test_chamfer.py 30.2 KB
Newer Older
1
# Copyright (c) Meta Platforms, Inc. and affiliates.
Patrick Labatut's avatar
Patrick Labatut committed
2
3
4
5
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
facebook-github-bot's avatar
facebook-github-bot committed
6
7

import unittest
Nikhila Ravi's avatar
Nikhila Ravi committed
8
from collections import namedtuple
9

Nikhila Ravi's avatar
Nikhila Ravi committed
10
import numpy as np
facebook-github-bot's avatar
facebook-github-bot committed
11
12
import torch
import torch.nn.functional as F
13
from pytorch3d.loss import chamfer_distance
Nikhila Ravi's avatar
Nikhila Ravi committed
14
15
from pytorch3d.structures.pointclouds import Pointclouds

Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
16
17
from .common_testing import get_random_cuda_device, TestCaseMixin

Nikhila Ravi's avatar
Nikhila Ravi committed
18
19
20
21
22

# Output of init_pointclouds
points_normals = namedtuple(
    "points_normals", "p1_lengths p2_lengths cloud1 cloud2 p1 p2 n1 n2 weights"
)
facebook-github-bot's avatar
facebook-github-bot committed
23

Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
24
25

class TestChamfer(TestCaseMixin, unittest.TestCase):
Nikhila Ravi's avatar
Nikhila Ravi committed
26
27
28
29
30
    def setUp(self) -> None:
        super().setUp()
        torch.manual_seed(1)

    @staticmethod
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
31
32
33
    def init_pointclouds(
        N, P1, P2, device, requires_grad: bool = True, allow_empty: bool = True
    ):
Nikhila Ravi's avatar
Nikhila Ravi committed
34
35
36
37
38
39
        """
        Create 2 pointclouds object and associated padded points/normals tensors by
        starting from lists. The clouds and tensors have the same data. The
        leaf nodes for the clouds are a list of tensors. The padded tensor can be
        used directly as a leaf node.
        """
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
40
41
42
        low = 0 if allow_empty else 1
        p1_lengths = torch.randint(low, P1, size=(N,), dtype=torch.int64, device=device)
        p2_lengths = torch.randint(low, P2, size=(N,), dtype=torch.int64, device=device)
Nikhila Ravi's avatar
Nikhila Ravi committed
43
44
        P1 = p1_lengths.max().item()
        P2 = p2_lengths.max().item()
Nikhila Ravi's avatar
Nikhila Ravi committed
45
46
47
        weights = torch.rand((N,), dtype=torch.float32, device=device)

        # list of points and normals tensors
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
48
49
50
51
52
53
54
        p1 = torch.rand((N, P1, 3), dtype=torch.float32, device=device)
        p2 = torch.rand((N, P2, 3), dtype=torch.float32, device=device)
        n1 = torch.rand((N, P1, 3), dtype=torch.float32, device=device)
        n2 = torch.rand((N, P2, 3), dtype=torch.float32, device=device)
        n1 /= n1.norm(dim=-1, p=2, keepdim=True)
        n2 /= n2.norm(dim=-1, p=2, keepdim=True)

Nikhila Ravi's avatar
Nikhila Ravi committed
55
56
57
58
59
60
61
        p1_list = []
        p2_list = []
        n1_list = []
        n2_list = []
        for i in range(N):
            l1 = p1_lengths[i]
            l2 = p2_lengths[i]
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
62
63
64
65
            p1_list.append(p1[i, :l1].clone())
            p2_list.append(p2[i, :l2].clone())
            n1_list.append(n1[i, :l1].clone())
            n2_list.append(n2[i, :l2].clone())
Nikhila Ravi's avatar
Nikhila Ravi committed
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89

        # Set requires_grad for all tensors in the lists and
        # padded tensors.
        if requires_grad:
            for p in p2_list + p1_list + n1_list + n2_list + [p1, p2, n1, n2]:
                p.requires_grad = True

        # Create pointclouds objects
        cloud1 = Pointclouds(points=p1_list, normals=n1_list)
        cloud2 = Pointclouds(points=p2_list, normals=n2_list)

        # Return pointclouds objects and padded tensors
        return points_normals(
            p1_lengths=p1_lengths,
            p2_lengths=p2_lengths,
            cloud1=cloud1,
            cloud2=cloud2,
            p1=p1,
            p2=p2,
            n1=n1,
            n2=n2,
            weights=weights,
        )

facebook-github-bot's avatar
facebook-github-bot committed
90
    @staticmethod
91
    def chamfer_distance_naive_pointclouds(p1, p2, norm: int = 2, device="cpu"):
facebook-github-bot's avatar
facebook-github-bot committed
92
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
93
94
95
96
        Naive iterative implementation of nearest neighbor and chamfer distance.
        x and y are assumed to be pointclouds objects with points and optionally normals.
        This functions supports heterogeneous pointclouds in a batch.
        Returns lists of the unreduced loss and loss_normals.
facebook-github-bot's avatar
facebook-github-bot committed
97
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
        x = p1.points_padded()
        y = p2.points_padded()
        N, P1, D = x.shape
        P2 = y.size(1)
        x_lengths = p1.num_points_per_cloud()
        y_lengths = p2.num_points_per_cloud()
        x_normals = p1.normals_padded()
        y_normals = p2.normals_padded()

        return_normals = x_normals is not None and y_normals is not None

        # Initialize all distances to + inf
        dist = torch.ones((N, P1, P2), dtype=torch.float32, device=device) * np.inf

        x_mask = (
            torch.arange(P1, device=x.device)[None] >= x_lengths[:, None]
        )  # shape [N, P1]
        y_mask = (
            torch.arange(P2, device=y.device)[None] >= y_lengths[:, None]
        )  # shape [N, P2]

