File size: 38,136 Bytes
5f0e48f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Cross-domain compression transfer for power-grid detection (Task 2).

Student YOLO = detection loss
  + logit KD from a big teacher (binary KL on class scores, KL on DFL box distributions)
  + projection-alignment branch: student backbone features -> teacher class-logit space (training only)
  + optional multi-down / shared-up LoRA with a style router (multi-domain adaptation, mergeable per site)

  python gridkd.py train --data D.yaml --model yolo11n.pt --teacher T.pt -- epochs=100 device=0 name=s2
  python gridkd.py eval  --weights a.pt b.pt --data A.yaml B.yaml
  python gridkd.py merge --weights lora.pt --gates 1,0,0,0 --out site.pt   (or --gates fog.yaml: mean gates of that domain)
  python gridkd.py prune --weights s2.pt --ratio 0.3 --out s2_p30.pt   (then train --model s2_p30.pt to recover)
  python gridkd.py lsl --weights lt.pt --pos br.yaml --neg ref_1c.yaml --out lt_lsl.pt   (LoRA on only where inputs look like br)
  python gridkd.py bnsw --weights lt.pt --base s2.pt --out lt_bnsw.pt   (same, decided by the two models' BN statistics)
  python gridkd.py selftest
"""

import argparse
from copy import deepcopy
import math
import sys
from pathlib import Path

import torch
import torch.nn as nn
import torch.nn.functional as F
from ultralytics.models.yolo.detect import DetectionTrainer
from ultralytics.nn.modules import C2f, Conv
from ultralytics.nn.tasks import DetectionModel, attempt_load_one_weight
from ultralytics.utils.loss import v8DetectionLoss
from ultralytics.utils.torch_utils import de_parallel, fuse_conv_and_bn, get_flops

LAYER_ATTRS = ("f", "i", "type", "np")  # ultralytics routes layers by these


def _copy_layer_attrs(dst, src):
    for a in LAYER_ATTRS:
        if hasattr(src, a):
            setattr(dst, a, getattr(src, a))


# ----------------------------------------------------------------------------- LoRA + style router
class StyleRouter(nn.Module):
    """Style code = channel mean/std of stem features (AdaIN statistics) -> softmax gates over K experts."""

    def __init__(self, c, k, tau=1.0):
        super().__init__()
        self.fc, self.tau = nn.Linear(2 * c, k), tau
        self.logits = self.gates = None
        self.fixed = None  # (k,) tensor pins the gates: one site / one style, or before merging

    def forward(self, f):
        s = torch.cat([f.mean((2, 3)), f.flatten(2).std(-1)], 1)
        self.logits = self.fc(s)
        g = (self.logits.float() / self.tau).softmax(-1)
        self.gates = g if self.fixed is None else self.fixed.to(g).expand_as(g)
        return f

    def __getstate__(self):  # per-forward tensors are not state (non-leaf ones also break EMA deepcopy)
        return {**self.__dict__, "logits": None, "gates": None}


class RoutedStem(nn.Module):
    """Layer 0 + router, so gates exist before any LoRA layer runs."""

    def __init__(self, stem, router):
        super().__init__()
        self.stem, self.router = stem, router
        _copy_layer_attrs(self, stem)

    def forward(self, x):
        return self.router(self.stem(x))


class MultiLoRAConv(nn.Module):
    """y = act(BN(Wx) + s * B(sum_k g_k A_k x)): K style-specific down-projections A_k, one shared up-projection B."""

    def __init__(self, m: Conv, router: StyleRouter, r=4, alpha=8):
        super().__init__()
        c, k = m.conv, router.fc.out_features
        self.m, self.k, self.r, self.s = m, k, r, alpha / r
        self.A = nn.Conv2d(c.in_channels, k * r, c.kernel_size, c.stride, c.padding, c.dilation, bias=False)
        self.B = nn.Conv2d(r, c.out_channels, 1, bias=False)
        nn.init.kaiming_uniform_(self.A.weight, a=math.sqrt(5))
        nn.init.zeros_(self.B.weight)  # starts exactly at the base model
        self._router = [router]  # shared reference, deliberately not a submodule
        _copy_layer_attrs(self, m)

    def forward(self, x):
        m = self.m
        y = m.bn(m.conv(x)) if hasattr(m, "bn") else m.conv(x)  # survives ultralytics fuse()
        a = self.A(x)
        n, _, h, w = a.shape
        g = self._router[0].gates.to(a.dtype).view(n, self.k, 1, 1, 1)
        d = self.s * self.B((a.view(n, self.k, self.r, h, w) * g).sum(1))
        sw = getattr(self, "switch", None)  # local-support switch: adapter only on inputs from its own domain
        if sw is None:
            return m.act(y + d)
        on = sw(x, y).to(d).view(n, 1, 1, 1)
        return m.act(on * (y + d) + (1 - on) * sw.off(y))

