []
        seen = set()
        for dd in self.data_dirs:
            demo_paths = sorted(glob.glob(os.path.join(dd, "*.pt")))
            for p in demo_paths:
                t = os.path.splitext(os.path.basename(p))[0]
                if t not in seen:
                    seen.add(t)
                    found_tasks.append(t)

        if self.tasks_filter is not None:
            requested = [t for t in tasks if t in self.tasks_filter] if tasks is not None else []
            if len(requested) == 0:
                requested = [t for t in found_tasks if t in self.tasks_filter]
            tasks = requested

            if self.verbose:
                missing = [t for t in self.tasks_filter if t not in set(found_tasks)]
                print(f"[WMDataset] Task filter: keeping {len(tasks)}/{len(found_tasks)} tasks")
                if missing:
                    msg = f"[WMDataset] WARNING: {len(missing)} requested tasks not found in data_dir(s) (e.g. {missing[:5]})"
                    if self.strict_tasks:
                        raise FileNotFoundError(msg)
                    else:
                        print(msg)
        else:
            tasks = found_tasks

        # DDP-aware strided task partitioning: each rank handles every N-th task
        # so shard pools are disjoint and per-worker caches stay effective.
        if ddp_partition:
            rank = int(os.environ.get("RANK", "0"))
            world_size = int(os.environ.get("WORLD_SIZE", "1"))
            if world_size > 1 and len(tasks) >= world_size:
                tasks = [tasks[i] for i in range(rank, len(tasks), world_size)]
                if self.verbose:
                    print(f"[WMDataset] DDP partition: rank {rank}/{world_size} "
                          f"→ {len(tasks)} tasks (strided)")

        self.iid_sampling = bool(iid_sampling)
        self.samples_per_shard = int(samples_per_shard)
        # Per-worker state for iid_sampling locality (initialized lazily)
        self._cur_shard_entry: Optional[int] = None
        self._cur_shard_draws: int = 0

        # Stored per-task
        self.tasks = []
        self.demo_paths = []     # stores list of per-task demo paths (joined)
        self.shard_lists = []    # per task -> list[segments], each segment is list[shard_paths]
        self.shard_cum = []      # per task -> list[segments], each segment is list[int] cumulative frame counts per shard
        self.seg_cum_frames = [] # per task -> cumulative frame counts across segments (for indexing)

        self.ep = []
        self.act = []
        self.rew = []
        self.valid_starts = []
        self._cum_counts = []

        # Precomputed per-task metadata used by __getitem__
        self._emb_ids = []
        self._act_dims = []
        self._act_mask_1d = []
        self._lang_embs = []

        total = 0
        for task in tasks:
            # --- gather segments for this task from each (data_dir, frames_dir) source ---
            seg_eps = []
            seg_acts = []
            seg_rews = []
            seg_shards = []
            seg_shard_cums = []
            seg_num_frames = []
            seg_demo_paths = []

            ep_offset = 0  # ensures episode ids are unique across segments

            for (dd, fd) in self.sources:
                dp = os.path.join(dd, f"{task}.pt")
                shard_glob = os.path.join(fd, task, "*shard*.pt")
                shards = sorted(glob.glob(shard_glob))
                if not os.path.exists(dp) or len(shards) == 0:
                    continue

                try:
                    td = torch.load(dp, map_location="cpu", weights_only=False)
                except Exception as e:
                    if self.verbose:
                        print(f"[WMDataset] Skipping task={task} source=({dd},{fd}): torch.load demo failed: {e}")
                    continue

                try:
                    ep = td["episode"].to(torch.int64).cpu()
                    act = td["action"].cpu()
                    rew = td["reward"].cpu()
                except Exception as e:
                    if self.verbose:
                        print(f"[WMDataset] Skipping task={task} source=({dd},{fd}): missing keys in demo: {e}")
                    continue

                if rew.ndim == 2 and rew.shape[-1] == 1:
                    rew = rew.squeeze(-1)
                rew = rew.to(torch.float32)

                if act.ndim == 1:
                    act = act.unsqueeze(-1)
                act = act.to(torch.float32)