Nikhila Ravi's avatar
Nikhila Ravi committed
119
120
        is_x_heterogeneous = (x_lengths != P1).any()
        is_y_heterogeneous = (y_lengths != P2).any()
Nikhila Ravi's avatar
Nikhila Ravi committed
121
122
123
124
        # Only calculate the distances for the points which are not masked
        for n in range(N):
            for i1 in range(x_lengths[n]):
                for i2 in range(y_lengths[n]):
125
126
127
128
129
130
131
132
                    if norm == 2:
                        dist[n, i1, i2] = torch.sum((x[n, i1, :] - y[n, i2, :]) ** 2)
                    elif norm == 1:
                        dist[n, i1, i2] = torch.sum(
                            torch.abs(x[n, i1, :] - y[n, i2, :])
                        )
                    else:
                        raise ValueError("No support for norm %d" % (norm))
Nikhila Ravi's avatar
Nikhila Ravi committed
133
134
135

        x_dist = torch.min(dist, dim=2)[0]  # (N, P1)
        y_dist = torch.min(dist, dim=1)[0]  # (N, P2)
facebook-github-bot's avatar
facebook-github-bot committed
136

Nikhila Ravi's avatar
Nikhila Ravi committed
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
        if is_x_heterogeneous:
            x_dist[x_mask] = 0.0
        if is_y_heterogeneous:
            y_dist[y_mask] = 0.0

        loss = [x_dist, y_dist]

        lnorm = [x.new_zeros(()), x.new_zeros(())]

        if return_normals:
            x_index = dist.argmin(2).view(N, P1, 1).expand(N, P1, 3)
            y_index = dist.argmin(1).view(N, P2, 1).expand(N, P2, 3)
            lnorm1 = 1 - torch.abs(
                F.cosine_similarity(
                    x_normals, y_normals.gather(1, x_index), dim=2, eps=1e-6
                )
            )
            lnorm2 = 1 - torch.abs(
                F.cosine_similarity(
                    y_normals, x_normals.gather(1, y_index), dim=2, eps=1e-6
                )
            )

            if is_x_heterogeneous:
                lnorm1[x_mask] = 0.0
            if is_y_heterogeneous:
                lnorm2[y_mask] = 0.0

            lnorm = [lnorm1, lnorm2]  # [(N, P1), (N, P2)]

        return loss, lnorm
facebook-github-bot's avatar
facebook-github-bot committed
168
169

    @staticmethod
170
    def chamfer_distance_naive(x, y, x_normals=None, y_normals=None, norm: int = 2):
facebook-github-bot's avatar
facebook-github-bot committed
171
172
        """
        Naive iterative implementation of nearest neighbor and chamfer distance.
Nikhila Ravi's avatar
Nikhila Ravi committed
173
174
        Returns lists of the unreduced loss and loss_normals. This naive
        version only supports homogeneous pointcouds in a batch.
facebook-github-bot's avatar
facebook-github-bot committed
175
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
176
177
        N, P1, D = x.shape
        P2 = y.size(1)
Nikhila Ravi's avatar
Nikhila Ravi committed
178
        device = x.device
Nikhila Ravi's avatar
Nikhila Ravi committed
179
        return_normals = x_normals is not None and y_normals is not None
facebook-github-bot's avatar
facebook-github-bot committed
180
181
182
183
184
        dist = torch.zeros((N, P1, P2), dtype=torch.float32, device=device)

        for n in range(N):
            for i1 in range(P1):
                for i2 in range(P2):
185
186
187
188
189
190
191
192
                    if norm == 2:
                        dist[n, i1, i2] = torch.sum((x[n, i1, :] - y[n, i2, :]) ** 2)
                    elif norm == 1:
                        dist[n, i1, i2] = torch.sum(
                            torch.abs(x[n, i1, :] - y[n, i2, :])
                        )
                    else:
                        raise ValueError("No support for norm %d" % (norm))
facebook-github-bot's avatar
facebook-github-bot committed
193
194
195
196
197

        loss = [
            torch.min(dist, dim=2)[0],  # (N, P1)
            torch.min(dist, dim=1)[0],  # (N, P2)
        ]
Nikhila Ravi's avatar
Nikhila Ravi committed
198
        lnorm = [x.new_zeros(()), x.new_zeros(())]
facebook-github-bot's avatar
facebook-github-bot committed
199
200

        if return_normals:
Nikhila Ravi's avatar
Nikhila Ravi committed
201
202
            x_index = dist.argmin(2).view(N, P1, 1).expand(N, P1, 3)
            y_index = dist.argmin(1).view(N, P2, 1).expand(N, P2, 3)
facebook-github-bot's avatar
facebook-github-bot committed
203
204
            lnorm1 = 1 - torch.abs(
                F.cosine_similarity(
Nikhila Ravi's avatar
Nikhila Ravi committed
205
                    x_normals, y_normals.gather(1, x_index), dim=2, eps=1e-6
facebook-github-bot's avatar
facebook-github-bot committed
206
207
208
209
                )
            )
            lnorm2 = 1 - torch.abs(
                F.cosine_similarity(
Nikhila Ravi's avatar
Nikhila Ravi committed
210
                    y_normals, x_normals.gather(1, y_index), dim=2, eps=1e-6
facebook-github-bot's avatar
facebook-github-bot committed
211
212
213
214
215
216
                )
            )
            lnorm = [lnorm1, lnorm2]  # [(N, P1), (N, P2)]

