Repository navigation
<
8000
div class="treeToggleWrapper__9kKsad">
Expand file tree
/
Copy pathtrain.py
More file actions
792 lines (729 loc) · 34.8 KB
/
Copy pathtrain.py
File metadata and controls
792 lines (729 loc) · 34.8 KB
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
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
"""Weighted any-to-any cotraining with FloMo."""
import datetime
import os
import time
from collections import defaultdict
from typing import Dict, List
import hydra
import numpy as np
import torch
import wandb
from einops import rearrange, repeat
from omegaconf import OmegaConf
from tqdm import tqdm
from torch.utils.data import ConcatDataset, DataLoader, RandomSampler
from flomo.loaders.cotrain_loader import (
CotrainCollate,
build_cotrain_dataloader,
build_cotrain_dataset,
cotrain_batch_to_device,
derive_active_modalities,
)
from flomo.models.modalities import (
ActionModality,
MotionTokensWanPretrainedModality,
)
from flomo.models.FloMo import FloMo
from flomo.utils.log_utils import BrokenSampleLog
from flomo.utils.grad_norm_logging import log_module_gradients
from flomo.utils.config_validation import validate_train_config
from flomo.utils.distributed import (
create_accelerator,
is_main_process,
log_main,
prepare_training_components,
register_grad_comm_hook,
save_training_checkpoint,
suppress_non_main_prints,
wandb_log,
)
from flomo.utils.motion_stats import _install_motion_token_stats
from flomo.utils.train_utils import (
compute_mid_epoch_resume,
get_checkpoint_dir,
get_lr_scheduler,
get_optim,
get_root_dir,
latest_checkpoint_from_dir,
load_checkpoint,
safe_torch_compile,
segfault_resilient_iter,
setup_wandb_tmp_dirs,
unwrap_model_for_checkpoint,
)
def _compute_ar_metrics(pred_pos, gt_pos):
"""MSE/L1 + moving/nonmoving variants on (B, V, T, N, D) positions.
Moving/nonmoving keys are skipped when their mask is empty."""
from flomo.utils.data_utils import compute_moving_indices
diff = pred_pos.float() - gt_pos.float()
out = {"mse": diff.pow(2).mean().item(), "l1": diff.abs().mean().item()}
moving = compute_moving_indices(gt_pos).bool() # (B, V, N)
moving_mask = repeat(moving, "b v n -> b v t n d", t=gt_pos.shape[2], d=gt_pos.shape[-1])
for tag, mask in (("moving", moving_mask), ("nonmoving", ~moving_mask)):
if mask.any():
d = diff[mask]
out[f"mse_{tag}"] = d.pow(2).mean().item()
out[f"l1_{tag}"] = d.abs().mean().item()
return out
@torch.no_grad()
def _ar_metrics_from_pred_latents(inner, batch, pred_latents):
"""Return per-modality AR MSE/L1; skip modalities without position decoders.
Cumulative-displacement tracks are zero-anchored. Reconstruct GT and
predicted positions from `batch["init_pts"]` in physical units.
"""
out = {}
for name, pred_latent in pred_latents.items():
mod = inner.modalities[name]
if isinstance(mod, ActionModality):
gt_raw, real_idx = mod.gather_real(batch[name], batch[f"{name}_mask"]) # (N_real, T_a, action_dim)
if gt_raw is None:
continue
diff = pred_latent[real_idx].float() - gt_raw.float()
out[name] = {"mse": diff.pow(2).mean().item(), "l1": diff.abs().mean().item()}
continue
gt_raw, real_idx = mod.gather_real(batch[name], batch[f"{name}_mask"]) # (N_real, V, T, N, D)
if gt_raw is None:
continue
ds_idx = None
if isinstance(mod, MotionTokensWanPretrainedModality):
# Ramen encoding skips ds_idx, but decoding needs per-dataset bounds.
ds_idx = torch.tensor(
[mod._ds_to_idx[mod._batch_ds_names[i]] for i in real_idx.tolist()],
device=pred_latent.device, dtype=torch.long,
)
if getattr(mod, "supervision", None) == "cumulative_disp":
if "init_pts" not in batch:
continue
init_real = torch.stack(
[batch["init_pts"][i] for i in real_idx.tolist()], dim=0,
).to(gt_raw) # (N_real, V, 1, N, D)
