File size: 21,995 Bytes
9e15541
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import sys
import math
from typing import Callable, Mapping

import skimage.metrics as sk_metrics
import torch
import torch.nn.functional as F
from ignite.engine import Engine
from ignite.exceptions import NotComputableError
from ignite.metrics import Metric
from ignite.metrics.metric import reinit__is_reduced, sync_all_reduce

import pulp


def median_scaling(
    depth_gt: torch.Tensor,
    depth_pred: torch.Tensor,
):
    # TODO: ensure this works for any batch size
    mask = depth_gt > 0

    depth_gt[mask] = torch.nan
    depth_pred[mask] = torch.nan
    scaling = torch.nanmedian(depth_gt.flatten(-2, -1), dim=-1) / torch.nanmedian(
        depth_pred.flatten(-2, -1), dim=-1
    )
    depth_pred = scaling[..., None, None] * depth_pred
    return depth_pred


def l2_scaling(
    depth_gt: torch.Tensor,
    depth_pred: torch.Tensor,
):
    # TODO: ensure this works for any batch size
    mask = depth_gt > 0
    depth_pred = depth_pred
    depth_gt_ = depth_gt[mask]
    depth_pred_ = depth_pred[mask]
    depth_pred_ = torch.stack((depth_pred_, torch.ones_like(depth_pred_)), dim=-1)
    x = torch.linalg.lstsq(
        depth_pred_.to(torch.float32), depth_gt_.unsqueeze(-1).to(torch.float32)
    ).solution.squeeze()
    depth_pred = depth_pred * x[0] + x[1]
    return depth_pred


def compute_depth_metrics(
    depth_gt: torch.Tensor,
    depth_pred: torch.Tensor,
    scaling_fn: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] | None,
):
    # TODO: find out if dim -3 is dummy dimension or part of the batch
    # TODO: Test if works for batches of images
    if scaling_fn:
        depth_pred = scaling_fn(depth_gt, depth_pred)

    depth_pred = torch.clamp(depth_pred, 1e-3, 80)
    mask = depth_gt != 0

    max_ratio = torch.maximum((depth_gt / depth_pred), (depth_pred / depth_gt))
    a_scores = {}
    for name, thresh in {"a1": 1.25, "a2": 1.25**2, "a3": 1.25**3}.items():
        within_thresh = (max_ratio < thresh).to(torch.float)
        within_thresh[~mask] = 0.0
        a_scores[name] = within_thresh.flatten(-2, -1).sum(dim=-1) / mask.to(
            torch.float
        ).flatten(-2, -1).sum(dim=-1)

    square_error = (depth_gt - depth_pred) ** 2
    square_error[~mask] = 0.0

    log_square_error = (torch.log(depth_gt) - torch.log(depth_pred)) ** 2
    log_square_error[~mask] = 0.0

    abs_error = torch.abs(depth_gt - depth_pred)
    abs_error[~mask] = 0.0

    rmse = (
        square_error.flatten(-2, -1).sum(dim=-1)
        / mask.to(torch.float).flatten(-2, -1).sum(dim=-1)
    ) ** 0.5

    rmse_log = (
        log_square_error.flatten(-2, -1).sum(dim=-1)
        / mask.to(torch.float).flatten(-2, -1).sum(dim=-1)
    ) ** 0.5

    abs_rel = abs_error / depth_gt
    abs_rel[~mask] = 0.0
    abs_rel = (
        abs_rel.flatten(-2, -1).sum(dim=-1)
        / mask.to(torch.float).flatten(-2, -1).sum(dim=-1)
    ) ** 0.5

    sq_rel = square_error / depth_gt
    sq_rel[~mask] = 0.0
    sq_rel = (
        sq_rel.flatten(-2, -1).sum(dim=-1)
        / mask.to(torch.float).flatten(-2, -1).sum(dim=-1)
    ) ** 0.5

    metrics_dict = {
        "abs_rel": abs_rel,
        "sq_rel": sq_rel,
        "rmse": rmse,
        "rmse_log": rmse_log,
        "a1": a_scores["a1"],
        "a2": a_scores["a2"],
        "a3": a_scores["a3"],
    }
    return metrics_dict


def compute_occ_metrics(
    occupancy_pred: torch.Tensor, occupancy_gt: torch.Tensor, is_visible: torch.Tensor
):
    # Only not visible points can be occupied
    occupancy_gt &= ~is_visible

    is_occupied_acc = (occupancy_pred == occupancy_gt).float().mean().item()
    is_occupied_prec = occupancy_gt[occupancy_pred].float().mean().item()
    is_occupied_rec = occupancy_pred[occupancy_gt].float().mean().item()