        return loss, lnorm

Nikhila Ravi's avatar
Nikhila Ravi committed
217
    def test_chamfer_point_batch_reduction_mean(self):
facebook-github-bot's avatar
facebook-github-bot committed
218
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
219
220
221
222
        Compare output of vectorized chamfer loss with naive implementation
        for the default settings (point_reduction = "mean" and batch_reduction = "mean")
        and no normals.
        This tests only uses homogeneous pointclouds.
facebook-github-bot's avatar
facebook-github-bot committed
223
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
224
        N, max_P1, max_P2 = 7, 10, 18
Nikhila Ravi's avatar
Nikhila Ravi committed
225
        device = get_random_cuda_device()
Nikhila Ravi's avatar
Nikhila Ravi committed
226

227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
        for norm in [1, 2]:
            points_normals = TestChamfer.init_pointclouds(N, max_P1, max_P2, device)
            p1 = points_normals.p1
            p2 = points_normals.p2
            weights = points_normals.weights
            p11 = p1.detach().clone()
            p22 = p2.detach().clone()
            p11.requires_grad = True
            p22.requires_grad = True
            P1 = p1.shape[1]
            P2 = p2.shape[1]

            pred_loss, pred_loss_norm = TestChamfer.chamfer_distance_naive(
                p1, p2, norm=norm
            )
Nikhila Ravi's avatar
Nikhila Ravi committed
242

243
244
245
246
247
            # point_reduction = "mean".
            loss, loss_norm = chamfer_distance(p11, p22, weights=weights, norm=norm)
            pred_loss = pred_loss[0].sum(1) / P1 + pred_loss[1].sum(1) / P2
            pred_loss *= weights
            pred_loss = pred_loss.sum() / weights.sum()
Nikhila Ravi's avatar
Nikhila Ravi committed
248

249
250
            self.assertClose(loss, pred_loss)
            self.assertTrue(loss_norm is None)
facebook-github-bot's avatar
facebook-github-bot committed
251

252
253
            # Check gradients
            self._check_gradients(loss, None, pred_loss, None, p1, p11, p2, p22)
Nikhila Ravi's avatar
Nikhila Ravi committed
254
255

    def test_chamfer_vs_naive_pointcloud(self):
facebook-github-bot's avatar
facebook-github-bot committed
256
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
257
258
259
260
        Test the default settings for chamfer_distance
        (point reduction = "mean" and batch_reduction="mean") but with heterogeneous
        pointclouds as input. Compare with the naive implementation of chamfer
        which supports heterogeneous pointcloud objects.
facebook-github-bot's avatar
facebook-github-bot committed
261
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
262
        N, max_P1, max_P2 = 3, 70, 70
Nikhila Ravi's avatar
Nikhila Ravi committed
263
        device = get_random_cuda_device()
facebook-github-bot's avatar
facebook-github-bot committed
264

265
266
267
268
269
        for norm in [1, 2]:
            points_normals = TestChamfer.init_pointclouds(N, max_P1, max_P2, device)
            weights = points_normals.weights
            x_lengths = points_normals.p1_lengths
            y_lengths = points_normals.p2_lengths
Nikhila Ravi's avatar
Nikhila Ravi committed
270

271
272
273
274
275
276
277
278
279
280
281
            # Chamfer with tensors as input for heterogeneous pointclouds.
            cham_tensor, norm_tensor = chamfer_distance(
                points_normals.p1,
                points_normals.p2,
                x_normals=points_normals.n1,
                y_normals=points_normals.n2,
                x_lengths=points_normals.p1_lengths,
                y_lengths=points_normals.p2_lengths,
                weights=weights,
                norm=norm,
            )
Nikhila Ravi's avatar
Nikhila Ravi committed
282

283
284
285
286
            # Chamfer with pointclouds as input.
            pred_loss, pred_norm_loss = TestChamfer.chamfer_distance_naive_pointclouds(
                points_normals.cloud1, points_normals.cloud2, norm=norm, device=device
            )
Nikhila Ravi's avatar
Nikhila Ravi committed
287

288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
            # Mean reduction point loss.
            pred_loss[0] *= weights.view(N, 1)
            pred_loss[1] *= weights.view(N, 1)
            pred_loss_mean = (
                pred_loss[0].sum(1) / x_lengths + pred_loss[1].sum(1) / y_lengths
            )
            pred_loss_mean = pred_loss_mean.sum()
            pred_loss_mean /= weights.sum()

            # Mean reduction norm loss.
            pred_norm_loss[0] *= weights.view(N, 1)
            pred_norm_loss[1] *= weights.view(N, 1)
            pred_norm_loss_mean = (
                pred_norm_loss[0].sum(1) / x_lengths
                + pred_norm_loss[1].sum(1) / y_lengths
            )
            pred_norm_loss_mean = pred_norm_loss_mean.sum() / weights.sum()
Nikhila Ravi's avatar
Nikhila Ravi committed
305