    @torch.no_grad()
    def merged(self, g):
        """Fold LoRA for fixed gates g (k,) into a plain fused Conv -> zero deployment overhead."""
        m = self.m
        if hasattr(m, "bn"):
            m.conv = fuse_conv_and_bn(m.conv, m.bn)
            delattr(m, "bn")
            m.forward = m.forward_fuse
        A = (self.A.weight.view(self.k, self.r, *self.A.weight.shape[1:]) * g.view(-1, 1, 1, 1, 1).to(self.A.weight)).sum(0)
        m.conv.weight += self.s * torch.einsum("or,rihw->oihw", self.B.weight[:, :, 0, 0], A).to(m.conv.weight)
        return m


def inject_lora(model: DetectionModel, r=4, k=4, alpha=8, tau=1.0):
    seq = model.model
    stem = seq[0]
    router = StyleRouter(stem.conv.out_channels, k, tau)
    seq[0] = RoutedStem(stem, router)
    for mod in list(model.modules()):
        for name, child in list(mod.named_children()):
            if type(child) is Conv and child.conv.groups == 1 and child is not stem:  # skips DWConv / stem
                setattr(mod, name, MultiLoRAConv(child, router, r, alpha))
    return router


def find_router(model):
    return next((m.router for m in model.modules() if isinstance(m, RoutedStem)), None)


@torch.no_grad()
def merge_lora(model, gates):
    """Replace every LoRA wrapper by its merged Conv and drop the router (in place)."""
    g = torch.as_tensor(gates, dtype=torch.float32)
    g = g / g.sum()
    for mod in list(model.modules()):
        for name, child in list(mod.named_children()):
            if isinstance(child, MultiLoRAConv):
                setattr(mod, name, child.merged(g))
            elif isinstance(child, RoutedStem):
                setattr(mod, name, child.stem)
    return model


def lora_params(model):
    return [p for m in model.modules() if isinstance(m, (MultiLoRAConv, StyleRouter)) for n, p in m.named_parameters() if not n.startswith("m.")]


# ----------------------------------------------------------------------------- local-support switch (after LSL, arXiv 2610.02126)
def style_stats(x):
    """Per-image channel mean and log-std (AdaIN statistics) of a feature map -> (n, 2c)."""
    x = x.float().flatten(2)
    return torch.cat([x.mean(-1), x.std(-1).clamp_min(1e-6).log()], 1)


class DiagGMM(nn.Module):
    """Diagonal Gaussian mixture, fit by EM."""

    def __init__(self, mu, var, logw):
        super().__init__()
        self.register_buffer("mu", mu)
        self.register_buffer("var", var)
        self.register_buffer("logw", logw)

    def _ll(self, x):  # (n, d) -> (n, k): log w_k + log N(x | mu_k, var_k)
        return self.logw - 0.5 * ((x[:, None] - self.mu) ** 2 / self.var + self.var.log() + math.log(2 * math.pi)).sum(-1)

    def logp(self, x):
        return self._ll(x.to(self.mu)).logsumexp(-1)

    @classmethod
    @torch.no_grad()
    def fit(cls, x, k=4, iters=100, seed=0):
        x = x.double()
        n, d = x.shape
        k = min(k, n)
        floor = 1e-3 * x.var(0).mean() + 1e-8  # ponytail: one variance floor for all dims; per-dim floors if a dim is near-constant
        idx = torch.randperm(n, generator=torch.Generator().manual_seed(seed))[:k].to(x.device)
        g = cls(x[idx].clone(), x.var(0).expand(k, d).clone() + floor, torch.full((k,), -math.log(k), dtype=x.dtype, device=x.device))
        for _ in range(iters):
            r = g._ll(x).softmax(-1)  # responsibilities
            nk = r.sum(0) + 1e-12
            mu = r.T @ x / nk[:, None]
            g = cls(mu, (r.T @ x**2 / nk[:, None] - mu**2).clamp_min(0) + floor, (nk / n).log())
        return cls(g.mu.float(), g.var.float(), g.logw.float())


class DensitySwitch(nn.Module):
    """Adapter on for an image iff its style statistics at this layer are likelier under the adapter's own training
    data (pos) than under a small set of generic images (neg). Needs no data from earlier domains."""

    def __init__(self, pos, neg):
        super().__init__()
        self.pos, self.neg = pos, neg
        self.hits = self.seen = 0.0  # on-rate bookkeeping for eval

    def forward(self, x, y=None):
        f = style_stats(x)
        on = (self.pos.logp(f) > self.neg.logp(f)).float()
        self.hits, self.seen = self.hits + float(on.sum()), self.seen + len(on)
        return on

    def off(self, y):
        return y


class ConstSwitch(nn.Module):  # adapter forced on (1) or off (0)
    def __init__(self, v):
        super().__init__()
        self.v = v

    def forward(self, x, y=None):
        return x.new_full((x.shape[0],), self.v)

    def off(self, y):
        return y


class BNSwitch(nn.Module):
    """Adapter on for an image iff this layer's base output is likelier under the adapted model's BN statistics
    (target) than under the base model's (source); off also restores the source normalisation, so a closed switch is
    exactly the base layer. Both densities are read from the two checkpoints: no data needed. With a shared `acc`, each
    layer decides on the log-likelihood ratio summed over all layers run so far (one consistent decision per image)."""