    not_occupied_not_visible_ratio = (
        ((~occupancy_gt) & (~is_visible)).float().mean().item()
    )

    total_ie = ((~occupancy_gt) & (~is_visible)).float().sum().item()

    ie_acc = (occupancy_pred == occupancy_gt)[(~is_visible)].float().mean().item()
    ie_prec = (~occupancy_gt)[(~occupancy_pred) & (~is_visible)].float().mean()
    ie_rec = (~occupancy_pred)[(~occupancy_gt) & (~is_visible)].float().mean()
    total_no_nop_nv = (
        ((~occupancy_gt) & (~occupancy_pred))[(~is_visible) & (~occupancy_gt)]
        .float()
        .sum()
    )

    return {
        "o_acc": is_occupied_acc,
        "o_rec": is_occupied_rec,
        "o_prec": is_occupied_prec,
        "ie_acc": ie_acc,
        "ie_rec": ie_rec,
        "ie_prec": ie_prec,
        "ie_r": not_occupied_not_visible_ratio,
        "t_ie": total_ie,
        "t_no_nop_nv": total_no_nop_nv,
    }


def compute_nvs_metrics(data, lpips):
    # TODO: This is only correct for batchsize 1!
    # Following tucker et al. and others, we crop 5% on all sides

    # idx of stereo frame (the target frame is always the "stereo" frame).
    sf_id = data["rgb_gt"].shape[1] // 2

    imgs_gt = data["rgb_gt"][:1, sf_id : sf_id + 1]
    imgs_pred = data["fine"][0]["rgb"][:1, sf_id : sf_id + 1]

    imgs_gt = imgs_gt.squeeze(0).permute(0, 3, 1, 2)
    imgs_pred = imgs_pred.squeeze(0).squeeze(-2).permute(0, 3, 1, 2)

    n, c, h, w = imgs_gt.shape
    y0 = int(math.ceil(0.05 * h))
    y1 = int(math.floor(0.95 * h))
    x0 = int(math.ceil(0.05 * w))
    x1 = int(math.floor(0.95 * w))

    imgs_gt = imgs_gt[:, :, y0:y1, x0:x1]
    imgs_pred = imgs_pred[:, :, y0:y1, x0:x1]

    imgs_gt_np = imgs_gt.detach().squeeze().permute(1, 2, 0).cpu().numpy()
    imgs_pred_np = imgs_pred.detach().squeeze().permute(1, 2, 0).cpu().numpy()

    ssim_score = sk_metrics.structural_similarity(
        imgs_pred_np, imgs_gt_np, multichannel=True, data_range=1, channel_axis=-1
    )
    psnr_score = sk_metrics.peak_signal_noise_ratio(
        imgs_pred_np, imgs_gt_np, data_range=1
    )
    lpips_score = lpips(imgs_pred, imgs_gt, normalize=False).mean()

    metrics_dict = {
        "ssim": torch.tensor([ssim_score], device=imgs_gt.device),
        "psnr": torch.tensor([psnr_score], device=imgs_gt.device),
        "lpips": torch.tensor([lpips_score], device=imgs_gt.device),
    }
    return metrics_dict


def compute_dino_metrics(data):
    dino_gt = data["dino_gt"]
    if "dino_features_downsampled" in data["coarse"][0]:
        dino_pred = data["coarse"][0]["dino_features_downsampled"].squeeze(-2)
    else:
        dino_pred = data["coarse"][0]["dino_features"].squeeze(-2)

    l1_loss = F.l1_loss(dino_pred, dino_gt, reduction="none").mean(dim=(0, 2, 3, 4))
    l2_loss = F.mse_loss(dino_pred, dino_gt, reduction="none").mean(dim=(0, 2, 3, 4))
    cos_sim = F.cosine_similarity(dino_pred, dino_gt, dim=-1).mean(dim=(0, 2, 3))

    metrics_dict = {
        "l1": torch.tensor([l1_loss.mean()], device=dino_gt.device),
        "l2": torch.tensor([l2_loss.mean()], device=dino_gt.device),
        "cos_sim": torch.tensor([cos_sim.mean()], device=dino_gt.device)
    }
    for i in range(len(l1_loss)):
        metrics_dict[f"l1_{i}"] = torch.tensor([l1_loss[i]], device=dino_gt.device)
        metrics_dict[f"l2_{i}"] = torch.tensor([l2_loss[i]], device=dino_gt.device)
        metrics_dict[f"cos_sim_{i}"] = torch.tensor([cos_sim[i]], device=dino_gt.device)
    return metrics_dict


