Skip to content

cli

sleap_nn.export.cli

CLI entry points for export workflows.

Functions:

Name Description
export

Export trained models to ONNX/TensorRT formats.

export(model_paths, output, fmt, opset_version, max_instances, max_batch_size, input_scale, input_height, input_width, crop_size, max_peaks_per_node, n_line_points, max_edge_length_ratio, dist_penalty_weight, device, precision, peak_threshold, workspace_size_gb, verify)

Export trained models to ONNX/TensorRT formats.

Source code in sleap_nn/export/cli.py
 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
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
@click.command(context_settings=CONTEXT_SETTINGS)
@click.argument(
    "model_paths",
    nargs=-1,
    type=click.Path(exists=True, file_okay=False, path_type=Path),
)
@click.option(
    "--output",
    "-o",
    type=click.Path(file_okay=False, path_type=Path),
    default=None,
    help="Output directory for exported model files.",
)
@click.option(
    "--format",
    "-f",
    "fmt",
    type=click.Choice(["onnx", "tensorrt", "both"], case_sensitive=False),
    default="onnx",
    show_default=True,
)
@click.option("--opset-version", type=int, default=17, show_default=True)
@click.option("--max-instances", type=int, default=20, show_default=True)
@click.option("--max-batch-size", type=int, default=8, show_default=True)
@click.option("--input-scale", type=float, default=None)
@click.option("--input-height", type=int, default=None)
@click.option("--input-width", type=int, default=None)
@click.option("--crop-size", type=int, default=None)
@click.option("--max-peaks-per-node", type=int, default=20, show_default=True)
@click.option("--n-line-points", type=int, default=10, show_default=True)
@click.option("--max-edge-length-ratio", type=float, default=0.25, show_default=True)
@click.option("--dist-penalty-weight", type=float, default=1.0, show_default=True)
@click.option("--device", type=str, default="cpu", show_default=True)
@click.option(
    "--precision",
    type=click.Choice(["fp32", "fp16", "tf32"], case_sensitive=False),
    default="fp16",
    show_default=True,
    help="TensorRT precision mode.",
)
@click.option(
    "--peak-threshold",
    type=float,
    default=0.2,
    show_default=True,
    help="Minimum confidence threshold for peak detection (baked into ONNX graph).",
)
@click.option(
    "--workspace-size-gb",
    type=float,
    default=None,
    help="TensorRT builder workspace size in GB. Uses exporter default if unset.",
)
@click.option("--verify/--no-verify", default=True, show_default=True)
def export(
    model_paths: tuple[Path, ...],
    output: Optional[Path],
    fmt: str,
    opset_version: int,
    max_instances: int,
    max_batch_size: int,
    input_scale: Optional[float],
    input_height: Optional[int],
    input_width: Optional[int],
    crop_size: Optional[int],
    max_peaks_per_node: int,
    n_line_points: int,
    max_edge_length_ratio: float,
    dist_penalty_weight: float,
    device: str,
    precision: str,
    peak_threshold: float,
    workspace_size_gb: Optional[float],
    verify: bool,
) -> None:
    """Export trained models to ONNX/TensorRT formats."""
    import torch

    from sleap_nn.export.exporters import export_to_onnx, export_to_tensorrt
    from sleap_nn.export.metadata import (
        build_base_metadata,
        embed_metadata_in_onnx,
        hash_file,
    )
    from sleap_nn.export.utils import (
        load_training_config,
        warn_on_tiled_export,
        resolve_anchor_part,
        resolve_backbone_source,
        resolve_centroid_method,
        resolve_backbone_type,
        resolve_background_fill,
        resolve_burn_in,
        resolve_class_maps_output_stride,
        resolve_class_names,
        resolve_crop_size,
        resolve_edge_inds,
        resolve_embedding_dim,
        resolve_embedding_input_channels,
        resolve_input_channels,
        resolve_input_scale,
        resolve_input_shape,
        resolve_model_type,
        resolve_n_classes,
        resolve_node_names,
        resolve_normalize,
        resolve_output_stride,
        resolve_pafs_output_stride,
    )
    from sleap_nn.export.wrappers import (
        BottomUpMultiClassONNXWrapper,
        BottomUpONNXWrapper,
        CenteredInstanceONNXWrapper,
        CentroidONNXWrapper,
        EmbeddingONNXWrapper,
        SingleInstanceONNXWrapper,
        TopDownMultiClassCombinedONNXWrapper,
        TopDownMultiClassONNXWrapper,
        TopDownONNXWrapper,
    )

    fmt = fmt.lower()

    trt_workspace_kwargs = (
        {"workspace_size": int(workspace_size_gb * (1 << 30))}
        if workspace_size_gb is not None
        else {}
    )

    if not model_paths:
        raise click.ClickException("Provide at least one model path to export.")