    def __init__(self, bn_t, bn_s, acc=None, first=False):
        super().__init__()
        g, b = bn_t.weight.detach(), bn_t.bias.detach()  # BN affine parameters were frozen with the base: same in both
        st, ss = (bn_t.running_var + bn_t.eps).sqrt(), (bn_s.running_var + bn_s.eps).sqrt()
        # base output o = g (z - m_t) / s_t + b; each model's BN statistics give a diagonal Gaussian over o
        self.register_buffer("mu", torch.stack([b, b + g * (bn_s.running_mean - bn_t.running_mean) / st]))
        self.register_buffer("var", torch.stack([bn_t.running_var, bn_s.running_var]) * (g / st) ** 2 + 1e-6)
        self.register_buffer("a", st / ss)  # o -> source-normalised output: a * o + c
        self.register_buffer("c", b * (1 - st / ss) + g * (bn_t.running_mean - bn_s.running_mean) / ss)
        self.acc, self.first = acc, first
        self.hits = self.seen = 0.0

    def forward(self, x, y):
        f = y.float().flatten(2)
        m, v = f.mean(-1)[:, None], f.var(-1, unbiased=False)[:, None]  # (n, 1, c)
        ll = -0.5 * (((m - self.mu) ** 2 + v) / self.var + self.var.log()).sum(-1)  # (n, 2): mean per-pixel log-density
        r = ll[:, 0] - ll[:, 1]
        if self.acc is not None:
            r = self.acc["r"] = r if self.first else self.acc["r"] + r
        on = (r > 0).float()
        self.hits, self.seen = self.hits + float(on.sum()), self.seen + len(on)
        return on

    def off(self, y):
        return y * self.a.view(1, -1, 1, 1).to(y) + self.c.view(1, -1, 1, 1).to(y)


class SwitchedConv(nn.Module):
    """A conv without adapter whose BN statistics follow the same switch (on: adapted model's, off: base model's)."""

    def __init__(self, m, switch):
        super().__init__()
        self.m, self.switch = m, switch

    def forward(self, x):
        m = self.m
        y = m.bn(m.conv(x)) if hasattr(m, "bn") else m.conv(x)
        on = self.switch(x, y).to(y).view(-1, 1, 1, 1)
        return m.act(on * y + (1 - on) * self.switch.off(y))


def attach_bn_switches(model, base, cumulative=True):
    """BN switches on every BN layer of an adapted model, against its base model (same architecture, unfused)."""
    ref, mods = dict(base.named_modules()), dict(model.named_modules())
    acc = {} if cumulative else None
    inner = {id(x.m) for x in mods.values() if isinstance(x, MultiLoRAConv)}
    for n, x in mods.items():
        if isinstance(x, MultiLoRAConv):
            x.switch = BNSwitch(x.m.bn, ref[n].bn, acc)
        elif isinstance(x, Conv) and id(x) not in inner:
            parent, _, leaf = n.rpartition(".")
            stem = leaf == "stem"  # RoutedStem holds layer 0's conv as .stem (the base has it at model.0); it runs first
            setattr(mods[parent], leaf, SwitchedConv(x, BNSwitch(x.bn, ref[parent if stem else n].bn, acc, first=stem)))
    sws = [x for x in model.modules() if isinstance(x, BNSwitch)]
    assert acc is None or sum(x.first for x in sws) == 1, "cumulative mode needs the routed stem to reset the sum"
    return len(sws)


# ----------------------------------------------------------------------------- KD criterion
def _bkl(s_logit, t_logit, T):
    """Binary KL(sigmoid(t/T) || sigmoid(s/T)) per element."""
    q = (t_logit / T).sigmoid()
    return F.binary_cross_entropy_with_logits(s_logit / T, q, reduction="none") - F.binary_cross_entropy_with_logits(
        t_logit / T, q, reduction="none"
    )


class KDLoss(v8DetectionLoss):
    """v8 detection loss + [logit KD, projection alignment, router CE]. Items: box cls dfl kd pa rt.

    teacher: one model, or one expert per domain (aligned with `domains`, path substrings of im_file)."""

    def __init__(self, model, teacher=None, proj=None, feats=None, layers=(), T=2.0, w_kd=1.0, w_pa=1.0, router=None,
                 w_rt=0.1, domains=(), route_ce=True):
        super().__init__(model)
        self.teachers = [teacher] if isinstance(teacher, nn.Module) else list(teacher or ())
        self.pa_proj, self.feats, self.layers, self.T = proj, feats, tuple(layers), T  # self.proj is v8's DFL
        self.w_kd, self.w_pa, self.router, self.w_rt, self.domains, self.route_ce = w_kd, w_pa, router, w_rt, domains, route_ce

    def _ids(self, batch):  # domain index per image, -1 if none matches
        return torch.tensor([next((j for j, d in enumerate(self.domains) if d in f), -1) for f in batch["im_file"]], device=self.device)

    @torch.no_grad()
    def _teach(self, img, ids):
        if len(self.teachers) == 1:
            return self._flat(self.teachers[0](img)[1])
        k, out = ids.clamp(min=0), None  # unmatched images go to the first expert
        for j, te in enumerate(self.teachers):
            m = k == j
            if m.any():
                d, c = self._flat(te(img[m])[1])
                out = out or (d.new_zeros(len(img), *d.shape[1:]), c.new_zeros(len(img), *c.shape[1:]))
                out[0][m], out[1][m] = d, c
        return out

    def _flat(self, maps):
        b = maps[0].shape[0]
        d, c = torch.cat([x.view(b, self.no, -1) for x in maps], 2).split((self.reg_max * 4, self.nc), 1)
        return d.permute(0, 2, 1).float(), c.permute(0, 2, 1).float()  # (B,A,4R) (B,A,nc)