def compute_stego_metrics(data):
    if "stego_corr" not in data["segmentation"]:
        return {}

    metrics_dict = {
        "stego_self_corr": data["segmentation"]["stego_corr"]["stego_self_corr"],
        "stego_nn_corr": data["segmentation"]["stego_corr"]["stego_nn_corr"],
        "stego_random_corr": data["segmentation"]["stego_corr"]["stego_random_corr"],
    }
    return metrics_dict


def compute_seg_metrics(data, n_classes, gt_classes):
    segs_gt = data["segmentation"]["target"].flatten()
    valid_mask = segs_gt >= 0
    segs_gt = segs_gt[valid_mask]

    metrics_dict = {}
    for result_key, result in data["segmentation"]["results"].items():
        if "pseudo_segs_pred" in result:
            segs_pred = result["pseudo_segs_pred"][:, 0].flatten()
        else:
            segs_pred = result["segs_pred"][:, 0].flatten()

        segs_pred = segs_pred[valid_mask]
        confusion_matrix = torch.bincount(n_classes * segs_gt + segs_pred,
                                          minlength=n_classes * gt_classes).reshape(gt_classes, n_classes)
        metrics_dict[result_key] = confusion_matrix
    
    return metrics_dict


class MeanMetric(Metric):
    def __init__(self, output_transform=lambda x: x["output"], device="cpu"):
        super(MeanMetric, self).__init__(
            output_transform=output_transform, device=device
        )
        self._sum = torch.tensor(0, device=self._device, dtype=torch.float32)
        self._num_examples = 0
        self.required_output_keys = ()

    @reinit__is_reduced
    def reset(self):
        self._sum = torch.tensor(0, device=self._device, dtype=torch.float32)
        self._num_examples = 0
        super(MeanMetric, self).reset()

    @reinit__is_reduced
    def update(self, value):
        if torch.any(torch.isnan(torch.tensor(value))):
            raise ValueError("NaN values present in metric!")
        self._sum += value
        self._num_examples += 1

    @sync_all_reduce("_num_examples:SUM", "_sum:SUM")
    def compute(self):
        if self._num_examples == 0:
            raise NotComputableError(
                "CustomAccuracy must have at least one example before it can be computed."
            )
        return self._sum.item() / self._num_examples

    @torch.no_grad()
    def iteration_completed(self, engine: Engine) -> None:
        output = self._output_transform(
            engine.state.output
        )  ## engine.state.output.keys() == dict_keys(['output', 'loss_dict', 'timings_dict', 'metrics_dict'])
        self.update(output)


class DictMeanMetric(Metric):
    def __init__(self, name: str, output_transform=lambda x: x["output"], device="cpu"):
        self._name = name
        self._sums: dict[str, torch.Tensor] = {}
        self._num_examples = 0
        self.required_output_keys = ()
        super(DictMeanMetric, self).__init__(
            output_transform=output_transform, device=device
        )

    @reinit__is_reduced
    def reset(self):
        self._sums = {}
        self._num_examples = 0
        super(DictMeanMetric, self).reset()

    @reinit__is_reduced
    def update(self, value):
        num_examples = None
        for key, metric in value.items():
            if not key in self._sums:
                self._sums[key] = torch.tensor(
                    0, device=self._device, dtype=torch.float32
                )
            if torch.any(torch.isnan(metric)):
                # TODO: integrate into logging
                print(f"Warining: Metric {self._name}/{key} has a nan value")
                continue
            self._sums[key] += metric.sum().to(self._device)
            # TODO: check if this works with batches
            if num_examples is None:
                num_examples = metric.shape[0]
        self._num_examples += 1

    @sync_all_reduce("_num_examples:SUM", "_sum:SUM")
    def compute(self):
        if self._num_examples == 0:
            raise NotComputableError(
                "CustomAccuracy must have at least one example before it can be computed."
            )
        return {
            f"{self._name}_{key}": metric.item() / self._num_examples
            for key, metric in self._sums.items()
        }

    @torch.no_grad()
    def iteration_completed(self, engine: Engine) -> None:
        output = self._output_transform(engine.state.output["output"])
        self.update(output)

    def completed(self, engine: Engine, name: str) -> None:
        """Helper method to compute metric's value and put into the engine. It is automatically attached to the
        `engine` with :meth:`~ignite.metrics.metric.Metric.attach`. If metrics' value is torch tensor, it is
        explicitly sent to CPU device.