306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
            self.assertClose(pred_loss_mean, cham_tensor)
            self.assertClose(pred_norm_loss_mean, norm_tensor)

            self._check_gradients(
                cham_tensor,
                norm_tensor,
                pred_loss_mean,
                pred_norm_loss_mean,
                points_normals.cloud1.points_list(),
                points_normals.p1,
                points_normals.cloud2.points_list(),
                points_normals.p2,
                points_normals.cloud1.normals_list(),
                points_normals.n1,
                points_normals.cloud2.normals_list(),
                points_normals.n2,
                x_lengths,
                y_lengths,
            )
Nikhila Ravi's avatar
Nikhila Ravi committed
325
326
327
328

    def test_chamfer_pointcloud_object_withnormals(self):
        N = 5
        P1, P2 = 100, 100
Nikhila Ravi's avatar
Nikhila Ravi committed
329
        device = get_random_cuda_device()
Nikhila Ravi's avatar
Nikhila Ravi committed
330
331
332
333
334
335
336
337
338
339
340
341
342

        reductions = [
            ("sum", "sum"),
            ("mean", "sum"),
            ("sum", "mean"),
            ("mean", "mean"),
            ("sum", None),
            ("mean", None),
        ]
        for (point_reduction, batch_reduction) in reductions:

            # Reinitialize all the tensors so that the
            # backward pass can be computed.
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
343
344
345
            points_normals = TestChamfer.init_pointclouds(
                N, P1, P2, device, allow_empty=False
            )
Nikhila Ravi's avatar
Nikhila Ravi committed
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

            # Chamfer with pointclouds as input.
            cham_cloud, norm_cloud = chamfer_distance(
                points_normals.cloud1,
                points_normals.cloud2,
                point_reduction=point_reduction,
                batch_reduction=batch_reduction,
            )

            # Chamfer with tensors as input.
            cham_tensor, norm_tensor = chamfer_distance(
                points_normals.p1,
                points_normals.p2,
                x_lengths=points_normals.p1_lengths,
                y_lengths=points_normals.p2_lengths,
                x_normals=points_normals.n1,
                y_normals=points_normals.n2,
                point_reduction=point_reduction,
                batch_reduction=batch_reduction,
            )

            self.assertClose(cham_cloud, cham_tensor)
            self.assertClose(norm_cloud, norm_tensor)
            self._check_gradients(
                cham_tensor,
                norm_tensor,
                cham_cloud,
                norm_cloud,
                points_normals.cloud1.points_list(),
                points_normals.p1,
                points_normals.cloud2.points_list(),
                points_normals.p2,
                points_normals.cloud1.normals_list(),
                points_normals.n1,
                points_normals.cloud2.normals_list(),
                points_normals.n2,
                points_normals.p1_lengths,
                points_normals.p2_lengths,
            )

    def test_chamfer_pointcloud_object_nonormals(self):
        N = 5
        P1, P2 = 100, 100
Nikhila Ravi's avatar
Nikhila Ravi committed
389
        device = get_random_cuda_device()
Nikhila Ravi's avatar
Nikhila Ravi committed
390
391
392
393
394
395
396
397
398
399
400
401
402

        reductions = [
            ("sum", "sum"),
            ("mean", "sum"),
            ("sum", "mean"),
            ("mean", "mean"),
            ("sum", None),
            ("mean", None),
        ]
        for (point_reduction, batch_reduction) in reductions:

            # Reinitialize all the tensors so that the
            # backward pass can be computed.
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
403
404
405
            points_normals = TestChamfer.init_pointclouds(
                N, P1, P2, device, allow_empty=False
            )
Nikhila Ravi's avatar
Nikhila Ravi committed
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

            # Chamfer with pointclouds as input.
            cham_cloud, _ = chamfer_distance(
                points_normals.cloud1,
                points_normals.cloud2,
                point_reduction=point_reduction,
                batch_reduction=batch_reduction,
            )

            # Chamfer with tensors as input.
            cham_tensor, _ = chamfer_distance(
                points_normals.p1,
                points_normals.p2,
                x_lengths=points_normals.p1_lengths,
                y_lengths=points_normals.p2_lengths,
                point_reduction=point_reduction,
                batch_reduction=batch_reduction,
            )

            self.assertClose(cham_cloud, cham_tensor)
            self._check_gradients(
                cham_tensor,
                None,
                cham_cloud,
                None,
                points_normals.cloud1.points_list(),
                points_normals.p1,
                points_normals.cloud2.points_list(),
                points_normals.p2,
                lengths1=points_normals.p1_lengths,
                lengths2=points_normals.p2_lengths,
            )

    def test_chamfer_point_reduction_mean(self):
        """
        Compare output of vectorized chamfer loss with naive implementation
        for point_reduction = "mean" and batch_reduction = None.
        """
        N, max_P1, max_P2 = 7, 10, 18
Nikhila Ravi's avatar
Nikhila Ravi committed
445
        device = get_random_cuda_device()
Nikhila Ravi's avatar
Nikhila Ravi committed
446
447
448
449
450
451
452
453
454
455
456
457
458
        points_normals = TestChamfer.init_pointclouds(N, max_P1, max_P2, device)
        p1 = points_normals.p1
        p2 = points_normals.p2
        p1_normals = points_normals.n1
        p2_normals = points_normals.n2
        weights = points_normals.weights
        p11 = p1.detach().clone()
        p22 = p2.detach().clone()
        p11.requires_grad = True
        p22.requires_grad = True
        P1 = p1.shape[1]
        P2 = p2.shape[1]

facebook-github-bot's avatar
facebook-github-bot committed
459
        pred_loss, pred_loss_norm = TestChamfer.chamfer_distance_naive(
Nikhila Ravi's avatar
Nikhila Ravi committed
460
            p1, p2, x_normals=p1_normals, y_normals=p2_normals
facebook-github-bot's avatar
facebook-github-bot committed
461
462
463
464
        )