    model_paths = list(model_paths)
    cfgs = [load_training_config(path) for path in model_paths]
    for cfg in cfgs:
        warn_on_tiled_export(cfg)
    model_types = [resolve_model_type(cfg) for cfg in cfgs]
    backbone_types = [resolve_backbone_type(cfg) for cfg in cfgs]

    # A standalone centroid model is exported from a SINGLE directory; passing two
    # centroid dirs is almost always a mistake (the user likely meant centroid +
    # centered_instance for a top-down bundle). Catch it with a clear message
    # rather than letting it fall through to the generic combination error.
    if len(model_paths) == 2 and all(mt == "centroid" for mt in model_types):
        raise click.ClickException(
            "Received two centroid model directories. A standalone centroid model "
            "is exported from a single directory; a top-down bundle pairs a centroid "
            "directory with a centered_instance (or multi_class_topdown) directory. "
            "Pass one centroid directory for a standalone centroid export."
        )

    if len(model_paths) == 1:
        model_path = model_paths[0]
        cfg = cfgs[0]
        model_type = model_types[0]
        backbone_type = backbone_types[0]

        if model_type not in (
            "centroid",
            "centered_instance",
            "bottomup",
            "single_instance",
            "multi_class_topdown",
            "multi_class_bottomup",
            "embedding",
        ):
            raise click.ClickException(
                f"Model type '{model_type}' is not supported for export yet."
            )

        ckpt_path = model_path / "best.ckpt"
        if not ckpt_path.exists():
            raise click.ClickException(f"Checkpoint not found: {ckpt_path}")

        lightning_model = _load_lightning_model(
            model_type=model_type,
            backbone_type=backbone_type,
            cfg=cfg,
            ckpt_path=ckpt_path,
            device=device,
        )

        torch_model = lightning_model.model
        torch_model.eval()
        torch_model.to(device)

        export_dir = output or (model_path / "exported")
        export_dir.mkdir(parents=True, exist_ok=True)

        resolved_scale = (
            input_scale if input_scale is not None else resolve_input_scale(cfg)
        )
        output_stride = resolve_output_stride(cfg, model_type)
        resolved_crop_size = (
            (crop_size, crop_size) if crop_size is not None else resolve_crop_size(cfg)
        )
        metadata_max_instances = None
        metadata_max_peaks = None
        metadata_n_classes = None
        metadata_class_names = None
        metadata_embedding_dim = None
        metadata_normalize = None
        metadata_backbone_source = None
        metadata_burn_in = None
        metadata_background_fill = None
        metadata_normalization = "0_to_1"