                N = int(rew.shape[0])
                if act.shape[0] != N or ep.shape[0] != N:
                    if self.verbose:
                        print(f"[WMDataset] Skipping task={task} source=({dd},{fd}): length mismatch ep/act/rew.")
                    continue

                # Determine per-shard frame counts from index file or last shard
                index_path = os.path.join(fd, task, f"{task}_index.json")
                shard_sizes = None
                if os.path.exists(index_path):
                    try:
                        with open(index_path, "r") as f:
                            index = json.load(f)
                        shard_sizes = [int(index[os.path.basename(s)]) for s in shards]
                    except Exception:
                        shard_sizes = None
                if shard_sizes is None:
                    # Fallback: load each shard to get its size
                    shard_sizes = []
                    fallback_ok = True
                    for s in shards:
                        try:
                            td_s = torch.load(s, map_location="cpu", weights_only=False)
                            shard_sizes.append(int(td_s["frames"].shape[0]))
                        except Exception as e:
                            if self.verbose:
                                print(f"[WMDataset] Skipping task={task} source=({dd},{fd}): failed to read shard {s}: {e}")
                            fallback_ok = False
                            break
                    if not fallback_ok:
                        continue

                N_frames_avail = sum(shard_sizes)
                N_eff = min(N, N_frames_avail)
                if N_eff < (self.T + 1):
                    if self.verbose:
                        print(f"[WMDataset] Skipping task={task} source=({dd},{fd}): not enough frames (N_eff={N_eff}) for T={self.T}.")
                    continue

                ep = ep[:N_eff]
                act = act[:N_eff]
                rew = rew[:N_eff]

                # Make episode IDs unique across segments to prevent windows crossing boundaries
                if ep.numel() > 0:
                    seg_max = int(ep.max().item())
                else:
                    seg_max = 0
                ep = ep + ep_offset
                ep_offset += seg_max + 1

                # Build cumulative shard frame counts for this segment, capped at N_eff
                seg_shard_cum = []
                running = 0
                for sz in shard_sizes:
                    running += sz
                    seg_shard_cum.append(min(running, N_eff))
                    if running >= N_eff:
                        break

                seg_eps.append(ep)
                seg_acts.append(act)
                seg_rews.append(rew)
                seg_shards.append(shards)
                seg_num_frames.append(int(N_eff))
                seg_demo_paths.append(dp)
                seg_shard_cums.append(seg_shard_cum)

            if len(seg_eps) == 0:
                if self.verbose:
                    print(f"[WMDataset] Skipping task={task}: missing demo+shards across all sources.")
                continue

            # Concatenate segments for this task
            ep = torch.cat(seg_eps, dim=0)
            act = torch.cat(seg_acts, dim=0)
            rew = torch.cat(seg_rews, dim=0)

            N_eff = int(rew.shape[0])

            # Valid starts: need obs indices i..i+T and transitions at indices i+1..i+T
            start_count = N_eff - self.T  # i in [0, start_count-1]

            # episode consistency: ensure the whole window is within same episode
            ep_ok = (ep[:start_count] == ep[self.T:self.T + start_count])

            # filter invalid transitions (nan action or nan reward)
            act_nan = torch.isnan(act).any(dim=-1)
            rew_nan = torch.isnan(rew)
            step_ok = ~(act_nan | rew_nan)  # length N_eff

            # transitions live at indices 1..N_eff-1
            step_ok2 = step_ok[1:]          # length N_eff-1

            # for each start i, need step_ok at indices (i+1 .. i+T) all true
            cs = torch.cumsum(step_ok2.to(torch.int32), dim=0)
            end = torch.arange(start_count) + (self.T - 1)
            prev = torch.arange(start_count) - 1
            prev_cs = torch.zeros(start_count, dtype=cs.dtype)
            m = prev >= 0
            prev_cs[m] = cs[prev[m]]
            window_sum = cs[end] - prev_cs
            window_ok = (window_sum == self.T)

            valid = ep_ok & window_ok
            valid_idx = valid.nonzero(as_tuple=False).flatten()
            if valid_idx.numel() == 0:
                if self.verbose:
                    print(f"[WMDataset] Skipping task={task}: no valid windows after filtering.")
                continue