    def __call__(self, preds, batch):
        loss, items = super().__call__(preds, batch)
        bs, T = batch["img"].shape[0], self.T
        extra = torch.zeros(3, device=self.device)  # kd, pa, rt
        ids = self._ids(batch) if self.domains else None
        if self.teachers:
            t_dist, t_cls = self._teach(batch["img"], ids)
            s_dist, s_cls = self._flat(preds[1] if isinstance(preds, tuple) else preds)
            # anchors where either model sees an object; keeps background false positives in the loss
            w = torch.maximum(t_cls.sigmoid(), s_cls.detach().sigmoid()).amax(-1)
            wt = t_cls.sigmoid().amax(-1)
            kd_cls = (_bkl(s_cls, t_cls, T).sum(-1) * w).sum() / w.sum().clamp(min=1)
            sd, td = s_dist.view(bs, -1, 4, self.reg_max), t_dist.view(bs, -1, 4, self.reg_max)
            kd_box = F.kl_div((sd / T).log_softmax(-1), (td / T).log_softmax(-1), reduction="none", log_target=True)
            kd_box = (kd_box.sum(-1).mean(-1) * wt).sum() / wt.sum().clamp(min=1)
            extra[0] = (kd_cls + kd_box) * T * T
            if self.pa_proj is not None:
                p = torch.cat([pj(self.feats[i]).flatten(2) for i, pj in zip(self.layers, self.pa_proj)], 2)
                extra[1] = (_bkl(p.permute(0, 2, 1).float(), t_cls, T).sum(-1) * w).sum() / w.sum().clamp(min=1) * T * T
        if self.router is not None and self.route_ce and self.domains:
            if (ids >= 0).any():
                extra[2] = F.cross_entropy(self.router.logits.float()[ids >= 0], ids[ids >= 0])
        wvec = torch.tensor([self.w_kd, self.w_pa, self.w_rt], device=self.device)
        return torch.cat([loss, extra * wvec * bs]), torch.cat([items, extra.detach()])


# ----------------------------------------------------------------------------- trainer
class KDTrainer(DetectionTrainer):
    def __init__(self, kd=None, **kw):
        self.kd = kd or {}
        super().__init__(**kw)
        self.add_callback("on_pretrain_routine_end", KDTrainer._attach)

    def get_model(self, cfg=None, weights=None, verbose=True):
        # a pruned net no longer matches its yaml, so train the loaded module itself
        model = weights if getattr(weights, "pruned", False) else super().get_model(cfg, weights, verbose)
        kd = self.kd
        if kd.get("lora_r"):
            inject_lora(model, kd["lora_r"], kd["lora_k"], kd["lora_alpha"])
        if kd.get("teacher") and kd.get("w_pa"):
            layers, ch = kd["pa_layers"], {}
            hs = [model.model[i].register_forward_hook(lambda m, a, o, i=i: ch.__setitem__(i, o.shape[1])) for i in layers]
            model.eval()(torch.zeros(1, 3, 64, 64))  # probe channel counts
            model.train()
            [h.remove() for h in hs]
            nc = model.model[-1].nc
            model.kd_proj = nn.ModuleList(nn.Sequential(Conv(ch[i], ch[i], 1), nn.Conv2d(ch[i], nc, 1)) for i in layers)
        return model

    def get_validator(self):
        v = super().get_validator()
        self.loss_names = "box_loss", "cls_loss", "dfl_loss", "kd_loss", "pa_loss", "rt_loss"
        return v

    @staticmethod
    def _attach(trainer):
        """After EMA/optimizer exist: hooks + KD criterion on the live model only (keeps checkpoints clean)."""
        kd, model = trainer.kd, de_parallel(trainer.model)
        teachers, feats = [], None
        for w in filter(None, (kd.get("teacher") or "").split(",")):  # one teacher, or one expert per domain
            t, _ = attempt_load_one_weight(w, device=trainer.device, fuse=True)
            assert t.model[-1].nc == model.model[-1].nc, "teacher/student class sets differ"
            teachers.append(t.requires_grad_(False).eval())
        if hasattr(model, "kd_proj"):
            feats = {}
            for i in kd["pa_layers"]:
                model.model[i].register_forward_hook(lambda m, a, o, i=i: feats.__setitem__(i, o))
        router = find_router(model)
        if router is not None:  # LoRA stage: freeze the compressed student, train adapters + router (+ projector)
            keep = {id(p) for p in lora_params(model)} | {id(p) for p in getattr(model, "kd_proj", nn.Module()).parameters()}
            for p in model.parameters():
                p.requires_grad_(id(p) in keep)
        doms = tuple(kd.get("domains", ()))
        assert len(teachers) <= 1 or len(teachers) == len(doms), "per-domain experts need one teacher per domain"
        model.criterion = KDLoss(
            model, teachers, getattr(model, "kd_proj", None), feats, kd.get("pa_layers", ()), kd.get("T", 2.0), kd.get("w_kd", 1.0),
            kd.get("w_pa", 1.0), router, kd.get("w_rt", 0.1), doms,
            route_ce=router is not None and router.fc.out_features >= len(doms),  # K=1: plain LoRA, no router CE
        )
        trainer.ema.ema.criterion = KDLoss(trainer.ema.ema)  # val loss needs the same 6 items; stripped in save_model

    def save_model(self):
        ema = self.ema.ema
        c = ema.__dict__.pop("criterion", None)  # keep checkpoints loadable without gridkd.KDLoss
        try:
            super().save_model()
        finally:
            ema.criterion = c