        # point_reduction = "mean".
        loss, loss_norm = chamfer_distance(
Nikhila Ravi's avatar
Nikhila Ravi committed
465
466
467
468
            p11,
            p22,
            x_normals=p1_normals,
            y_normals=p2_normals,
facebook-github-bot's avatar
facebook-github-bot committed
469
            weights=weights,
Nikhila Ravi's avatar
Nikhila Ravi committed
470
            batch_reduction=None,
facebook-github-bot's avatar
facebook-github-bot committed
471
472
473
474
            point_reduction="mean",
        )
        pred_loss_mean = pred_loss[0].sum(1) / P1 + pred_loss[1].sum(1) / P2
        pred_loss_mean *= weights
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
475
        self.assertClose(loss, pred_loss_mean)
facebook-github-bot's avatar
facebook-github-bot committed
476
477
478
479
480

        pred_loss_norm_mean = (
            pred_loss_norm[0].sum(1) / P1 + pred_loss_norm[1].sum(1) / P2
        )
        pred_loss_norm_mean *= weights
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
481
        self.assertClose(loss_norm, pred_loss_norm_mean)
facebook-github-bot's avatar
facebook-github-bot committed
482

Nikhila Ravi's avatar
Nikhila Ravi committed
483
484
485
486
487
488
489
490
491
492
493
        # Check gradients
        self._check_gradients(
            loss, loss_norm, pred_loss_mean, pred_loss_norm_mean, p1, p11, p2, p22
        )

    def test_chamfer_point_reduction_sum(self):
        """
        Compare output of vectorized chamfer loss with naive implementation
        for point_reduction = "sum" and batch_reduction = None.
        """
        N, P1, P2 = 7, 10, 18
Nikhila Ravi's avatar
Nikhila Ravi committed
494
        device = get_random_cuda_device()
Nikhila Ravi's avatar
Nikhila Ravi committed
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
        points_normals = TestChamfer.init_pointclouds(N, P1, P2, device)
        p1 = points_normals.p1
        p2 = points_normals.p2
        p1_normals = points_normals.n1
        p2_normals = points_normals.n2
        weights = points_normals.weights
        p11 = p1.detach().clone()
        p22 = p2.detach().clone()
        p11.requires_grad = True
        p22.requires_grad = True

        pred_loss, pred_loss_norm = TestChamfer.chamfer_distance_naive(
            p1, p2, x_normals=p1_normals, y_normals=p2_normals
        )

facebook-github-bot's avatar
facebook-github-bot committed
510
        loss, loss_norm = chamfer_distance(
Nikhila Ravi's avatar
Nikhila Ravi committed
511
512
513
514
            p11,
            p22,
            x_normals=p1_normals,
            y_normals=p2_normals,
facebook-github-bot's avatar
facebook-github-bot committed
515
            weights=weights,
Nikhila Ravi's avatar
Nikhila Ravi committed
516
            batch_reduction=None,
facebook-github-bot's avatar
facebook-github-bot committed
517
518
519
520
            point_reduction="sum",
        )
        pred_loss_sum = pred_loss[0].sum(1) + pred_loss[1].sum(1)
        pred_loss_sum *= weights
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
521
        self.assertClose(loss, pred_loss_sum)
facebook-github-bot's avatar
facebook-github-bot committed
522
523
524

        pred_loss_norm_sum = pred_loss_norm[0].sum(1) + pred_loss_norm[1].sum(1)
        pred_loss_norm_sum *= weights
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
525
        self.assertClose(loss_norm, pred_loss_norm_sum)
facebook-github-bot's avatar
facebook-github-bot committed
526

Nikhila Ravi's avatar
Nikhila Ravi committed
527
528
529
530
        # Check gradients
        self._check_gradients(
            loss, loss_norm, pred_loss_sum, pred_loss_norm_sum, p1, p11, p2, p22
        )
facebook-github-bot's avatar
facebook-github-bot committed
531

Nikhila Ravi's avatar
Nikhila Ravi committed
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
    def _check_gradients(
        self,
        loss,
        loss_norm,
        pred_loss,
        pred_loss_norm,
        x1,
        x2,
        y1,
        y2,
        xn1=None,  # normals
        xn2=None,  # normals
        yn1=None,  # normals
        yn2=None,  # normals
        lengths1=None,
        lengths2=None,
    ):
facebook-github-bot's avatar
facebook-github-bot committed
549
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
550
551
552
        x1 and x2 can have different types based on the leaf node used in the calculation:
        e.g. x1 may be a list of tensors whereas x2 is a padded tensor.
        This also applies for the pairs: (y1, y2), (xn1, xn2), (yn1, yn2).
facebook-github-bot's avatar
facebook-github-bot committed
553
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
554
        grad_loss = torch.rand(loss.shape, device=loss.device, dtype=loss.dtype)
facebook-github-bot's avatar
facebook-github-bot committed
555

Nikhila Ravi's avatar
Nikhila Ravi committed
556
557
558
559
560
561
562
563
        # Loss for normals is optional. Iniitalize to 0.
        norm_loss_term = pred_norm_loss_term = 0.0
        if loss_norm is not None and pred_loss_norm is not None:
            grad_normals = torch.rand(
                loss_norm.shape, device=loss.device, dtype=loss.dtype
            )
            norm_loss_term = loss_norm * grad_normals
            pred_norm_loss_term = pred_loss_norm * grad_normals
facebook-github-bot's avatar
facebook-github-bot committed
564