        Args:
            engine: the engine to which the metric must be attached
            name: the name of the metric used as key in dict `engine.state.metrics`

        .. changes from default implementation:
            don't add whole result dict to engine state, but only the values

        """
        result = self.compute()
        if isinstance(result, Mapping):
            if name in result.keys():
                raise ValueError(
                    f"Argument name '{name}' is conflicting with mapping keys: {list(result.keys())}"
                )

            for key, value in result.items():
                engine.state.metrics[key] = value
        else:
            if isinstance(result, torch.Tensor):
                if len(result.size()) == 0:
                    result = result.item()
                elif "cpu" not in result.device.type:
                    result = result.cpu()

            engine.state.metrics[name] = result


class SegmentationMetric(DictMeanMetric):
    def __init__(self, name: str, output_transform=lambda x: x["output"], device="cpu", assign_pseudo=True):
        super(SegmentationMetric, self).__init__(
            name, output_transform, device
        )
        self.assign_pseudo = assign_pseudo

        # [road, sidewalk, building, wall, fence, pole, traffic light, traffic sign, vegetation, terrain, sky, person, rider, car, truck, bus, train, motorcycle, bicycle]
        self.weights = torch.Tensor([4, 2, 2, 1, 1, 1, 1, 1, 2, 2, 1, 1, 1, 2, 1, 1, 1, 1, 1])
        self.weights = self.weights / self.weights.mean()
    
    @reinit__is_reduced
    def update(self, value):
        for key, metric in value.items():
            if not key in self._sums:
                self._sums[key] = torch.zeros(metric.shape, device=self._device, dtype=torch.int32)
            if torch.any(torch.isnan(metric)):
                print(f"Warining: Metric {self._name}/{key} has a nan value")
                continue
            self._sums[key] += metric.to(self._device)
        self._num_examples += 1

    @sync_all_reduce("_num_examples:SUM", "_sum:SUM")
    def compute(self):
        if self._num_examples == 0:
            raise NotComputableError(
                "CustomAccuracy must have at least one example before it can be computed."
            )
        result = {}
        for key, _sum in self._sums.items():
            if self.assign_pseudo:
                assignment = self._calculate_pseudo_label_assignment(_sum)
                gt_classes = _sum.size(0)
                confusion_matrix = torch.zeros((gt_classes, gt_classes), dtype=_sum.dtype)
                confusion_matrix.scatter_add_(
                    1,
                    assignment.unsqueeze(0).expand(gt_classes, -1),
                    _sum
                )
                result[key + "_assignment"] = assignment
            else:
                confusion_matrix = _sum

            # confusion_matrix axes: (actual, prediction)

            true_positives = confusion_matrix.diag()
            false_negatives = torch.sum(confusion_matrix, dim=1) - true_positives
            false_positives = torch.sum(confusion_matrix, dim=0) - true_positives
            denominator = true_positives + false_positives + false_negatives
            per_class_iou = torch.where(denominator > 0, true_positives / denominator,
                                        torch.zeros_like(denominator))

            result[key + "_per_class_iou"] = per_class_iou
            result[key + "_miou"] = per_class_iou.mean().item()
            result[key + "_weighted_miou"] = (per_class_iou * self.weights).mean().item()
            result[key + "_acc"] = confusion_matrix.diag().sum().item() / confusion_matrix.sum().item()

            result[key + "_confusion_matrix"] = confusion_matrix

        return result

    def _calculate_pseudo_label_assignment(self, metric_matrix):
        """Implemented this way to generalize to over-segmentation"""
        gt_classes, n_classes = metric_matrix.size()
        costs = metric_matrix.cpu().numpy()

        problem = pulp.LpProblem("CapacitatedAssignment", pulp.LpMaximize)
        x = [[pulp.LpVariable(f"x_{i}_{j}", cat="Binary") for j in range(n_classes)] for i in
             range(gt_classes)]

        problem += pulp.lpSum(costs[i][j] * x[i][j] for i in range(gt_classes) for j in range(n_classes))
        for j in range(n_classes):
            problem += pulp.lpSum(x[i][j] for i in range(gt_classes)) == 1, f"AssignPseudoLabel_{j}"

        for i in range(gt_classes):
            problem += pulp.lpSum(x[i][j] for j in range(n_classes)) >= 1, f"MinAssignActualLabel_{i}"

        problem.solve()

        print("Status:", pulp.LpStatus[problem.status])
        print("Objective:", pulp.value(problem.objective))

        assignment = torch.zeros(n_classes, dtype=torch.int64)
        for j in range(n_classes):
            assignment[j] = next(i for i in range(gt_classes) if pulp.value(x[i][j]) == 1)