        if model_type == "centroid":
            wrapper = CentroidONNXWrapper(
                torch_model,
                max_instances=max_instances,
                output_stride=output_stride,
                input_scale=resolved_scale,
                peak_threshold=peak_threshold,
            )
            output_names = ["centroids", "centroid_vals", "instance_valid"]
            metadata_max_instances = max_instances
            node_names = resolve_node_names(cfg, model_type)
            edge_inds = resolve_edge_inds(cfg, node_names)
        elif model_type == "centered_instance":
            wrapper = CenteredInstanceONNXWrapper(
                torch_model,
                output_stride=output_stride,
                input_scale=resolved_scale,
                peak_threshold=peak_threshold,
            )
            output_names = ["peaks", "peak_vals"]
            node_names = resolve_node_names(cfg, model_type)
            edge_inds = resolve_edge_inds(cfg, node_names)
        elif model_type == "bottomup":
            node_names = resolve_node_names(cfg, model_type)
            edge_inds = resolve_edge_inds(cfg, node_names)
            pafs_output_stride = resolve_pafs_output_stride(cfg)
            wrapper = BottomUpONNXWrapper(
                torch_model,
                skeleton_edges=edge_inds,
                n_nodes=len(node_names),
                max_peaks_per_node=max_peaks_per_node,
                n_line_points=n_line_points,
                cms_output_stride=output_stride,
                pafs_output_stride=pafs_output_stride,
                max_edge_length_ratio=max_edge_length_ratio,
                dist_penalty_weight=dist_penalty_weight,
                input_scale=resolved_scale,
                peak_threshold=peak_threshold,
            )
            output_names = [
                "peaks",
                "peak_vals",
                "peak_mask",
                "line_scores",
                "candidate_mask",
            ]
            metadata_max_peaks = max_peaks_per_node
        elif model_type == "single_instance":
            wrapper = SingleInstanceONNXWrapper(
                torch_model,
                output_stride=output_stride,
                input_scale=resolved_scale,
                peak_threshold=peak_threshold,
            )
            output_names = ["peaks", "peak_vals"]
            node_names = resolve_node_names(cfg, model_type)
            edge_inds = resolve_edge_inds(cfg, node_names)
        elif model_type == "multi_class_topdown":
            n_classes = resolve_n_classes(cfg, model_type)
            class_names = resolve_class_names(cfg, model_type)
            wrapper = TopDownMultiClassONNXWrapper(
                torch_model,
                output_stride=output_stride,
                input_scale=resolved_scale,
                n_classes=n_classes,
                peak_threshold=peak_threshold,
            )
            output_names = ["peaks", "peak_vals", "class_logits"]
            node_names = resolve_node_names(cfg, model_type)
            edge_inds = resolve_edge_inds(cfg, node_names)
            metadata_n_classes = n_classes
            metadata_class_names = class_names
        elif model_type == "multi_class_bottomup":
            node_names = resolve_node_names(cfg, model_type)
            edge_inds = resolve_edge_inds(cfg, node_names)
            n_classes = resolve_n_classes(cfg, model_type)
            class_names = resolve_class_names(cfg, model_type)
            class_maps_output_stride = resolve_class_maps_output_stride(cfg)
            wrapper = BottomUpMultiClassONNXWrapper(
                torch_model,
                n_nodes=len(node_names),
                n_classes=n_classes,
                max_peaks_per_node=max_peaks_per_node,
                cms_output_stride=output_stride,
                class_maps_output_stride=class_maps_output_stride,
                input_scale=resolved_scale,
                peak_threshold=peak_threshold,
            )
            output_names = ["peaks", "peak_vals", "peak_mask", "class_probs"]
            metadata_max_peaks = max_peaks_per_node
            metadata_n_classes = n_classes
            metadata_class_names = class_names
        elif model_type == "embedding":
            # Re-ID head: crop -> appearance vector. Simplest wrapper (single output,
            # no peak finding, no skeleton). Input is the grayscale crop the embedder
            # standardizes per-crop (the wrapper does not /255).
            wrapper = EmbeddingONNXWrapper(
                torch_model, normalize=resolve_normalize(cfg)
            )
            output_names = ["embedding"]
            # No skeleton semantics for an appearance model.
            node_names = []
            edge_inds = []
            metadata_embedding_dim = resolve_embedding_dim(cfg)
            metadata_normalize = resolve_normalize(cfg)
            metadata_backbone_source = resolve_backbone_source(cfg)
            metadata_burn_in = resolve_burn_in(cfg)
            metadata_background_fill = resolve_background_fill(cfg)
            metadata_normalization = "per_crop_standardize"
            if metadata_burn_in:
                # The single-input ONNX graph standardizes over the WHOLE crop; a
                # burn_in model's native inference standardizes over the foreground
                # (mask) only and fills the background. The exported embeddings will
                # therefore DIVERGE from native masked inference. Recorded in metadata.
                logger.warning(
                    "Exporting a mask-burn-in embedding model "
                    f"(background_fill='{metadata_background_fill}'): the ONNX graph "
                    "does a MASKLESS whole-crop standardize and cannot reproduce the "
                    "masked (foreground-only) standardize used at training/native "
                    "inference, so exported embeddings will diverge. Use the native "
                    "`sleap-nn predict ... --save_embeddings slp` path for exact "
                    "parity, or train/export a burn_in=False model for a faithful "
                    "single-input ONNX embedder."
                )
        else:
            raise click.ClickException(
                f"Model type '{model_type}' is not supported for export yet."
            )