# ----------------------------------------------------------------------------- CLI
def _overrides(pairs):
    out = {}
    for s in pairs:
        k, v = s.split("=", 1)
        try:
            v = eval(v, {}, {})  # numbers / bools / lists
        except Exception:
            pass
        out[k] = v
    return out


def load_plain(weights):
    from ultralytics import YOLO

    y = YOLO(weights)
    if hasattr(y.model, "kd_proj"):
        del y.model.kd_proj  # training-only branch
    return y


def cmd_train(a):
    kd = dict(teacher=a.teacher, T=a.T, w_kd=a.w_kd, w_pa=a.w_pa, pa_layers=[int(i) for i in a.pa_layers.split(",")],
              lora_r=a.lora_r, lora_k=a.lora_k, lora_alpha=a.lora_alpha, w_rt=a.w_rt,
              domains=[d for d in a.domains.split(",") if d])
    ov = {**dict(model=a.model, data=a.data, workers=4, exist_ok=True), **_overrides(a.rest)}  # CLI overrides win
    KDTrainer(kd=kd, overrides=ov).train()


def cmd_eval(a):
    rows = []
    for w in a.weights:
        y = load_plain(w)
        router = find_router(y.model)
        if router is not None and a.gates:
            router.fixed = torch.tensor([float(x) for x in a.gates.split(",")])
        n, gf = sum(p.numel() for p in y.model.parameters()) / 1e6, get_flops(y.model, a.imgsz)
        tag = Path(w).parent.parent.name if Path(w).parent.name == "weights" else Path(w).stem
        sws = [s for s in y.model.modules() if isinstance(s, (DensitySwitch, BNSwitch))]
        for d in a.data:
            for s in sws:
                s.hits = s.seen = 0.0
            m = y.val(data=d, split=a.split, imgsz=a.imgsz, batch=a.batch, device=a.device, workers=4, plots=False, verbose=False)
            on = sum(s.hits for s in sws) / max(sum(s.seen for s in sws), 1) if sws else None  # share of (image, layer) switches on
            rows.append((tag, Path(d).stem, n, gf, m.box.map50, m.box.map, on))
    print(f"\n{'model':24s} {'domain':10s} {'params(M)':>9s} {'GFLOPs':>7s} {'mAP50':>7s} {'mAP50-95':>8s} {'on-rate':>7s}")
    for r in rows:
        print(f"{r[0]:24s} {r[1]:10s} {r[2]:9.2f} {r[3]:7.1f} {r[4]:7.3f} {r[5]:8.3f} {'' if r[6] is None else f'{r[6]:7.3f}'}")
    if a.out:
        Path(a.out).write_text("model,domain,params_M,gflops,map50,map50_95,on_rate\n" + "".join(
            f"{r[0]},{r[1]},{r[2]:.3f},{r[3]:.2f},{r[4]:.4f},{r[5]:.4f},{'' if r[6] is None else f'{r[6]:.4f}'}\n" for r in rows))


class C2fSplit(nn.Module):
    """C2f/C3k2 with cv1 cut into two convs (exact): chunk+cat of one tensor breaks DepGraph's channel indexing."""

    def __init__(self, m):
        super().__init__()
        self.cv0, self.cv1 = (_conv_slice(m.cv1, sl) for sl in (slice(0, m.c), slice(m.c, 2 * m.c)))
        self.m, self.cv2 = m.m, m.cv2
        _copy_layer_attrs(self, m)

    def forward(self, x):
        y = [self.cv0(x), self.cv1(x)]
        y.extend(m(y[-1]) for m in self.m)
        return self.cv2(torch.cat(y, 1))


def _conv_slice(conv, sl):  # Conv keeping output channels sl
    c = deepcopy(conv)
    c.conv.weight = nn.Parameter(conv.conv.weight[sl].clone())
    c.conv.out_channels = c.conv.weight.shape[0]
    for k in ("weight", "bias"):
        setattr(c.bn, k, nn.Parameter(getattr(conv.bn, k)[sl].clone()))
    c.bn.running_mean, c.bn.running_var = conv.bn.running_mean[sl].clone(), conv.bn.running_var[sl].clone()
    c.bn.num_features = c.conv.out_channels
    return c


def cmd_prune(a):
    """DepGraph structured channel pruning (L2 group magnitude). C2PSA attention and the head's output convs are kept."""
    import torch_pruning as tp