T_vel = gt_raw.shape[2] - 1
# Undo loader Ramen normalization so GT and predictions share physical units.
if getattr(mod, "tracks_ramen_normalized", False) and ds_idx is not None:
side = int(round(gt_raw.shape[3] ** 0.5))
gt_grid = rearrange(gt_raw[:, 0, 1:], "b t (h w) c -> b c t h w", h=side, w=side)
gt_phys = torch.stack([
mod._un_ramen(gt_grid[b], int(ds_idx[b]), t_lo=1, t_hi=1 + T_vel)
for b in range(gt_grid.shape[0])
], dim=0)
gt_cd = rearrange(gt_phys, "b c t h w -> b 1 t (h w) c").to(init_real)
else:
gt_cd = gt_raw[:, :, 1:]
gt_pos = torch.cat([init_real, init_real + gt_cd], dim=2)
init_pts = init_real
else:
init_pts = gt_raw[:, :, 0:1]
gt_pos = gt_raw
pred_real = pred_latent[real_idx]
T_vel = gt_pos.shape[2] - 1
if isinstance(mod, MotionTokensWanPretrainedModality):
pred_pos = mod.decode_positions(pred_real, ds_idx, init_pts, T_vel)
else:
continue
out[name] = _compute_ar_metrics(pred_pos, gt_pos[:, :, : pred_pos.shape[2]])
return out
def _share_normalizer_within_type(per_ds, ds_cfgs, accelerator):
"""Share one action normalizer over each dataset type's combined train data.
Types with at most one normalizer need no recomputation.
"""
by_type: Dict[str, List[int]] = {}
for i, (d, c) in enumerate(zip(per_ds, ds_cfgs)):
if getattr(d, "normalizer", None) is None:
continue
by_type.setdefault(str(c.type), []).append(i)
for type_name, idxs in by_type.items():
if len(idxs) < 2:
continue
# Aggregate unique (task, demo_id) across all instances' index_maps.
seen: set = set()
seqs: List[np.ndarray] = []
action_space = per_ds[idxs[0]].action_space
horizon = per_ds[idxs[0]].true_horizon
for i in idxs:
d = per_ds[i]
for entry in d.index_map:
key = (entry["task"], entry["demo_id"])
if key in seen:
continue
seen.add(key)
with d._open(entry["demo_path"]) as f:
seqs.append(f[f"actions/{action_space}"][:].astype(np.float32))
from flomo.utils.action_normalizer import compute_normalizer_stats
shared = compute_normalizer_stats(
action_sequences=seqs, reference_sequences=seqs,
action_space=action_space, horizon=horizon,
)
shared.metadata = {
"tasks": sorted({t for t, _ in seen}),
"num_demos": len(seqs),
"shared_across": [str(ds_cfgs[i].name) for i in idxs],
}
for i in idxs:
per_ds[i].set_normalizer(shared)
log_main(
f"[normalizer-share] type={type_name}: shared normalizer across "
f"{len(idxs)} train datasets ({[str(ds_cfgs[i].name) for i in idxs]}); "
f"{len(seqs)} demos, {len(seen)} unique (task, demo).",
accelerator=accelerator,
)
def _normalizers_by_type(per_ds, ds_cfgs):
"""Pick each type's normalizer after `_share_normalizer_within_type`."""
by_type = {}
for d, c in zip(per_ds, ds_cfgs):
if c.type in by_type:
continue
nrm = getattr(d, "normalizer", None)
if nrm is not None:
by_type[c.type] = nrm
return by_type
def _build_val_loader(cfg, universal, partial, type_normalizers):
"""Sequential val loader. Each val dataset takes its type's normalizer so
train and val share one action scale per dataset type."""
if not cfg.val_datasets:
return None
val_ds = [
build_cotrain_dataset(d, cfg.forward_dynamics, None)
for d in cfg.val_datasets
]
for val_d, val_cfg in zip(val_ds, cfg.val_datasets):
if not getattr(val_d, "normalize_actions", False):
continue
if val_cfg.type not in type_normalizers:
raise RuntimeError(
f"Val dataset '{val_cfg.name}' (type={val_cfg.type}) has "
f"normalize_actions=True but no train dataset of type "
f"{val_cfg.type} provides a normalizer."