            # --- per-task action_dim + mask from tasks.json ---
            act_dim = self.A
            if self.task_meta is not None and task in self.task_meta:
                md = self.task_meta[task]
                if "action_dim" in md:
                    try:
                        act_dim = int(md["action_dim"])
                    except Exception:
                        act_dim = self.A
            act_dim = max(0, min(act_dim, self.A))

            act_mask_1d = torch.zeros(self.A, dtype=torch.float32)
            if act_dim > 0:
                act_mask_1d[:act_dim] = 1.0

            # --- per-task language embedding from tasks.json ---
            lang = self._zero_lang
            if self.lang_dim > 0 and self.task_meta is not None and task in self.task_meta and "text_embedding" in self.task_meta[task]:
                te = self.task_meta[task]["text_embedding"]
                l = torch.tensor(te, dtype=torch.float32)
                if l.numel() != self.lang_dim:
                    raise RuntimeError(f"text_embedding dim mismatch for task {task}: {tuple(l.shape)} vs {self.lang_dim}")
                lang = l

            # Store
            self.tasks.append(task)

            self.demo_paths.append(seg_demo_paths)

            # per-task segments of shard lists + cumulative frame counts
            self.shard_lists.append(seg_shards)
            self.shard_cum.append(seg_shard_cums)
            cum = []
            s = 0
            for nf in seg_num_frames:
                s += int(nf)
                cum.append(s)
            self.seg_cum_frames.append(cum)

            self.ep.append(ep)
            self.act.append(act)
            self.rew.append(rew)
            self.valid_starts.append(valid_idx)

            # precomputed metadata per task index
            task_idx = len(self.tasks) - 1
            self._emb_ids.append(torch.tensor(task_idx, dtype=torch.long))
            self._act_dims.append(act_dim)
            self._act_mask_1d.append(act_mask_1d)
            self._lang_embs.append(lang)

            total += int(valid_idx.numel())
            self._cum_counts.append(total)

            if self.verbose:
                print(f"[WMDataset] task={task} segments={len(seg_shards)} N_eff={N_eff} "
                      f"valid={valid_idx.numel()} act_dim={act_dim} lang={'yes' if lang is not self._zero_lang else 'no'}")

        self.num_tasks = len(self.tasks)
        assert self.num_tasks > 0, "No tasks found with both demo .pt and frame shards."

        # Resolve optional per-task sampling weights (iid_sampling only).
        # task_weights is a dict {task_name: weight} specifying relative
        # task draw probabilities. Within a task we keep shard weights
        # proportional to per-shard valid_starts for cache locality.
        # Tasks not present in the dict default to weight 0 (excluded).
        # If None, fall back to the legacy behavior where shard weight =
        # its valid_starts count, which makes P(task) ∝ total valid_starts.
        if task_weights is not None:
            tw_list: list[float] = []
            missing: list[str] = []
            for t in self.tasks:
                if t in task_weights:
                    w = float(task_weights[t])
                    if w < 0:
                        raise ValueError(f"task_weights[{t!r}] is negative: {w}")
                    tw_list.append(w)
                else:
                    missing.append(t)
                    tw_list.append(0.0)
            if missing and self.verbose:
                print(f"[WMDataset] Warning: {len(missing)} loaded tasks have no "
                      f"task_weights entry and will be excluded from sampling "
                      f"(e.g. {missing[:5]})")
            self.task_weights: Optional[list[float]] = tw_list
        else:
            self.task_weights = None