Nikhila Ravi's avatar
Nikhila Ravi committed
565
566
567
568
        l1 = (loss * grad_loss) + norm_loss_term
        l1.sum().backward()
        l2 = (pred_loss * grad_loss) + pred_norm_loss_term
        l2.sum().backward()
facebook-github-bot's avatar
facebook-github-bot committed
569

Nikhila Ravi's avatar
Nikhila Ravi committed
570
571
        self._check_grad_by_type(x1, x2, lengths1)
        self._check_grad_by_type(y1, y2, lengths2)
facebook-github-bot's avatar
facebook-github-bot committed
572

Nikhila Ravi's avatar
Nikhila Ravi committed
573
574
575
576
        # If leaf nodes for normals are passed in, check their gradients.
        if all(n is not None for n in [xn1, xn2, yn1, yn2]):
            self._check_grad_by_type(xn1, xn2, lengths1)
            self._check_grad_by_type(yn1, yn2, lengths2)
facebook-github-bot's avatar
facebook-github-bot committed
577

Nikhila Ravi's avatar
Nikhila Ravi committed
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
    def _check_grad_by_type(self, x1, x2, lengths=None):
        """
        x1 and x2 can be of different types e.g. list or tensor - compare appropriately
        based on the types.
        """
        error_msg = "All values for gradient checks must be tensors or lists of tensors"

        if all(isinstance(p, list) for p in [x1, x2]):
            # Lists of tensors
            for i in range(len(x1)):
                self.assertClose(x1[i].grad, x2[i].grad)
        elif isinstance(x1, list) and torch.is_tensor(x2):
            self.assertIsNotNone(lengths)  # lengths is required

            # List of tensors vs padded tensor
            for i in range(len(x1)):
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
594
                self.assertClose(x1[i].grad, x2.grad[i, : lengths[i]], atol=1e-7)
Nikhila Ravi's avatar
Nikhila Ravi committed
595
596
597
598
599
600
                self.assertTrue(x2.grad[i, lengths[i] :].sum().item() == 0.0)
        elif all(torch.is_tensor(p) for p in [x1, x2]):
            # Two tensors
            self.assertClose(x1.grad, x2.grad)
        else:
            raise ValueError(error_msg)
facebook-github-bot's avatar
facebook-github-bot committed
601
602
603
604

    def test_chamfer_joint_reduction(self):
        """
        Compare output of vectorized chamfer loss with naive implementation
Nikhila Ravi's avatar
Nikhila Ravi committed
605
        when batch_reduction in ["mean", "sum"] and
facebook-github-bot's avatar
facebook-github-bot committed
606
607
        point_reduction in ["mean", "sum"].
        """
Nikhila Ravi's avatar
Nikhila Ravi committed
608
        N, max_P1, max_P2 = 7, 10, 18
Nikhila Ravi's avatar
Nikhila Ravi committed
609
        device = get_random_cuda_device()
Nikhila Ravi's avatar
Nikhila Ravi committed
610
611
612
613
614
615
616
617
618
619

        points_normals = TestChamfer.init_pointclouds(N, max_P1, max_P2, device)
        p1 = points_normals.p1
        p2 = points_normals.p2
        p1_normals = points_normals.n1
        p2_normals = points_normals.n2
        weights = points_normals.weights

        P1 = p1.shape[1]
        P2 = p2.shape[1]
facebook-github-bot's avatar
facebook-github-bot committed
620
621

        pred_loss, pred_loss_norm = TestChamfer.chamfer_distance_naive(
Nikhila Ravi's avatar
Nikhila Ravi committed
622
            p1, p2, x_normals=p1_normals, y_normals=p2_normals
facebook-github-bot's avatar
facebook-github-bot committed
623
624
625
626
627
628
        )

        # batch_reduction = "sum", point_reduction = "sum".
        loss, loss_norm = chamfer_distance(
            p1,
            p2,
Nikhila Ravi's avatar
Nikhila Ravi committed
629
630
            x_normals=p1_normals,
            y_normals=p2_normals,
facebook-github-bot's avatar
facebook-github-bot committed
631
632
633
634
635
636
637
638
            weights=weights,
            batch_reduction="sum",
            point_reduction="sum",
        )
        pred_loss[0] *= weights.view(N, 1)
        pred_loss[1] *= weights.view(N, 1)
        pred_loss_sum = pred_loss[0].sum(1) + pred_loss[1].sum(1)  # point sum
        pred_loss_sum = pred_loss_sum.sum()  # batch sum
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
639
        self.assertClose(loss, pred_loss_sum)
facebook-github-bot's avatar
facebook-github-bot committed
640
641
642
643
644
645
646

        pred_loss_norm[0] *= weights.view(N, 1)
        pred_loss_norm[1] *= weights.view(N, 1)
        pred_loss_norm_sum = pred_loss_norm[0].sum(1) + pred_loss_norm[1].sum(
            1
        )  # point sum.
        pred_loss_norm_sum = pred_loss_norm_sum.sum()  # batch sum
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
647
        self.assertClose(loss_norm, pred_loss_norm_sum)
facebook-github-bot's avatar
facebook-github-bot committed
648
649
650
651
652