)
val_d.set_normalizer(type_normalizers[val_cfg.type])
combined = ConcatDataset(val_ds)
collate = CotrainCollate(universal=universal, partial=partial)
n_steps = cfg.val_steps_per_epoch
sampler = RandomSampler(
combined,
replacement=False,
num_samples=int(n_steps) * int(cfg.batch_size) if n_steps is not None else len(combined),
)
return DataLoader(
combined,
batch_size=int(cfg.gpu_max_bs),
sampler=sampler,
num_workers=int(cfg.num_workers),
pin_memory=True,
drop_last=False,
collate_fn=collate,
)
def _reduce_batched(mean_groups, sum_groups, accelerator):
"""ONE all-reduce. `mean_groups`: {grp: ({k: sum}, {k: count})} → world means.
`sum_groups`: {grp: {k: sum}} → world sums. Mean entries with zero world count
are omitted. All ranks must agree on every key set."""
mean_meta = [(g, sorted(s.keys())) for g, (s, _) in mean_groups.items()]
sum_meta = [(g, sorted(s.keys())) for g, s in sum_groups.items()]
flat_s = [float(mean_groups[g][0][k]) for g, ks in mean_meta for k in ks]
flat_c = [float(mean_groups[g][1][k]) for g, ks in mean_meta for k in ks]
flat_p = [float(sum_groups[g][k]) for g, ks in sum_meta for k in ks]
if not flat_s and not flat_p:
return {g: {} for g, _ in mean_meta}, {g: {} for g, _ in sum_meta}
payload = torch.tensor(flat_s + flat_c + flat_p, device=accelerator.device, dtype=torch.float64)
payload = accelerator.reduce(payload, "sum")
n = len(flat_s)
means, i = {}, 0
for g, ks in mean_meta:
d = {}
for k in ks:
cnt = payload[i + n].item()
if cnt > 0:
d[k] = (payload[i] / payload[i + n]).item()
i += 1
means[g] = d
sums, j = {}, 2 * n
for g, ks in sum_meta:
sums[g] = {k: payload[j + off].item() for off, k in enumerate(ks)}
j += len(ks)
return means, sums
class _MicroBatchMeans:
"""Per-rank accumulator over micro-batches; `flush` does ONE all-reduce returning
world means (loss/bd/pd) and world sums (per-dataset sample counts).
`pd_keys` is the canonical (dataset, modality) cross-product; `ds_keys` is the
canonical dataset list — both fixed so all ranks agree on the key set."""
def __init__(self, pd_keys, ds_keys):
self._pd_keys = pd_keys
self._ds_keys = ds_keys
self.reset()
def reset(self):
self.loss_sum, self.loss_n = 0.0, 0
self.bd_sum, self.bd_n = defaultdict(float), defaultdict(int)
self.pd_sum = {k: 0.0 for k in self._pd_keys}
self.pd_n = {k: 0 for k in self._pd_keys}
self.ds_count = {k: 0 for k in self._ds_keys}
def add(self, loss, info, dataset_names):
self.loss_sum += float(loss.item())
self.loss_n += 1
for k, v in info["loss_breakdown"].items():
self.bd_sum[k] += float(v.item())
self.bd_n[k] += 1
for k, v in info["per_dataset_breakdown"].items():
if k in self.pd_sum:
self.pd_sum[k] += float(v.item())
self.pd_n[k] += 1
for name in dataset_names:
self.ds_count[name] += 1
def flush(self, accelerator):
means, sums = _reduce_batched(
mean_groups={
"loss": ({"_": self.loss_sum}, {"_": self.loss_n}),
"bd": (dict(self.bd_sum), dict(self.bd_n)),
"pd": (self.pd_sum, self.pd_n),
},
sum_groups={"ds": dict(self.ds_count)},
accelerator=accelerator,
)
loss = means["loss"]["_"]
bd, pd, ds = means["bd"], means["pd"], sums["ds"]
self.reset()
return loss, bd, pd, ds
def _reduce_ar(per_mod, inner, accelerator):
"""World-mean AR metrics. Canonical key set = decoder-bearing modalities × metric names."""