        wrapper.eval()
        wrapper.to(device)

        input_shape = resolve_input_shape(
            cfg, input_height=input_height, input_width=input_width
        )
        input_channels = resolve_input_channels(cfg)
        if model_type == "embedding":
            # The embedder consumes a fixed-size grayscale CROP, not a full frame:
            # size the export input from the crop size + the (grayscale) data
            # channels. A 3ch ImageNet backbone repeats gray->3ch internally.
            if resolved_crop_size is None:
                raise click.ClickException(
                    "Embedding export requires a crop size. Provide --crop-size or "
                    "set data_config.preprocessing.crop_size."
                )
            crop_h, crop_w = resolved_crop_size
            input_channels = resolve_embedding_input_channels(cfg)
            input_shape = (1, input_channels, crop_h, crop_w)
        model_out_path = export_dir / "model.onnx"

        export_to_onnx(
            wrapper,
            model_out_path,
            input_shape=input_shape,
            input_dtype=torch.uint8,
            opset_version=opset_version,
            output_names=output_names,
            verify=verify,
        )

        training_config_path = _copy_training_config(model_path, export_dir, None)
        if training_config_path is not None:
            training_config_hash = hash_file(training_config_path)
            training_config_text = training_config_path.read_text()
        else:
            training_config_hash = ""
            training_config_text = None

        metadata = build_base_metadata(
            export_format="onnx",
            model_type=model_type,
            model_name=model_path.name,
            checkpoint_path=str(ckpt_path),
            backbone=backbone_type,
            n_nodes=len(node_names),
            n_edges=len(edge_inds),
            node_names=node_names,
            edge_inds=edge_inds,
            input_scale=resolved_scale,
            input_channels=input_channels,
            output_stride=output_stride,
            crop_size=resolved_crop_size,
            max_instances=metadata_max_instances,
            max_peaks_per_node=metadata_max_peaks,
            max_batch_size=max_batch_size,
            precision="fp32",
            training_config_hash=training_config_hash,
            training_config_embedded=training_config_text is not None,
            input_dtype="uint8",
            normalization=metadata_normalization,
            n_classes=metadata_n_classes,
            class_names=metadata_class_names,
            peak_threshold=peak_threshold,
            anchor_part=resolve_anchor_part(cfg, model_type),
            embedding_dim=metadata_embedding_dim,
            normalize=metadata_normalize,
            backbone_source=metadata_backbone_source,
            burn_in=metadata_burn_in,
            background_fill=metadata_background_fill,
            centroid_method=resolve_centroid_method(cfg, model_type),
        )

        metadata.save(export_dir / "export_metadata.json")

        if training_config_text is not None:
            try:
                embed_metadata_in_onnx(model_out_path, metadata, training_config_text)
            except ImportError:
                pass

        # Export to TensorRT if requested
        if fmt in ("tensorrt", "both"):
            trt_out_path = export_dir / "model.trt"
            B, C, H, W = input_shape

            # For centered_instance, single_instance, and embedding models, use the
            # crop size for TensorRT shape profiles since inference uses cropped inputs
            if model_type in ("centered_instance", "single_instance", "embedding"):
                if resolved_crop_size is not None:
                    crop_h, crop_w = resolved_crop_size
                    trt_input_shape = (1, C, crop_h, crop_w)
                    # Use crop size for min/opt, allow flexibility for max
                    trt_min_shape = (1, C, crop_h, crop_w)
                    trt_opt_shape = (1, C, crop_h, crop_w)
                    trt_max_shape = (max_batch_size, C, crop_h * 2, crop_w * 2)
                else:
                    trt_input_shape = input_shape
                    trt_min_shape = None
                    trt_opt_shape = None
                    trt_max_shape = (max_batch_size, C, H * 2, W * 2)
            else:
                trt_input_shape = input_shape
                trt_min_shape = None
                trt_opt_shape = None
                trt_max_shape = (max_batch_size, C, H * 2, W * 2)