        return assignment


class ConcatenateMetric(DictMeanMetric):
    @reinit__is_reduced
    def update(self, value, every_nth=100):
        n_bins = 50
        for key, metric in value.items():
            if not key in self._sums:
                self._sums[key] = torch.zeros((n_bins,), device=self._device, dtype=torch.int32)
            if torch.any(torch.isnan(metric)):
                print(f"Warning: Metric {self._name}/{key} has a nan value")
                continue

            metric_flat = metric.flatten().to(self._device)[::every_nth]
            if key in self._sums:
                self._sums[key] = torch.cat([self._sums[key], metric_flat])
            else:
                self._sums[key] = metric_flat

        self._num_examples += 1

    @sync_all_reduce("_num_examples:SUM", "_sum:SUM")
    def compute(self):
        return self._sums


class FG_ARI(Metric):
    def __init__(self, output_transform=lambda x: x["output"], device="cpu"):
        self._sum_fg_aris = torch.tensor(0, device=self._device, dtype=torch.float32)
        self._num_examples = 0
        self.required_output_keys = ()
        super(FG_ARI, self).__init__(output_transform=output_transform, device=device)

    @reinit__is_reduced
    def reset(self):
        self._sum_fg_aris = torch.tensor(0, device=self._device, dtype=torch.float32)
        self._num_examples = 0
        super(FG_ARI, self).reset()

    @reinit__is_reduced
    def update(self, data):
        true_masks = data["segs"]  # fc [n, h, w]
        pred_masks = data["slot_masks"]  # n, fc, sc, h, w

        n, fc, sc, h, w = pred_masks.shape

        true_masks = [
            F.interpolate(tm.to(float).unsqueeze(1), (h, w), mode="nearest")
            .squeeze(1)
            .to(int)
            for tm in true_masks
        ]

        for i in range(n):
            for f in range(fc):
                true_mask = true_masks[f][i]
                pred_mask = pred_masks[i, f]

                true_mask = true_mask.view(-1)
                pred_mask = pred_mask.view(sc, -1)

                if torch.max(true_mask) == 0:
                    continue

                foreground = true_mask > 0
                true_mask = true_mask[foreground]
                pred_mask = pred_mask[:, foreground].permute(1, 0)

                true_mask = F.one_hot(true_mask)

                # Filter out empty true groups
                not_empty = torch.any(true_mask, dim=0)
                true_mask = true_mask[:, not_empty]

                # Filter out empty predicted groups
                not_empty = torch.any(pred_mask, dim=0)
                pred_mask = pred_mask[:, not_empty]

                true_mask.unsqueeze_(0)
                pred_mask.unsqueeze_(0)

                _, n_points, n_true_groups = true_mask.shape
                n_pred_groups = pred_mask.shape[-1]
                if n_points <= n_true_groups and n_points <= n_pred_groups:
                    print(
                        "adjusted_rand_index requires n_groups < n_points.",
                        file=sys.stderr,
                    )
                    continue

                true_group_ids = torch.argmax(true_mask, -1)
                pred_group_ids = torch.argmax(pred_mask, -1)
                true_mask_oh = true_mask.to(torch.float32)
                pred_mask_oh = F.one_hot(pred_group_ids, n_pred_groups).to(
                    torch.float32
                )

                n_points = torch.sum(true_mask_oh, dim=[1, 2]).to(torch.float32)

                nij = torch.einsum("bji,bjk->bki", pred_mask_oh, true_mask_oh)
                a = torch.sum(nij, dim=1)
                b = torch.sum(nij, dim=2)

                rindex = torch.sum(nij * (nij - 1), dim=[1, 2])
                aindex = torch.sum(a * (a - 1), dim=1)
                bindex = torch.sum(b * (b - 1), dim=1)
                expected_rindex = aindex * bindex / (n_points * (n_points - 1))
                max_rindex = (aindex + bindex) / 2
                ari = (rindex - expected_rindex) / (
                    max_rindex - expected_rindex + 0.000000000001
                )

                _all_equal = lambda values: torch.all(
                    torch.eq(values, values[..., :1]), dim=-1
                )
                both_single_cluster = torch.logical_and(
                    _all_equal(true_group_ids), _all_equal(pred_group_ids)
                )

                self._sum_fg_aris += torch.where(
                    both_single_cluster, torch.ones_like(ari), ari
                ).squeeze()
                self._num_examples += 1

    @sync_all_reduce("_num_examples:SUM", "_sum_fg_aris:SUM")
    def compute(self):
        if self._num_examples == 0:
            raise NotComputableError(
                "CustomAccuracy must have at least one example before it can be computed."
            )
        return self._sum_fg_aris.item() / self._num_examples

    @torch.no_grad()
    def iteration_completed(self, engine: Engine) -> None:
        output = self._output_transform(engine.state.output)
        self.update(output)