ar_keys = []
for name, mod in inner.modalities.items():
if isinstance(mod, MotionTokensWanPretrainedModality):
for k in ("mse", "l1", "mse_moving", "l1_moving", "mse_nonmoving", "l1_nonmoving"):
ar_keys.append((name, k))
elif isinstance(mod, ActionModality):
for k in ("mse", "l1"):
ar_keys.append((name, k))
local_sum = {k: 0.0 for k in ar_keys}
local_n = {k: 0 for k in ar_keys}
for name, m in per_mod.items():
for k, v in m.items():
key = (name, k)
if key in local_sum:
local_sum[key] += float(v)
local_n[key] += 1
return _reduce_batched({"_": (local_sum, local_n)}, {}, accelerator)[0]["_"]
def _val_sub_dataset_keys(val_loader, active_outputs):
"""Return `(dataset/subname, modality)` val keys, omitting None subnames."""
raw = val_loader.dataset
datasets = raw.datasets if hasattr(raw, "datasets") else [raw]
out = []
for d in datasets:
if not (hasattr(d, "cotrain_name") and hasattr(d, "index_map")):
continue
if not hasattr(d, "_cotrain_subnames_cache"):
d._cotrain_subnames_cache = sorted(
{d._cotrain_subname(idx) for idx in d.index_map} - {None}
)
for s in d._cotrain_subnames_cache:
composite = f"{d.cotrain_name}/{s}"
for m in active_outputs:
out.append((composite, m))
return out
@torch.no_grad()
def _validate(model, val_loader, accelerator, cfg, num_inf_steps, epoch, train_steps, val_steps, broken_log):
"""Log teacher-forced metrics and periodic AR metrics as world means.
Reduce at optimizer-step boundaries and keep val_steps monotonic across epochs.
"""
model.eval()
inner = unwrap_model_for_checkpoint(model)
pd_keys = [(str(d.name), m) for d in cfg.val_datasets for m in inner.active_outputs]
pd_keys += _val_sub_dataset_keys(val_loader, inner.active_outputs)
ds_keys = [str(d.name) for d in cfg.val_datasets]
tf_sum, tf_count = defaultdict(float), defaultdict(int)
pd_sum, pd_count = defaultdict(float), defaultdict(int)
ar_sum = defaultdict(lambda: defaultdict(float))
ar_count = defaultdict(lambda: defaultdict(int))
loss_sum, loss_count = 0.0, 0
ar_interval = int(cfg.val_ar_interval)
grad_accum = accelerator.gradient_accumulation_steps
n_batches = len(val_loader)
mtr = _MicroBatchMeans(pd_keys, ds_keys)
for batch_idx, batch in enumerate(tqdm(val_loader, desc="val",
disable=not is_main_process(accelerator))):
broken_log.ingest(batch.pop("_broken_samples", []), phase="val")
batch = cotrain_batch_to_device(batch, accelerator.device, non_blocking=True)
with accelerator.autocast():
_pred, loss = model(batch)
mtr.add(loss, inner.last_train_info, batch["dataset_name"])
is_opt_step = ((batch_idx + 1) % grad_accum == 0) or (batch_idx + 1 == n_batches)
if not is_opt_step:
continue
val_steps += 1
avg_loss, avg_breakdown, avg_pd_breakdown, _ = mtr.flush(accelerator)
log_dict = {
"val_steps": val_steps,
"val/loss": avg_loss,
"val/epoch": epoch,
**{f"val/loss/{k}": v for k, v in avg_breakdown.items()},
**{f"val/loss/{ds}/{mod}": v for (ds, mod), v in avg_pd_breakdown.items()},
}
loss_sum += avg_loss
loss_count += 1
for k, v in avg_breakdown.items():
tf_sum[k] += v
tf_count[k] += 1
for (ds, mod), v in avg_pd_breakdown.items():
pd_key = f"{ds}/{mod}"
pd_sum[pd_key] += v
pd_count[pd_key] += 1
if ar_interval > 0 and (val_steps % ar_interval == 0):
with accelerator.autocast():
pred_latents = inner.infer(batch, num_inf_steps)
per_mod = _ar_metrics_from_pred_latents(inner, batch, pred_latents)
avg_ar = _reduce_ar(per_mod, inner, accelerator)
for (name, k), v in avg_ar.items():
log_dict[f"val/ar/{name}/{k}"] = v
ar_sum[name][k] += v
ar_count[name][k] += 1
_log_pred_viz(model, batch, val_steps, accelerator, num_inf_steps,
phase="val", pred_latents=pred_latents)
_log_cond_img_viz(model, batch, val_steps, accelerator, phase="val")
wandb_log(log_dict, accelerator)
avg_dict = {"val_steps": val_steps, "train_steps": train_steps, "val/epoch": epoch}
if loss_count > 0:
v = loss_sum / loss_count
avg_dict["val/loss_avg"] = v
avg_dict["val/loss_avg_by_train"] = v
for k in tf_sum:
v = tf_sum[k] / tf_count[k]
avg_dict[f"val/loss_avg/{k}"] = v
avg_dict[f"val/loss_avg_by_train/{k}"] = v
for k in pd_sum:
v = pd_sum[k] / pd_count[k]
avg_dict[f"val/loss_avg/{k}"] = v
avg_dict[f"val/loss_avg_by_train/{k}"] = v
for name, ks in ar_sum.items():
for k, s in ks.items():
avg_dict[f"val/ar_avg/{name}/{k}"] = s / ar_count[name][k]
if len(avg_dict) > 1:
wandb_log(avg_dict, accelerator)
if is_main_process(accelerator) and broken_log.unique:
wandb_log({"skipped/table": broken_log.wandb_table()}, accelerator)
model.train()
return val_steps
def _log_pred_viz(model, batch, step, accelerator, num_inference_steps, phase, pred_latents=None):
"""Log source/GT/prediction/error panels and projected tracks per modality."""