            export_to_tensorrt(
                wrapper,
                trt_out_path,
                input_shape=trt_input_shape,
                input_dtype=torch.uint8,
                precision=precision,
                min_shape=trt_min_shape,
                opt_shape=trt_opt_shape,
                max_shape=trt_max_shape,
                **trt_workspace_kwargs,
                verbose=True,
            )
            # Update metadata for TensorRT
            trt_metadata = build_base_metadata(
                export_format="tensorrt",
                model_type=model_type,
                model_name=model_path.name,
                checkpoint_path=str(ckpt_path),
                backbone=backbone_type,
                n_nodes=len(node_names),
                n_edges=len(edge_inds),
                node_names=node_names,
                edge_inds=edge_inds,
                input_scale=resolved_scale,
                input_channels=input_channels,
                output_stride=output_stride,
                crop_size=resolved_crop_size,
                max_instances=metadata_max_instances,
                max_peaks_per_node=metadata_max_peaks,
                max_batch_size=max_batch_size,
                precision=precision,
                training_config_hash=training_config_hash,
                training_config_embedded=training_config_text is not None,
                input_dtype="uint8",
                normalization=metadata_normalization,
                n_classes=metadata_n_classes,
                class_names=metadata_class_names,
                peak_threshold=peak_threshold,
                anchor_part=resolve_anchor_part(cfg, model_type),
                embedding_dim=metadata_embedding_dim,
                normalize=metadata_normalize,
                backbone_source=metadata_backbone_source,
                burn_in=metadata_burn_in,
                background_fill=metadata_background_fill,
                centroid_method=resolve_centroid_method(cfg, model_type),
            )
            trt_metadata.save(export_dir / "model.trt.metadata.json")
        return

    if len(model_paths) == 2 and set(model_types) == {
        "centroid",
        "centered_instance",
    }:
        centroid_idx = model_types.index("centroid")
        instance_idx = model_types.index("centered_instance")

        centroid_path = model_paths[centroid_idx]
        instance_path = model_paths[instance_idx]
        centroid_cfg = cfgs[centroid_idx]
        instance_cfg = cfgs[instance_idx]
        centroid_backbone = backbone_types[centroid_idx]
        instance_backbone = backbone_types[instance_idx]

        centroid_ckpt = centroid_path / "best.ckpt"
        instance_ckpt = instance_path / "best.ckpt"
        if not centroid_ckpt.exists():
            raise click.ClickException(f"Checkpoint not found: {centroid_ckpt}")
        if not instance_ckpt.exists():
            raise click.ClickException(f"Checkpoint not found: {instance_ckpt}")

        centroid_model = _load_lightning_model(
            model_type="centroid",
            backbone_type=centroid_backbone,
            cfg=centroid_cfg,
            ckpt_path=centroid_ckpt,
            device=device,
        ).model
        instance_model = _load_lightning_model(
            model_type="centered_instance",
            backbone_type=instance_backbone,
            cfg=instance_cfg,
            ckpt_path=instance_ckpt,
            device=device,
        ).model

        centroid_model.eval()
        instance_model.eval()
        centroid_model.to(device)
        instance_model.to(device)

        export_dir = output or (centroid_path / "exported_topdown")
        export_dir.mkdir(parents=True, exist_ok=True)

        centroid_scale = (
            input_scale
            if input_scale is not None
            else resolve_input_scale(centroid_cfg)
        )
        instance_scale = (
            input_scale
            if input_scale is not None
            else resolve_input_scale(instance_cfg)
        )
        centroid_stride = resolve_output_stride(centroid_cfg, "centroid")
        instance_stride = resolve_output_stride(instance_cfg, "centered_instance")

        resolved_crop = resolve_crop_size(instance_cfg)
        if crop_size is not None:
            resolved_crop = (crop_size, crop_size)
        if resolved_crop is None:
            raise click.ClickException(
                "Top-down export requires crop_size. Provide --crop-size or ensure "
                "data_config.preprocessing.crop_size is set."
            )

        node_names = resolve_node_names(instance_cfg, "centered_instance")
        edge_inds = resolve_edge_inds(instance_cfg, node_names)