        # Precompute per-shard valid starts for iid_sampling.  Each entry maps
        # a single on-disk shard file to the valid_starts whose first frame
        # lives in that shard, so drawing samples_per_shard consecutive samples
        # from one entry keeps a single file hot in the LRU cache.
        self._shard_entries: list[dict] = []
        self._shard_weights: list[float] = []
        if self.iid_sampling:
            # Collect raw shard entries grouped by task first so we can
            # renormalize shard weights to honor task_weights without
            # changing the intra-task distribution.
            per_task_entries: list[list[tuple[dict, float]]] = [[] for _ in range(self.num_tasks)]
            for ti in range(self.num_tasks):
                seg_cum = self.seg_cum_frames[ti]
                vs = self.valid_starts[ti]
                for si in range(len(self.shard_cum[ti])):
                    seg_offset = 0 if si == 0 else seg_cum[si - 1]
                    for shi in range(len(self.shard_cum[ti][si])):
                        local_start = 0 if shi == 0 else self.shard_cum[ti][si][shi - 1]
                        local_end = self.shard_cum[ti][si][shi]
                        g_start = seg_offset + local_start
                        g_end = seg_offset + local_end
                        mask = (vs >= g_start) & (vs < g_end)
                        shard_vs = vs[mask]
                        if shard_vs.numel() > 0:
                            entry = {"task_idx": ti, "valid_starts": shard_vs}
                            per_task_entries[ti].append((entry, float(shard_vs.numel())))

            if self.task_weights is None:
                # Legacy path: shard weight = its valid_starts count.
                for ti in range(self.num_tasks):
                    for entry, w in per_task_entries[ti]:
                        self._shard_entries.append(entry)
                        self._shard_weights.append(w)
            else:
                # Rescale so that the total shard weight for task ti equals
                # task_weights[ti] (relative scale is what matters). Tasks
                # with weight 0 contribute no shard entries and are
                # effectively excluded.
                for ti in range(self.num_tasks):
                    entries = per_task_entries[ti]
                    if not entries:
                        continue
                    w_task = float(self.task_weights[ti])
                    if w_task <= 0:
                        continue
                    total_vs = sum(w for _, w in entries)
                    if total_vs <= 0:
                        continue
                    scale = w_task / total_vs
                    for entry, w in entries:
                        self._shard_entries.append(entry)
                        self._shard_weights.append(w * scale)

            if self.verbose:
                msg = f"[WMDataset] iid_sampling: {len(self._shard_entries)} shard entries " \
                      f"(samples_per_shard={self.samples_per_shard})"
                if self.task_weights is not None:
                    nz = sum(1 for w in self.task_weights if w > 0)
                    msg += f" | task_weights active ({nz}/{self.num_tasks} tasks with w>0)"
                print(msg)

        if self.verbose:
            print(f"[WMDataset] Total valid sequences: {self._cum_counts[-1]} across {self.num_tasks} tasks.")

    def __len__(self):
        return self._cum_counts[-1]

    def _lookup(self, idx: int):
        lo, hi = 0, len(self._cum_counts) - 1
        while lo < hi:
            mid = (lo + hi) // 2
            if idx < self._cum_counts[mid]:
                hi = mid
            else:
                lo = mid + 1
        task_idx = lo
        prev = 0 if task_idx == 0 else self._cum_counts[task_idx - 1]
        local = idx - prev
        start = int(self.valid_starts[task_idx][local].item())
        return task_idx, start

    def _cache_get(self, key):
        if key in self._cache:
            v = self._cache.pop(key)
            self._cache[key] = v
            return v
        return None

    def _cache_put(self, key, tensor):
        nbytes = tensor.nbytes
        while self._cache_nbytes + nbytes > self.cache_bytes and len(self._cache) > 0:
            _, v = self._cache.popitem(last=False)
            self._cache_nbytes -= v.nbytes
        self._cache[key] = tensor
        self._cache_nbytes += nbytes

    def _load_shard_frames(self, task_idx: int, seg_idx: int, shard_idx: int) -> torch.Tensor:
        key = (task_idx, seg_idx, shard_idx)
        cached = self._cache_get(key)
        if cached is not None:
            return cached

        path = self.shard_lists[task_idx][seg_idx][shard_idx]
        td = torch.load(path, map_location="cpu", weights_only=False)
        frames = td["frames"]

        # Normalize to (N,3,H,W)
        if frames.ndim == 4 and frames.shape[-1] == 3 and frames.shape[1] != 3:
            frames = frames.permute(0, 3, 1, 2).contiguous()