if not is_main_process(accelerator):
return
inner = unwrap_model_for_checkpoint(model)
if pred_latents is None:
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
pred_latents = inner.infer(batch, num_inference_steps)
log_dict = {f"{phase}_steps": step}
for name, pred_latent in pred_latents.items():
mod = inner.modalities[name]
flow_out = mod.viz_pred(batch, pred_latent)
if flow_out is not None:
video, caption = flow_out
log_dict[f"viz/{phase}_flow/{name}"] = wandb.Video(video, fps=4.0, format="mp4", caption=caption)
# Also log the modality's optional loss-target-space panel.
if hasattr(mod, "viz_pred_normalized"):
norm_out = mod.viz_pred_normalized(batch, pred_latent)
if norm_out is not None:
video, caption = norm_out
log_dict[f"viz/{phase}_flow_normalized/{name}"] = wandb.Video(
video, fps=4.0, format="mp4", caption=caption,
)
if hasattr(mod, "viz_tracks"):
tr_out = mod.viz_tracks(batch, pred_latent)
if tr_out is not None:
video, caption = tr_out
log_dict[f"viz/{phase}_tracks/{name}"] = wandb.Video(video, fps=4.0, format="mp4", caption=caption)
if len(log_dict) > 1:
wandb_log(log_dict, accelerator)
@torch.no_grad()
def _log_cond_img_viz(model, batch, step, accelerator, phase):
"""VAE-decode each `img_<view>` input stream for the first real row and log
under `viz/{phase}_cond_img/{view}`. Sanity check for the wrist cameras,
which the pred-viz panel doesn't cover (img is input-only)."""
if not is_main_process(accelerator):
return
inner = unwrap_model_for_checkpoint(model)
log_dict = {f"{phase}_steps": step}
for name, mod in inner.modalities.items():
if not name.startswith("img_"):
continue
mask = batch[f"{name}_mask"]
real_idx = mask.nonzero(as_tuple=True)[0]
if real_idx.numel() == 0:
continue
i = int(real_idx[0])
raw = batch[name][i].unsqueeze(0)
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
latent = mod.encode(raw)
rgb = mod.vae.single_decode(latent.to(torch.bfloat16)) # (1, 3, T_pix, S, S) in [-1, 1]
frame = rgb[0, :, 0].float() # (3, S, S) cond frame
u8 = ((frame + 1.0) * 0.5 * 255.0).clamp(0, 255).to(torch.uint8)
img = rearrange(u8, "c h w -> h w c").cpu().numpy()
view = name[len("img_"):]
log_dict[f"viz/{phase}_cond_img/{view}"] = wandb.Image(img, caption=view)
if len(log_dict) > 1:
wandb_log(log_dict, accelerator)
@hydra.main(config_path="cfg", config_name="train", version_base="1.2")
def main(cfg):
"""Any-to-any cotrain entry."""