    y = load_plain(a.weights)
    model, det = y.model.float(), y.model.model[-1]
    keep = [det.dfl, *(b[-1] for b in (*det.cv2, *det.cv3)), *(m for m in model.model if type(m).__name__ == "C2PSA")]
    model.requires_grad_(True).eval()
    x = torch.randn(1, 3, a.imgsz, a.imgsz)
    with torch.no_grad():
        ref = torch.cat([t.flatten() for t in model(x)[1]])
    for mod in list(model.modules()):
        for name, child in list(mod.named_children()):
            if isinstance(child, C2f):  # C3k2 included
                setattr(mod, name, C2fSplit(child))
    with torch.no_grad():
        err = (torch.cat([t.flatten() for t in model(x)[1]]) - ref).abs().max().item()
    assert err < 1e-4, f"C2f split changed outputs by {err}"
    det.training = True  # BN frozen, Detect returns raw maps for tracing
    n0, f0 = sum(p.numel() for p in model.parameters()), get_flops(model, a.imgsz)
    tp.pruner.MetaPruner(model, x, tp.importance.GroupMagnitudeImportance(p=2), pruning_ratio=a.ratio, ignored_layers=keep, round_to=8).step()
    model.eval()
    with torch.no_grad():
        model(x)  # still runs end to end
    model.pruned = True
    n1, f1 = sum(p.numel() for p in model.parameters()), get_flops(model, a.imgsz)
    torch.save({**y.ckpt, "model": model, "ema": None}, a.out)
    print(f"pruned {a.ratio:.0%} channels: params {n0 / 1e6:.2f}M -> {n1 / 1e6:.2f}M, GFLOPs {f0:.1f} -> {f1:.1f}, saved {a.out}")


@torch.no_grad()
def domain_gates(y, data, device):
    """Mean router gates over one domain's val images = the fixed gates of that site's merged model."""
    stem, g = next(m for m in y.model.modules() if isinstance(m, RoutedStem)), []

    def hook(m, a, o):
        if a[0].any():  # skip the validator's all-zero warmup image
            g.append(m.router.gates.float().cpu())

    h = stem.register_forward_hook(hook)
    y.val(data=data, split="val", device=device, workers=4, plots=False, verbose=False)
    h.remove()
    return torch.cat(g).mean(0)


def cmd_merge(a):
    y = load_plain(a.weights)
    if a.gates.endswith(".yaml"):
        g = domain_gates(y, a.gates, a.device)
        y = load_plain(a.weights)  # val left inference-mode (fused) tensors behind; merge a fresh copy
    else:
        g = [float(x) for x in a.gates.split(",")]
    print("gates", [round(float(v), 3) for v in g])
    merge_lora(y.model, g)
    torch.save({**y.ckpt, "model": y.model.float(), "ema": None}, a.out)  # fp32: BN-fused weights overflow fp16 (YOLO.save halves)
    print("saved", a.out, sum(p.numel() for p in y.model.parameters()) / 1e6, "M params")


@torch.no_grad()
def layer_stats(y, data, split, device):
    """Style statistics of every LoRA conv's input over one split, through the validator's own preprocessing."""
    mods = [m for m in y.model.modules() if isinstance(m, MultiLoRAConv)]
    out, live = [[] for _ in mods], [False]  # only inside the validator's batch loop: skips warmup and the 32x32 FLOPs probe

    def grab(i):
        def f(m, a):
            if live[0]:
                out[i].append(style_stats(a[0]).cpu())
        return f

    def start(_):
        live[0] = True

    def end(_):
        live[0] = False

    y.add_callback("on_val_batch_start", start)
    y.add_callback("on_val_batch_end", end)
    hs = [m.register_forward_pre_hook(grab(i)) for i, m in enumerate(mods)]
    try:
        y.val(data=data, split=split, device=device, workers=4, plots=False, verbose=False)
    finally:
        for h in hs:
            h.remove()
        y.callbacks["on_val_batch_start"].remove(start)
        y.callbacks["on_val_batch_end"].remove(end)
    out = [torch.cat(o) for o in out]
    assert all(o.isfinite().all() for o in out), "non-finite style statistics"
    return mods, out


def cmd_lsl(a):
    """Fit local-support switches for one domain adapter: pos = its own training split (adapter on, as in training),
    neg = generic images through the base model (adapter off). No data from earlier domains is used."""
    y = load_plain(a.weights)
    mods, pos = layer_stats(y, a.pos, "train", a.device)
    for m in mods:
        m.switch = ConstSwitch(0.0)
    _, neg = layer_stats(y, a.neg, "val", a.device)
    dev = "cuda" if torch.cuda.is_available() else "cpu"
    y = load_plain(a.weights)  # val left inference-mode (fused) tensors behind; attach switches to a fresh copy
    mods = [m for m in y.model.modules() if isinstance(m, MultiLoRAConv)]
    tpr, tnr = [], []
    for m, p, q in zip(mods, pos, neg):
        p, q = p.to(dev), q.to(dev)
        s = DensitySwitch(DiagGMM.fit(p, a.k), DiagGMM.fit(q, a.k_neg)).to(dev)
        tpr.append(float((s.pos.logp(p) > s.neg.logp(p)).float().mean()))
        tnr.append(float((s.pos.logp(q) <= s.neg.logp(q)).float().mean()))
        m.switch = s.cpu()
    print(f"{len(mods)} switches | {len(pos[0])} own images, {len(neg[0])} generic | fit-set on-rate {sum(tpr) / len(tpr):.3f}, off-rate on generic {sum(tnr) / len(tnr):.3f}")
    torch.save({**y.ckpt, "model": y.model.float(), "ema": None}, a.out)
    print("saved", a.out)