        # Ensure uint8 storage (robust to float in [0,1] or [0,255])
        if frames.dtype != torch.uint8:
            frames_f = frames.to(torch.float32)
            mx = float(frames_f.max().item()) if frames_f.numel() > 0 else 0.0
            if mx > 1.5:
                frames = frames_f.clamp(0, 255).to(torch.uint8)
            else:
                frames = (frames_f.clamp(0, 1) * 255.0).to(torch.uint8)

        if frames.shape[-2] != self.H or frames.shape[-1] != self.W:
            raise RuntimeError(f"Shard frame size {tuple(frames.shape[-2:])} != {(self.H, self.W)} in {path}")

        self._cache_put(key, frames)
        return frames

    # map global frame idx -> segment -> shard within segment
    def _get_frames(self, task_idx: int, start: int, length: int) -> torch.Tensor:
        out = []
        idx = int(start)
        remaining = int(length)

        seg_cum = self.seg_cum_frames[task_idx]

        while remaining > 0:
            seg_idx = bisect.bisect_right(seg_cum, idx)
            prev_cum = 0 if seg_idx == 0 else seg_cum[seg_idx - 1]
            seg_end = seg_cum[seg_idx]
            local_idx = idx - prev_cum

            # map local frame index within segment to shard using cumulative counts
            shard_cum = self.shard_cum[task_idx][seg_idx]
            shard_idx = bisect.bisect_right(shard_cum, local_idx)
            shard_start = 0 if shard_idx == 0 else shard_cum[shard_idx - 1]
            off = local_idx - shard_start

            frames = self._load_shard_frames(task_idx, seg_idx, shard_idx)

            take = min(remaining, frames.shape[0] - off)

            # don't read past the segment's N_eff (important if segment was truncated)
            seg_remaining = seg_end - idx
            take = min(take, seg_remaining)

            if take <= 0:
                raise RuntimeError(
                    f"Frame indexing error task={self.tasks[task_idx]} idx={idx} seg_idx={seg_idx} "
                    f"local_idx={local_idx} shard_idx={shard_idx} off={off} shard_len={frames.shape[0]}"
                )

            out.append(frames[off:off + take])
            idx += take
            remaining -= take

        return torch.cat(out, dim=0)  # (length,3,H,W) uint8

    def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
        if self.iid_sampling:
            # Locality-friendly sampling: draw samples_per_shard consecutive
            # samples from the same on-disk shard before switching, keeping
            # the file hot in the LRU cache and reducing I/O.
            if (self._cur_shard_entry is None
                    or self._cur_shard_draws >= self.samples_per_shard):
                self._cur_shard_entry = _random.choices(
                    range(len(self._shard_entries)),
                    weights=self._shard_weights,
                    k=1,
                )[0]
                self._cur_shard_draws = 0
            entry = self._shard_entries[self._cur_shard_entry]
            self._cur_shard_draws += 1
            task_idx = entry["task_idx"]
            vs = entry["valid_starts"]
            start = int(vs[_random.randrange(vs.numel())].item())
        else:
            task_idx, start = self._lookup(int(idx))

        obs = self._get_frames(task_idx, start, self.T + 1)  # (T+1,3,H,W) uint8


        act = self.act[task_idx][start + 1 : start + 1 + self.T]   # (T,16) float32 (padded)
        rew = self.rew[task_idx][start + 1 : start + 1 + self.T]   # (T,) float32

        Ad = int(self._act_dims[task_idx])
        act_padded = torch.zeros(self.T, self.A, dtype=torch.float32)
        if Ad > 0:
            act_padded[:, :Ad] = torch.nan_to_num(act[:, :Ad], nan=0.0)

        act_mask = self._act_mask_1d[task_idx][None, :].expand(self.T, self.A).contiguous()

        return {
            "obs": obs,
            "act": act_padded,
            "act_mask": act_mask,
            "rew": rew,
            "lang_emb": self._lang_embs[task_idx],
            "emb_id": self._emb_ids[task_idx],
        }


def collate_batch(batch):
    out = {}
    for k in batch[0].keys():
        out[k] = torch.stack([b[k] for b in batch], dim=0)
    return out