validate_train_config(cfg)
run_name = str(cfg.run_name) or datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
if cfg.root_dir is None:
cfg.root_dir = get_root_dir()
setup_wandb_tmp_dirs()
accelerator = create_accelerator(cfg)
suppress_non_main_prints(accelerator)
log_main("================== FINAL CONFIG (cotrain) ==================", accelerator=accelerator)
log_main(OmegaConf.to_yaml(cfg), accelerator=accelerator)
device = accelerator.device
# Per-dataset instances + sample weights.
per_ds = [
build_cotrain_dataset(d, cfg.forward_dynamics, cfg.augmentations)
for d in cfg.datasets
]
per_ds_weights = [float(d.weight) for d in cfg.datasets]
# When a type has multiple train instances with their own normalizers
# (e.g. realyam_train + realyam_play_train), compute one shared normalizer
# over the union and apply to all. Otherwise each instance's actions get
# normalized to its own scale and the action head sees mixed targets.
_share_normalizer_within_type(per_ds, cfg.datasets, accelerator)
# One normalizer per dataset type, used by val datasets of the same type
# and baked into the checkpoint keyed by dataset type. Inference picks the
# entry by type (default "realyam", overridable via --normalizer-type).
type_normalizers = _normalizers_by_type(per_ds, cfg.datasets)
normalizer_stats_dict = (
{t: nrm.to_dict() for t, nrm in type_normalizers.items()}
if type_normalizers else None
)
_, _, universal, partial = derive_active_modalities(
cfg.datasets, cfg.forward_dynamics.modalities,
)
log_main(
f"[cotrain] {len(per_ds)} datasets; universal={universal}; partial={partial}",
accelerator=accelerator,
)
steps_per_epoch = cfg.train_steps_per_epoch or (sum(len(d) for d in per_ds) // cfg.batch_size)
# DataLoader yields micro-batches of size gpu_max_bs; accelerator.accumulate
# collects (batch_size / gpu_max_bs) of them per optimizer step. Sampler draws
# train_steps_per_epoch * batch_size samples per epoch — i.e. one weighted-replacement
# draw per forward pass.
train_loader = build_cotrain_dataloader(
per_dataset_instances=per_ds,
per_dataset_weights=per_ds_weights,
universal_modalities=universal,
partial_modalities=partial,
batch_size=cfg.gpu_max_bs,
num_samples=steps_per_epoch * cfg.batch_size,
num_workers=cfg.num_workers,
drop_last=True,
pin_memory=True,
prefetch_factor=2,
persistent_workers=False,
)
val_loader = _build_val_loader(cfg, universal, partial, type_normalizers)
model = FloMo(cfg).to(device)
model.setup_model()
_install_motion_token_stats(model, cfg, per_ds, accelerator)
optim_cfg = cfg.optim
optimizer = get_optim(
model, lr=optim_cfg.lr, weight_decay=optim_cfg.weight_decay,
betas=optim_cfg.adam_betas,
)
scheduler = get_lr_scheduler(
optimizer, lr_schedule=optim_cfg.lr_schedule,
warmup_steps=optim_cfg.warmup_steps,
decay_steps=optim_cfg.decay_steps,
eta_min=optim_cfg.eta_min,
)
scaler = getattr(accelerator, "scaler", None)
lora_enabled = (cfg.forward_dynamics.wan.finetune_mode == "lora")
checkpoint_dir = get_checkpoint_dir(stage="forward_dynamics", run_name=run_name, resume=cfg.resume)
log_main(f"Checkpoint dir: {checkpoint_dir}", accelerator=accelerator)
if cfg.resume and cfg.checkpoint is None:
cfg.checkpoint = latest_checkpoint_from_dir(checkpoint_dir)
if cfg.checkpoint is not None:
log_main(f"[resume] auto-resume from {cfg.checkpoint}", accelerator=accelerator)
if cfg.compile:
log_main("[compile] wrapping model with torch.compile (first step will be slow)", accelerator=accelerator)
model = safe_torch_compile(model)
train_steps = 0
val_steps = 0
wandb_run_id = None
samples_seen_ckpt = None
first_epoch_remaining = None
if cfg.checkpoint is not None:
init_only = not cfg.resume
model, optimizer, scheduler, scaler, ckpt_info = load_checkpoint(
cfg.checkpoint, model, optimizer, scheduler, scaler,
skip_optimizer=init_only or cfg.resume_skip_optimizer,
)
if init_only:
start_epoch = 1
log_main(
f"[init] loaded model weights from {cfg.checkpoint}; starting a fresh run",
accelerator=accelerator,
)
else:
train_steps = ckpt_info["train_steps"]
val_steps = ckpt_info["val_steps"]