        wrapper = TopDownONNXWrapper(
            centroid_model=centroid_model,
            instance_model=instance_model,
            max_instances=max_instances,
            crop_size=resolved_crop,
            centroid_output_stride=centroid_stride,
            instance_output_stride=instance_stride,
            centroid_input_scale=centroid_scale,
            instance_input_scale=instance_scale,
            n_nodes=len(node_names),
            centroid_peak_threshold=peak_threshold,
            instance_peak_threshold=peak_threshold,
        )
        wrapper.eval()
        wrapper.to(device)

        input_shape = resolve_input_shape(
            centroid_cfg, input_height=input_height, input_width=input_width
        )
        model_out_path = export_dir / "model.onnx"

        export_to_onnx(
            wrapper,
            model_out_path,
            input_shape=input_shape,
            input_dtype=torch.uint8,
            opset_version=opset_version,
            output_names=[
                "centroids",
                "centroid_vals",
                "peaks",
                "peak_vals",
                "instance_valid",
            ],
            verify=verify,
        )

        centroid_cfg_path = _copy_training_config(centroid_path, export_dir, "centroid")
        instance_cfg_path = _copy_training_config(
            instance_path, export_dir, "centered_instance"
        )
        config_payload = {}
        config_hashes = []
        if centroid_cfg_path is not None:
            config_payload["centroid"] = centroid_cfg_path.read_text()
            config_hashes.append(f"centroid:{hash_file(centroid_cfg_path)}")
        if instance_cfg_path is not None:
            config_payload["centered_instance"] = instance_cfg_path.read_text()
            config_hashes.append(f"centered_instance:{hash_file(instance_cfg_path)}")

        training_config_hash = ";".join(config_hashes) if config_hashes else ""
        training_config_text = json.dumps(config_payload) if config_payload else None

        metadata = build_base_metadata(
            export_format="onnx",
            model_type="topdown",
            model_name=f"{centroid_path.name}+{instance_path.name}",
            checkpoint_path=(
                f"centroid:{centroid_ckpt};centered_instance:{instance_ckpt}"
            ),
            backbone=(
                f"centroid:{centroid_backbone};centered_instance:{instance_backbone}"
            ),
            n_nodes=len(node_names),
            n_edges=len(edge_inds),
            node_names=node_names,
            edge_inds=edge_inds,
            input_scale=centroid_scale,
            input_channels=resolve_input_channels(centroid_cfg),
            output_stride=instance_stride,
            crop_size=resolved_crop,
            max_instances=max_instances,
            max_batch_size=max_batch_size,
            precision="fp32",
            training_config_hash=training_config_hash,
            training_config_embedded=training_config_text is not None,
            input_dtype="uint8",
            normalization="0_to_1",
            peak_threshold=peak_threshold,
            anchor_part=resolve_anchor_part(centroid_cfg, "centroid"),
            centroid_method=resolve_centroid_method(centroid_cfg, "centroid"),
        )

        metadata.save(export_dir / "export_metadata.json")

        if training_config_text is not None:
            try:
                embed_metadata_in_onnx(model_out_path, metadata, training_config_text)
            except ImportError:
                pass

        # Export to TensorRT if requested
        if fmt in ("tensorrt", "both"):
            trt_out_path = export_dir / "model.trt"
            B, C, H, W = input_shape
            export_to_tensorrt(
                wrapper,
                trt_out_path,
                input_shape=input_shape,
                input_dtype=torch.uint8,
                precision=precision,
                max_shape=(max_batch_size, C, H * 2, W * 2),
                **trt_workspace_kwargs,
                verbose=True,
            )
            # Update metadata for TensorRT
            trt_metadata = build_base_metadata(
                export_format="tensorrt",
                model_type="topdown",
                model_name=f"{centroid_path.name}+{instance_path.name}",
                checkpoint_path=(
                    f"centroid:{centroid_ckpt};centered_instance:{instance_ckpt}"
                ),
                backbone=(
                    f"centroid:{centroid_backbone};centered_instance:{instance_backbone}"
                ),
                n_nodes=len(node_names),
                n_edges=len(edge_inds),
                node_names=node_names,
                edge_inds=edge_inds,
                input_scale=centroid_scale,
                input_channels=resolve_input_channels(centroid_cfg),
                output_stride=instance_stride,
                crop_size=resolved_crop,
                max_instances=max_instances,
                max_batch_size=max_batch_size,
                precision=precision,
                training_config_hash=training_config_hash,
                training_config_embedded=training_config_text is not None,
                input_dtype="uint8",
                normalization="0_to_1",
                peak_threshold=peak_threshold,
                anchor_part=resolve_anchor_part(centroid_cfg, "centroid"),
                centroid_method=resolve_centroid_method(centroid_cfg, "centroid"),
            )
            trt_metadata.save(export_dir / "model.trt.metadata.json")
        return