        # batch_reduction = "mean", point_reduction = "sum".
        loss, loss_norm = chamfer_distance(
            p1,
            p2,
Nikhila Ravi's avatar
Nikhila Ravi committed
653
654
            x_normals=p1_normals,
            y_normals=p2_normals,
facebook-github-bot's avatar
facebook-github-bot committed
655
656
657
658
659
            weights=weights,
            batch_reduction="mean",
            point_reduction="sum",
        )
        pred_loss_sum /= weights.sum()
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
660
        self.assertClose(loss, pred_loss_sum)
facebook-github-bot's avatar
facebook-github-bot committed
661
662

        pred_loss_norm_sum /= weights.sum()
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
663
        self.assertClose(loss_norm, pred_loss_norm_sum)
facebook-github-bot's avatar
facebook-github-bot committed
664
665
666
667
668

        # batch_reduction = "sum", point_reduction = "mean".
        loss, loss_norm = chamfer_distance(
            p1,
            p2,
Nikhila Ravi's avatar
Nikhila Ravi committed
669
670
            x_normals=p1_normals,
            y_normals=p2_normals,
facebook-github-bot's avatar
facebook-github-bot committed
671
672
673
674
675
676
            weights=weights,
            batch_reduction="sum",
            point_reduction="mean",
        )
        pred_loss_mean = pred_loss[0].sum(1) / P1 + pred_loss[1].sum(1) / P2
        pred_loss_mean = pred_loss_mean.sum()
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
677
        self.assertClose(loss, pred_loss_mean)
facebook-github-bot's avatar
facebook-github-bot committed
678
679
680
681
682

        pred_loss_norm_mean = (
            pred_loss_norm[0].sum(1) / P1 + pred_loss_norm[1].sum(1) / P2
        )
        pred_loss_norm_mean = pred_loss_norm_mean.sum()
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
683
        self.assertClose(loss_norm, pred_loss_norm_mean)
facebook-github-bot's avatar
facebook-github-bot committed
684
685
686
687
688

        # batch_reduction = "mean", point_reduction = "mean". This is the default.
        loss, loss_norm = chamfer_distance(
            p1,
            p2,
Nikhila Ravi's avatar
Nikhila Ravi committed
689
690
            x_normals=p1_normals,
            y_normals=p2_normals,
facebook-github-bot's avatar
facebook-github-bot committed
691
692
693
694
695
            weights=weights,
            batch_reduction="mean",
            point_reduction="mean",
        )
        pred_loss_mean /= weights.sum()
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
696
        self.assertClose(loss, pred_loss_mean)
facebook-github-bot's avatar
facebook-github-bot committed
697
698

        pred_loss_norm_mean /= weights.sum()
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
699
        self.assertClose(loss_norm, pred_loss_norm_mean)
facebook-github-bot's avatar
facebook-github-bot committed
700

Nikhila Ravi's avatar
Nikhila Ravi committed
701
702
703
704
705
706
707
708
        # Error when batch_reduction is not in ["mean", "sum"] or None.
        with self.assertRaisesRegex(ValueError, "batch_reduction must be one of"):
            chamfer_distance(p1, p2, weights=weights, batch_reduction="max")

        # Error when point_reduction is not in ["mean", "sum"].
        with self.assertRaisesRegex(ValueError, "point_reduction must be one of"):
            chamfer_distance(p1, p2, weights=weights, point_reduction=None)

facebook-github-bot's avatar
facebook-github-bot committed
709
710
    def test_incorrect_weights(self):
        N, P1, P2 = 16, 64, 128
Nikhila Ravi's avatar
Nikhila Ravi committed
711
        device = get_random_cuda_device()
facebook-github-bot's avatar
facebook-github-bot committed
712
713
714
715
716
717
718
719
720
721
722
        p1 = torch.rand(
            (N, P1, 3), dtype=torch.float32, device=device, requires_grad=True
        )
        p2 = torch.rand(
            (N, P2, 3), dtype=torch.float32, device=device, requires_grad=True
        )

        weights = torch.zeros((N,), dtype=torch.float32, device=device)
        loss, loss_norm = chamfer_distance(
            p1, p2, weights=weights, batch_reduction="mean"
        )
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
723
        self.assertClose(loss.cpu(), torch.zeros(()))
facebook-github-bot's avatar
facebook-github-bot committed
724
        self.assertTrue(loss.requires_grad)
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
725
        self.assertClose(loss_norm.cpu(), torch.zeros(()))
facebook-github-bot's avatar
facebook-github-bot committed
726
727
728
        self.assertTrue(loss_norm.requires_grad)

        loss, loss_norm = chamfer_distance(
Nikhila Ravi's avatar
Nikhila Ravi committed
729
            p1, p2, weights=weights, batch_reduction=None
facebook-github-bot's avatar
facebook-github-bot committed
730
        )
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
731
        self.assertClose(loss.cpu(), torch.zeros((N, N)))
facebook-github-bot's avatar
facebook-github-bot committed
732
        self.assertTrue(loss.requires_grad)
Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
733
        self.assertClose(loss_norm.cpu(), torch.zeros((N, N)))
facebook-github-bot's avatar
facebook-github-bot committed
734
735
736
737
738
739
740
741
742
743
        self.assertTrue(loss_norm.requires_grad)

        weights = torch.ones((N,), dtype=torch.float32, device=device) * -1
        with self.assertRaises(ValueError):
            loss, loss_norm = chamfer_distance(p1, p2, weights=weights)

        weights = torch.zeros((N - 1,), dtype=torch.float32, device=device)
        with self.assertRaises(ValueError):
            loss, loss_norm = chamfer_distance(p1, p2, weights=weights)