# Resume mid-epoch; pass the global (unprepared-loader) grad_accum.
start_epoch, first_epoch_remaining = compute_mid_epoch_resume(
train_steps, train_loader, cfg.batch_size // cfg.gpu_max_bs, ckpt_info,
)
if first_epoch_remaining == 0:
start_epoch, first_epoch_remaining = start_epoch + 1, None
wandb_run_id = ckpt_info["wandb_run_id"]
samples_seen_ckpt = ckpt_info["samples_seen"]
# When scheduler is None, the saved (possibly mid-decay) lr sticks; push cfg.optim.lr back.
if scheduler is None:
for pg in optimizer.param_groups:
pg["lr"] = float(cfg.optim.lr)
log_main(f"[resume] scheduler disabled; lr -> {cfg.optim.lr}", accelerator=accelerator)
else:
start_epoch = 1
if cfg.wandb_run_id is not None:
wandb_run_id = cfg.wandb_run_id
model, optimizer, train_loader, scheduler, val_loader = prepare_training_components(
accelerator, model, optimizer, train_loader, scheduler=scheduler, val_loader=val_loader,
)
register_grad_comm_hook(model, accelerator, cfg.grad_comm)
if is_main_process(accelerator):
wandb.init(
entity=cfg.wandb_entity, project=cfg.wandb_project, group=cfg.wandb_group,
name=run_name,
id=wandb_run_id,
resume="allow" if wandb_run_id is not None else None,
config=OmegaConf.to_container(cfg, resolve=True, throw_on_missing=True),
mode='disabled' if not cfg.use_wandb else 'online',
allow_val_change=True,
)
wandb.define_metric("train_steps")
wandb.define_metric("val_steps")
wandb.define_metric("train/*", step_metric="train_steps")
wandb.define_metric("val/loss_avg_by_train*", step_metric="train_steps")
wandb.define_metric("val/*", step_metric="val_steps")
wandb.define_metric("viz/train_*", step_metric="train_steps")
wandb.define_metric("viz/val_*", step_metric="val_steps")
wandb.define_metric("skipped/*", step_metric="train_steps")
broken_log = BrokenSampleLog(checkpoint_dir, accelerator.process_index)
num_inf_steps = int(cfg.forward_dynamics.wan.num_inference_steps)
interval = int(cfg.log_train_video_interval)
# Track global per-dataset draws for data_pass metrics and checkpoint resume.
ds_keys = [str(d.name) for d in cfg.datasets]
ds_sizes = [float(len(p)) for p in per_ds]
samples_seen = list(samples_seen_ckpt) if samples_seen_ckpt is not None else [0.0] * len(ds_keys)
# New dataset slots start at zero; discard removed slots on resume.
samples_seen = (samples_seen + [0.0] * len(ds_keys))[:len(ds_keys)]
inner = unwrap_model_for_checkpoint(model)
pd_keys = [(str(d.name), m) for d in cfg.datasets for m in inner.active_outputs]
for epoch in range(start_epoch, cfg.num_epochs + 1):
model.train()
t_step = time.perf_counter()
mtr = _MicroBatchMeans(pd_keys, ds_keys)
epoch_budget = first_epoch_remaining if epoch == start_epoch else None
epoch_opt_steps = 0
for batch in tqdm(segfault_resilient_iter(train_loader, accelerator),
desc=f"cotrain ep{epoch}", total=len(train_loader),
disable=not is_main_process(accelerator)):
try:
broken_log.ingest(batch.pop("_broken_samples", []), phase="train")
batch = cotrain_batch_to_device(batch, device, non_blocking=True)
with accelerator.accumulate(model):
with accelerator.autocast():
_pred, loss = model(batch)
accelerator.backward(loss)
grad_norm = None
grad_norms = {}
if accelerator.sync_gradients:
# Measure synchronized gradients before clipping. DDP averages
# across ranks; bf16 needs no GradScaler correction.
if cfg.logging.log_grad_norms and is_main_process(accelerator):
grad_norms = log_module_gradients(
model, train_steps, modules_to_monitor=list(cfg.logging.modules_to_monitor)
)
if optim_cfg.clip_grad > 0:
grad_norm = accelerator.clip_grad_norm_(model.parameters(), max_norm=optim_cfg.clip_grad)
optimizer.step()
optimizer.zero_grad(set_to_none=True)
if accelerator.sync_gradients and scheduler is not None:
scheduler.step()
mtr.add(loss, inner.last_train_info, batch["dataset_name"])
if accelerator.sync_gradients:
train_steps += 1
epoch_opt_steps += 1
ckpt_every = cfg.train_steps_per_epoch // cfg.ckpt_saves_per_epoch if cfg.ckpt_saves_per_epoch > 0 else 0
if ckpt_every > 0 and train_steps % ckpt_every == 0:
save_training_checkpoint(
os.path.join(checkpoint_dir, "latest.pt"),
epoch, cfg, model, optimizer, scaler,
train_loss=loss.item(), val_loss=None, scheduler=scheduler,
train_steps=train_steps, val_steps=val_steps, accelerator=accelerator,
steps_per_epoch=steps_per_epoch,
lora_enabled=lora_enabled,
normalizer_stats=normalizer_stats_dict,
samples_seen=list(samples_seen),
)
avg_loss, avg_breakdown, avg_pd_breakdown, ds_counts = mtr.flush(accelerator)
for i, k in enumerate(ds_keys):
samples_seen[i] += ds_counts[k]
if is_main_process(accelerator):
now = time.perf_counter()
log_dict = {
"train_steps": train_steps,
"train/loss": avg_loss,
"train/it_per_sec": 1.0 / max(now - t_step, 1e-9),
"train/epoch": epoch,
**{f"train/data_pass/{k}": samples_seen[i] / ds_sizes[i]
for i, k in enumerate(ds_keys)},
**{f"train/loss/{k}": v for k, v in avg_breakdown.items()},
**{f"train/loss/{ds}/{mod}": v for (ds, mod), v in avg_pd_breakdown.items()},
**broken_log.counter_log(),
}
if scheduler is not None:
log_dict["train/lr"] = scheduler.get_last_lr()[0]
if grad_norm is not None:
log_dict["train/grad_norm"] = float(grad_norm)
log_dict.update({f"grad_norms/{k}": v for k, v in grad_norms.items()})
t_step = now
wandb_log(log_dict, accelerator)
if cfg.generate_video and interval > 0 and (train_steps % interval == 0):
with accelerator.autocast():
pred_latents = inner.infer(batch, num_inf_steps)
per_mod = _ar_metrics_from_pred_latents(inner, batch, pred_latents)
avg_ar = _reduce_ar(per_mod, inner, accelerator)
if is_main_process(accelerator):
_log_pred_viz(model, batch, train_steps, accelerator, num_inf_steps,
phase="train", pred_latents=pred_latents)
_log_cond_img_viz(model, batch, train_steps, accelerator, phase="train")
ar_log = {"train_steps": train_steps,
**{f"train/ar/{name}/{k}": v for (name, k), v in avg_ar.items()}}
if len(ar_log) > 1:
wandb_log(ar_log, accelerator)
if epoch_budget is not None and epoch_opt_steps >= epoch_budget:
break
except RuntimeError as e:
if "DataLoader worker" not in str(e):
raise
log_main(f"[recover] worker died mid-step: {e}; zeroing grads and continuing",
accelerator=accelerator)
optimizer.zero_grad(set_to_none=True)
continue
if val_loader is not None and cfg.eval_interval > 0 and (epoch % cfg.eval_interval == 0):
val_steps = _validate(model, val_loader, accelerator, cfg, num_inf_steps, epoch, train_steps, val_steps, broken_log)
save_training_checkpoint(
os.path.join(checkpoint_dir, "latest.pt"),
epoch, cfg, model, optimizer, scaler,
train_loss=loss.item(), val_loss=None, scheduler=scheduler,
train_steps=train_steps, val_steps=val_steps, accelerator=accelerator,
steps_per_epoch=steps_per_epoch,
lora_enabled=lora_enabled,
normalizer_stats=normalizer_stats_dict,
samples_seen=list(samples_seen),
)
if epoch == 1 or epoch % cfg.save_interval == 0:
save_training_checkpoint(
os.path.join(checkpoint_dir, f"{epoch}.pt"),
epoch, cfg, model, optimizer, scaler,
train_loss=loss.item(), val_loss=None, scheduler=scheduler,
train_steps=train_steps, val_steps=val_steps, accelerator=accelerator,
steps_per_epoch=steps_per_epoch,
lora_enabled=lora_enabled,
normalizer_stats=normalizer_stats_dict,
samples_seen=list(samples_seen),
)
if __name__ == "__main__":
main()