    # Combined multiclass top-down export (centroid + multi_class_topdown)
    if len(model_paths) == 2 and set(model_types) == {
        "centroid",
        "multi_class_topdown",
    }:
        centroid_idx = model_types.index("centroid")
        instance_idx = model_types.index("multi_class_topdown")

        centroid_path = model_paths[centroid_idx]
        instance_path = model_paths[instance_idx]
        centroid_cfg = cfgs[centroid_idx]
        instance_cfg = cfgs[instance_idx]
        centroid_backbone = backbone_types[centroid_idx]
        instance_backbone = backbone_types[instance_idx]

        centroid_ckpt = centroid_path / "best.ckpt"
        instance_ckpt = instance_path / "best.ckpt"
        if not centroid_ckpt.exists():
            raise click.ClickException(f"Checkpoint not found: {centroid_ckpt}")
        if not instance_ckpt.exists():
            raise click.ClickException(f"Checkpoint not found: {instance_ckpt}")

        centroid_model = _load_lightning_model(
            model_type="centroid",
            backbone_type=centroid_backbone,
            cfg=centroid_cfg,
            ckpt_path=centroid_ckpt,
            device=device,
        ).model
        instance_model = _load_lightning_model(
            model_type="multi_class_topdown",
            backbone_type=instance_backbone,
            cfg=instance_cfg,
            ckpt_path=instance_ckpt,
            device=device,
        ).model

        centroid_model.eval()
        instance_model.eval()
        centroid_model.to(device)
        instance_model.to(device)

        export_dir = output or (centroid_path / "exported_multi_class_topdown")
        export_dir.mkdir(parents=True, exist_ok=True)

        centroid_scale = (
            input_scale
            if input_scale is not None
            else resolve_input_scale(centroid_cfg)
        )
        instance_scale = (
            input_scale
            if input_scale is not None
            else resolve_input_scale(instance_cfg)
        )
        centroid_stride = resolve_output_stride(centroid_cfg, "centroid")
        instance_stride = resolve_output_stride(instance_cfg, "multi_class_topdown")

        resolved_crop = resolve_crop_size(instance_cfg)
        if crop_size is not None:
            resolved_crop = (crop_size, crop_size)
        if resolved_crop is None:
            raise click.ClickException(
                "Multiclass top-down export requires crop_size. Provide --crop-size or "
                "ensure data_config.preprocessing.crop_size is set."
            )

        node_names = resolve_node_names(instance_cfg, "multi_class_topdown")
        edge_inds = resolve_edge_inds(instance_cfg, node_names)
        n_classes = resolve_n_classes(instance_cfg, "multi_class_topdown")
        class_names = resolve_class_names(instance_cfg, "multi_class_topdown")

        wrapper = TopDownMultiClassCombinedONNXWrapper(
            centroid_model=centroid_model,
            instance_model=instance_model,
            max_instances=max_instances,
            crop_size=resolved_crop,
            centroid_output_stride=centroid_stride,
            instance_output_stride=instance_stride,
            centroid_input_scale=centroid_scale,
            instance_input_scale=instance_scale,
            n_nodes=len(node_names),
            n_classes=n_classes,
            centroid_peak_threshold=peak_threshold,
            instance_peak_threshold=peak_threshold,
        )
        wrapper.eval()
        wrapper.to(device)

        input_shape = resolve_input_shape(
            centroid_cfg, input_height=input_height, input_width=input_width
        )
        model_out_path = export_dir / "model.onnx"

        export_to_onnx(
            wrapper,
            model_out_path,
            input_shape=input_shape,
            input_dtype=torch.uint8,
            opset_version=opset_version,
            output_names=[
                "centroids",
                "centroid_vals",
                "peaks",
                "peak_vals",
                "class_logits",
                "instance_valid",
            ],
            verify=verify,
        )