def cmd_bnsw(a):
    """BN switches for an adapter trained on the new domain only, against its base model: no data needed.
    adabn / srcbn are attribution controls: adapted BN statistics without LoRA / LoRA with the base model's statistics."""
    y = load_plain(a.weights)
    base = load_plain(a.base).model.float()
    if a.mode == "adabn":
        for m in y.model.modules():
            if isinstance(m, MultiLoRAConv):
                m.switch = ConstSwitch(0.0)
        n = "LoRA removed, adapted BN statistics kept"
    elif a.mode == "srcbn":
        ref = dict(base.named_modules())
        bns = [(x, ref[k.replace(".stem.bn", ".bn").replace(".m.bn", ".bn")]) for k, x in y.model.named_modules() if isinstance(x, nn.BatchNorm2d)]
        for x, r in bns:
            x.running_mean.copy_(r.running_mean), x.running_var.copy_(r.running_var)
        n = f"LoRA kept, base BN statistics on all {len(bns)} BN layers"
    else:
        n = f"{attach_bn_switches(y.model, base, a.mode == 'cum')} BN switches ({a.mode})"
    torch.save({**y.ckpt, "model": y.model.float(), "ema": None}, a.out)
    print(f"{n} | saved {a.out}")


def selftest():
    """LoRA merge must be exact; KD loss must run forward/backward. No downloads needed."""
    torch.manual_seed(0)
    s = DetectionModel("yolo11n.yaml", nc=2, verbose=False)
    router = inject_lora(s, r=4, k=3)
    s.eval()
    x = torch.rand(2, 3, 128, 128)
    neck = []  # compare head inputs: a random model's head output is just its bias
    s.model[-1].register_forward_pre_hook(lambda m, a: neck.append(torch.cat([t.flatten(1) for t in a[0]], 1)))
    router.fixed = torch.tensor([0.2, 0.5, 0.3])
    s(x)  # B = 0 -> base model
    for m in s.modules():
        if isinstance(m, MultiLoRAConv):
            nn.init.normal_(m.B.weight, std=0.1)
    s(x)
    merge_lora(s, router.fixed)(x)
    base, y0, y1 = neck
    assert not any(isinstance(m, (MultiLoRAConv, RoutedStem)) for m in s.modules())
    change, err = (y0 - base).abs().max(), (y0 - y1).abs().max()
    assert change > 0.05 * base.abs().max() and err < 1e-3 * change, (change, err)

    from types import SimpleNamespace

    st = DetectionModel("yolo11n.yaml", nc=2, verbose=False)
    st.args = SimpleNamespace(box=7.5, cls=0.5, dfl=1.5)
    te = DetectionModel("yolo11s.yaml", nc=2, verbose=False).eval()
    feats = {}
    for i in (4, 6, 10):
        st.model[i].register_forward_hook(lambda m, a, o, i=i: feats.__setitem__(i, o))
    proj = nn.ModuleList(nn.Sequential(Conv(c, c, 1), nn.Conv2d(c, 2, 1)) for c in (128, 128, 256))
    crit = KDLoss(st, te, proj, feats, (4, 6, 10))
    batch = dict(img=torch.rand(2, 3, 128, 128), batch_idx=torch.tensor([0.0, 1.0]), cls=torch.tensor([[0.0], [1.0]]),
                 bboxes=torch.tensor([[0.5, 0.5, 0.2, 0.3], [0.3, 0.6, 0.1, 0.2]]), im_file=["a", "b"])
    st.train()
    loss, items = crit(st(batch["img"]), batch)
    loss.sum().backward()
    assert items.shape == (6,) and torch.isfinite(items).all() and items[3] > 0 and items[4] > 0, items
    assert all(p.grad is not None for p in proj.parameters())
    # per-domain experts: two copies of one teacher must reproduce single-teacher KD exactly
    two = KDLoss(st, [te, te], proj, feats, (4, 6, 10), domains=("a", "b"))
    with torch.no_grad():
        preds = st(batch["img"])
        kd1, kd2 = crit(preds, batch)[1][3:5], two(preds, batch)[1][3:5]
    assert torch.allclose(kd1, kd2, atol=1e-5), (kd1, kd2)
    # local-support switch: the likelihood ratio separates two domains; a closed switch gives exactly the base model
    ga, gb = DiagGMM.fit(torch.randn(500, 6), 2), DiagGMM.fit(torch.randn(500, 6) + 3, 2)
    za, zb = torch.randn(200, 6), torch.randn(200, 6) + 3
    acc = float(((ga.logp(za) > gb.logp(za)).float().mean() + (gb.logp(zb) > ga.logp(zb)).float().mean()) / 2)
    assert acc > 0.97, acc
    s3 = DetectionModel("yolo11n.yaml", nc=2, verbose=False)
    inject_lora(s3, r=4, k=1)
    s3.eval()
    lo, hs = [m for m in s3.modules() if isinstance(m, MultiLoRAConv)], []
    s3.model[-1].register_forward_pre_hook(lambda m, a: hs.append(torch.cat([t.flatten(1) for t in a[0]], 1)))
    with torch.no_grad():
        s3(x)
        for m in lo:
            nn.init.normal_(m.B.weight, std=0.1)
            m.switch = ConstSwitch(0.0)
        s3(x)
        for m in lo:
            m.switch = ConstSwitch(1.0)
        s3(x)
    b0, off, on = hs
    assert torch.equal(b0, off) and (on - b0).abs().max() > 0.05 * b0.abs().max()

    def neck(model):
        h = []
        k = model.model[-1].register_forward_pre_hook(lambda m, a: h.append(torch.cat([t.flatten(1) for t in a[0]], 1)))
        with torch.no_grad():
            model(x)
        k.remove()
        return h[0]