Nikhila Ravi's avatar
Nikhila Ravi committed
744
745
    def test_incorrect_inputs(self):
        N, P1, P2 = 7, 10, 18
Nikhila Ravi's avatar
Nikhila Ravi committed
746
        device = get_random_cuda_device()
Nikhila Ravi's avatar
Nikhila Ravi committed
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
        points_normals = TestChamfer.init_pointclouds(N, P1, P2, device)
        p1 = points_normals.p1
        p2 = points_normals.p2
        p1_normals = points_normals.n1

        # Normals of wrong shape
        with self.assertRaisesRegex(ValueError, "Expected normals to be of shape"):
            chamfer_distance(p1, p2, x_normals=p1_normals[None])

        # Points of wrong shape
        with self.assertRaisesRegex(ValueError, "Expected points to be of shape"):
            chamfer_distance(p1[None], p2)

        # Lengths of wrong shape
        with self.assertRaisesRegex(ValueError, "Expected lengths to be of shape"):
            chamfer_distance(p1, p2, x_lengths=torch.tensor([1, 2, 3], device=device))

        # Points are not a tensor or Pointclouds
        with self.assertRaisesRegex(ValueError, "Pointclouds objects or torch.Tensor"):
            chamfer_distance(x=[1, 1, 1], y=[1, 1, 1])

768
769
770
771
772
773
774
775
776
777
778
779
780
    def test_invalid_norm(self):
        N, P1, P2 = 7, 10, 18
        device = get_random_cuda_device()
        points_normals = TestChamfer.init_pointclouds(N, P1, P2, device)
        p1 = points_normals.p1
        p2 = points_normals.p2

        with self.assertRaisesRegex(ValueError, "Support for 1 or 2 norm."):
            chamfer_distance(p1, p2, norm=0)

        with self.assertRaisesRegex(ValueError, "Support for 1 or 2 norm."):
            chamfer_distance(p1, p2, norm=3)

781
782
783
784
785
786
787
788
789
790
791
    def test_empty_clouds(self):
        # Check that point_reduction doesn't divide by zero
        points1 = Pointclouds(points=[torch.zeros(0, 3), torch.zeros(10, 3)])
        points2 = Pointclouds(points=torch.ones(2, 40, 3))
        loss, _ = chamfer_distance(points1, points2, batch_reduction=None)
        self.assertClose(loss, torch.tensor([0.0, 6.0]))

        # Check that batch_reduction doesn't divide by zero
        loss2, _ = chamfer_distance(Pointclouds([]), Pointclouds([]))
        self.assertClose(loss2, torch.tensor(0.0))

facebook-github-bot's avatar
facebook-github-bot committed
792
    @staticmethod
Nikhila Ravi's avatar
Nikhila Ravi committed
793
    def chamfer_with_init(
Nikhila Ravi's avatar
Nikhila Ravi committed
794
795
796
797
798
799
        batch_size: int,
        P1: int,
        P2: int,
        return_normals: bool,
        homogeneous: bool,
        device="cpu",
Nikhila Ravi's avatar
Nikhila Ravi committed
800
    ):
Nikhila Ravi's avatar
Nikhila Ravi committed
801
802
803
        points_normals = TestChamfer.init_pointclouds(batch_size, P1, P2, device=device)
        l1 = points_normals.p1_lengths
        l2 = points_normals.p2_lengths
Nikhila Ravi's avatar
Nikhila Ravi committed
804
805
806
807
808
        if homogeneous:
            # Set lengths to None so in Chamfer it assumes
            # there is no padding.
            l1 = l2 = None

facebook-github-bot's avatar
facebook-github-bot committed
809
810
811
812
        torch.cuda.synchronize()

        def loss():
            loss, loss_normals = chamfer_distance(
Nikhila Ravi's avatar
Nikhila Ravi committed
813
814
                points_normals.p1,
                points_normals.p2,
Nikhila Ravi's avatar
Nikhila Ravi committed
815
816
                x_lengths=l1,
                y_lengths=l2,
Nikhila Ravi's avatar
Nikhila Ravi committed
817
818
819
                x_normals=points_normals.n1,
                y_normals=points_normals.n2,
                weights=points_normals.weights,
facebook-github-bot's avatar
facebook-github-bot committed
820
821
822
823
824
825
826
            )
            torch.cuda.synchronize()

        return loss

    @staticmethod
    def chamfer_naive_with_init(
Nikhila Ravi's avatar
Nikhila Ravi committed
827
        batch_size: int, P1: int, P2: int, return_normals: bool, device="cpu"
facebook-github-bot's avatar
facebook-github-bot committed
828
    ):
Nikhila Ravi's avatar
Nikhila Ravi committed
829
        points_normals = TestChamfer.init_pointclouds(batch_size, P1, P2, device=device)
facebook-github-bot's avatar
facebook-github-bot committed
830
831
832
833
        torch.cuda.synchronize()

        def loss():
            loss, loss_normals = TestChamfer.chamfer_distance_naive(
Nikhila Ravi's avatar
Nikhila Ravi committed
834
835
836
837
                points_normals.p1,
                points_normals.p2,
                x_normals=points_normals.n1,
                y_normals=points_normals.n2,
facebook-github-bot's avatar
facebook-github-bot committed
838
839
840
841
            )
            torch.cuda.synchronize()

        return loss