        centroid_cfg_path = _copy_training_config(centroid_path, export_dir, "centroid")
        instance_cfg_path = _copy_training_config(
            instance_path, export_dir, "multi_class_topdown"
        )
        config_payload = {}
        config_hashes = []
        if centroid_cfg_path is not None:
            config_payload["centroid"] = centroid_cfg_path.read_text()
            config_hashes.append(f"centroid:{hash_file(centroid_cfg_path)}")
        if instance_cfg_path is not None:
            config_payload["multi_class_topdown"] = instance_cfg_path.read_text()
            config_hashes.append(f"multi_class_topdown:{hash_file(instance_cfg_path)}")

        training_config_hash = ";".join(config_hashes) if config_hashes else ""
        training_config_text = json.dumps(config_payload) if config_payload else None

        metadata = build_base_metadata(
            export_format="onnx",
            model_type="multi_class_topdown_combined",
            model_name=f"{centroid_path.name}+{instance_path.name}",
            checkpoint_path=(
                f"centroid:{centroid_ckpt};multi_class_topdown:{instance_ckpt}"
            ),
            backbone=(
                f"centroid:{centroid_backbone};multi_class_topdown:{instance_backbone}"
            ),
            n_nodes=len(node_names),
            n_edges=len(edge_inds),
            node_names=node_names,
            edge_inds=edge_inds,
            input_scale=centroid_scale,
            input_channels=resolve_input_channels(centroid_cfg),
            output_stride=instance_stride,
            crop_size=resolved_crop,
            max_instances=max_instances,
            max_batch_size=max_batch_size,
            training_config_hash=training_config_hash,
            training_config_embedded=training_config_text is not None,
            input_dtype="uint8",
            normalization="0_to_1",
            n_classes=n_classes,
            class_names=class_names,
            peak_threshold=peak_threshold,
            anchor_part=resolve_anchor_part(centroid_cfg, "centroid"),
            centroid_method=resolve_centroid_method(centroid_cfg, "centroid"),
        )
        metadata.save(export_dir / "export_metadata.json")
        click.echo(f"ONNX model exported to: {model_out_path}")
        click.echo(f"Metadata saved to: {export_dir / 'export_metadata.json'}")

        # TensorRT export for combined multiclass top-down
        if fmt in ("tensorrt", "both"):
            trt_out_path = export_dir / "model.trt"
            B, C, H, W = input_shape
            export_to_tensorrt(
                wrapper,
                trt_out_path,
                input_shape=input_shape,
                input_dtype=torch.uint8,
                precision=precision,
                max_shape=(max_batch_size, C, H * 2, W * 2),
                **trt_workspace_kwargs,
                verbose=True,
            )
            trt_metadata = build_base_metadata(
                export_format="tensorrt",
                model_type="multi_class_topdown_combined",
                model_name=f"{centroid_path.name}+{instance_path.name}",
                checkpoint_path=(
                    f"centroid:{centroid_ckpt};multi_class_topdown:{instance_ckpt}"
                ),
                backbone=(
                    f"centroid:{centroid_backbone};multi_class_topdown:{instance_backbone}"
                ),
                n_nodes=len(node_names),
                n_edges=len(edge_inds),
                node_names=node_names,
                edge_inds=edge_inds,
                input_scale=centroid_scale,
                input_channels=resolve_input_channels(centroid_cfg),
                output_stride=instance_stride,
                crop_size=resolved_crop,
                max_instances=max_instances,
                max_batch_size=max_batch_size,
                precision=precision,
                training_config_hash=training_config_hash,
                training_config_embedded=training_config_text is not None,
                input_dtype="uint8",
                normalization="0_to_1",
                n_classes=n_classes,
                class_names=class_names,
                peak_threshold=peak_threshold,
                anchor_part=resolve_anchor_part(centroid_cfg, "centroid"),
                centroid_method=resolve_centroid_method(centroid_cfg, "centroid"),
            )
            trt_metadata.save(export_dir / "model.trt.metadata.json")
        return

    raise click.ClickException(
        "Provide one model path for centroid/centered-instance/bottom-up export, "
        "or two paths (centroid + centered_instance or centroid + multi_class_topdown) "
        "for combined top-down export."
    )