    # BN switch: closed it must be exactly the base model (source BN restored everywhere), open the adapted one
    base = DetectionModel("yolo11n.yaml", nc=2, verbose=False).eval()
    for v in base.modules():
        if isinstance(v, nn.BatchNorm2d):
            v.running_mean.normal_(0, 0.5), v.running_var.uniform_(0.5, 2)
    ad = deepcopy(base)
    inject_lora(ad, r=4, k=1)
    ad.eval()
    for v in ad.modules():
        if isinstance(v, nn.BatchNorm2d):
            v.running_mean.add_(torch.randn_like(v.running_mean)), v.running_var.mul_(2)
        if isinstance(v, MultiLoRAConv):
            nn.init.normal_(v.B.weight, std=0.1)
    want_on, want_off = neck(ad), neck(base)
    nsw = attach_bn_switches(ad, base)
    bsw = [v for v in ad.modules() if isinstance(v, BNSwitch)]
    for v in bsw:
        v.forward = lambda x, y: y.new_zeros(len(y))
    got_off = neck(ad)
    for v in bsw:
        v.forward = lambda x, y: y.new_ones(len(y))
    got_on = neck(ad)
    e_off, e_on = ((g - w).abs().max() / w.abs().max() for g, w in ((got_off, want_off), (got_on, want_on)))
    assert e_off < 1e-4 and e_on < 1e-5 and (want_on - want_off).abs().max() > 0.05 * want_off.abs().max(), (e_off, e_on)
    print(f"selftest ok | LoRA change {float(change):.3g}, merge err {float(err):.3g} | switch acc {acc:.3f} | "
          f"{nsw} BN switches, closed err {float(e_off):.2g}, open err {float(e_on):.2g} | loss items", [round(float(v), 4) for v in items])


def main():
    p = argparse.ArgumentParser()
    sp = p.add_subparsers(dest="cmd", required=True)
    t = sp.add_parser("train")
    t.add_argument("--data", required=True)
    t.add_argument("--model", default="yolo11n.pt")
    t.add_argument("--teacher", default="", help="weights, or comma list = one expert per --domains entry")
    t.add_argument("--T", type=float, default=2.0)
    t.add_argument("--w-kd", type=float, default=1.0)
    t.add_argument("--w-pa", type=float, default=1.0)
    t.add_argument("--pa-layers", default="4,6,10")  # YOLO11 backbone P3/P4/P5
    t.add_argument("--lora-r", type=int, default=0)
    t.add_argument("--lora-k", type=int, default=4)
    t.add_argument("--lora-alpha", type=float, default=8)
    t.add_argument("--w-rt", type=float, default=0.1)
    t.add_argument("--domains", default="")  # path substrings -> router labels
    t.add_argument("rest", nargs="*", help="ultralytics overrides key=value")
    e = sp.add_parser("eval")
    e.add_argument("--weights", nargs="+", required=True)
    e.add_argument("--data", nargs="+", required=True)
    e.add_argument("--split", default="test")
    e.add_argument("--gates", default="")
    e.add_argument("--imgsz", type=int, default=640)
    e.add_argument("--batch", type=int, default=32)
    e.add_argument("--device", default="0")
    e.add_argument("--out", default="", help="also write the table as CSV")
    m = sp.add_parser("merge")
    m.add_argument("--weights", required=True)
    m.add_argument("--gates", required=True, help="comma list, or a domain data.yaml -> mean gates on its val split")
    m.add_argument("--out", required=True)
    m.add_argument("--device", default="0")
    r = sp.add_parser("prune")
    r.add_argument("--weights", required=True)
    r.add_argument("--ratio", type=float, default=0.3, help="fraction of channels removed per layer group")
    r.add_argument("--out", required=True)
    r.add_argument("--imgsz", type=int, default=640)
    s = sp.add_parser("lsl")
    s.add_argument("--weights", required=True, help="LoRA run trained on one domain only")
    s.add_argument("--pos", required=True, help="that domain's data.yaml; its train split fits the on-density")
    s.add_argument("--neg", required=True, help="generic images data.yaml (val split), e.g. ref_coco128/ref_1c.yaml")
    s.add_argument("--k", type=int, default=4, help="GMM components, own domain")
    s.add_argument("--k-neg", type=int, default=1, help="GMM components, generic images")
    s.add_argument("--out", required=True)
    s.add_argument("--device", default="0")
    s = sp.add_parser("bnsw")
    s.add_argument("--weights", required=True, help="LoRA run trained on one domain only")
    s.add_argument("--base", required=True, help="the model it was adapted from (its BN statistics describe the old domain)")
    s.add_argument("--mode", choices=("cum", "local", "adabn", "srcbn"), default="cum",
                   help="decide on the ratio summed over layers so far, or per layer; adabn/srcbn: attribution controls")
    s.add_argument("--out", required=True)
    sp.add_parser("selftest")
    a = p.parse_args()
    {"train": cmd_train, "eval": cmd_eval, "merge": cmd_merge, "prune": cmd_prune, "lsl": cmd_lsl, "bnsw": cmd_bnsw, "selftest": lambda _: selftest()}[a.cmd](a)


if __name__ == "__main__":
    sys.path.insert(0, str(Path(__file__).parent))
    __import__("gridkd").main()  # classes pickle as gridkd.*, not __main__.*