test_rotation_conversions.py 11.5 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
8
9
10


import itertools
import math
import unittest
11
from distutils.version import LooseVersion
12
from typing import Optional, Union
facebook-github-bot's avatar
facebook-github-bot committed
13

14
import numpy as np
15
import torch
facebook-github-bot's avatar
facebook-github-bot committed
16
from pytorch3d.transforms.rotation_conversions import (
17
18
    axis_angle_to_matrix,
    axis_angle_to_quaternion,
facebook-github-bot's avatar
facebook-github-bot committed
19
    euler_angles_to_matrix,
20
    matrix_to_axis_angle,
facebook-github-bot's avatar
facebook-github-bot committed
21
22
    matrix_to_euler_angles,
    matrix_to_quaternion,
23
    matrix_to_rotation_6d,
facebook-github-bot's avatar
facebook-github-bot committed
24
25
    quaternion_apply,
    quaternion_multiply,
26
    quaternion_to_axis_angle,
facebook-github-bot's avatar
facebook-github-bot committed
27
28
29
30
    quaternion_to_matrix,
    random_quaternions,
    random_rotation,
    random_rotations,
31
    rotation_6d_to_matrix,
facebook-github-bot's avatar
facebook-github-bot committed
32
33
)

Jeremy Reizenstein's avatar
Jeremy Reizenstein committed
34
35
from .common_testing import TestCaseMixin

facebook-github-bot's avatar
facebook-github-bot committed
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

class TestRandomRotation(unittest.TestCase):
    def setUp(self) -> None:
        super().setUp()
        torch.manual_seed(1)

    def test_random_rotation_invariant(self):
        """The image of the x-axis isn't biased among quadrants."""
        N = 1000
        base = random_rotation()
        quadrants = list(itertools.product([False, True], repeat=3))

        matrices = random_rotations(N)
        transformed = torch.matmul(base, matrices)
        transformed2 = torch.matmul(matrices, base)

        for k, results in enumerate([matrices, transformed, transformed2]):
            counts = {i: 0 for i in quadrants}
            for j in range(N):
                counts[tuple(i.item() > 0 for i in results[j, 0])] += 1
            average = N / 8.0
            counts_tensor = torch.tensor(list(counts.values()))
            chisquare_statistic = torch.sum(
                (counts_tensor - average) * (counts_tensor - average) / average
            )
            # The 0.1 significance level for chisquare(8-1) is
            # scipy.stats.chi2(7).ppf(0.9) == 12.017.
63
            self.assertLess(chisquare_statistic, 12, (counts, chisquare_statistic, k))
facebook-github-bot's avatar
facebook-github-bot committed
64
65


66
class TestRotationConversion(TestCaseMixin, unittest.TestCase):
facebook-github-bot's avatar
facebook-github-bot committed
67
68
69
70
71
72
73
74
    def setUp(self) -> None:
        super().setUp()
        torch.manual_seed(1)

    def test_from_quat(self):
        """quat -> mtx -> quat"""
        data = random_quaternions(13, dtype=torch.float64)
        mdata = matrix_to_quaternion(quaternion_to_matrix(data))
75
        self._assert_quaternions_close(data, mdata)
facebook-github-bot's avatar
facebook-github-bot committed
76
77
78
79
80

    def test_to_quat(self):
        """mtx -> quat -> mtx"""
        data = random_rotations(13, dtype=torch.float64)
        mdata = quaternion_to_matrix(matrix_to_quaternion(data))
81
        self.assertClose(data, mdata)
facebook-github-bot's avatar
facebook-github-bot committed
82
83
84

    def test_quat_grad_exists(self):
        """Quaternion calculations are differentiable."""
85
86
        rotation = random_rotation()
        rotation.requires_grad = True
facebook-github-bot's avatar
facebook-github-bot committed
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
        modified = quaternion_to_matrix(matrix_to_quaternion(rotation))
        [g] = torch.autograd.grad(modified.sum(), rotation)
        self.assertTrue(torch.isfinite(g).all())

    def _tait_bryan_conventions(self):
        return map("".join, itertools.permutations("XYZ"))

    def _proper_euler_conventions(self):
        letterpairs = itertools.permutations("XYZ", 2)
        return (l0 + l1 + l0 for l0, l1 in letterpairs)

    def _all_euler_angle_conventions(self):
        return itertools.chain(
            self._tait_bryan_conventions(), self._proper_euler_conventions()
        )

    def test_conventions(self):
        """The conventions listings have the right length."""
        all = list(self._all_euler_angle_conventions())
        self.assertEqual(len(all), 12)
        self.assertEqual(len(set(all)), 12)

    def test_from_euler(self):
        """euler -> mtx -> euler"""
        n_repetitions = 10
        # tolerance is how much we keep the middle angle away from the extreme
        # allowed values which make the calculation unstable (Gimbal lock).
        tolerance = 0.04
        half_pi = math.pi / 2
        data = torch.zeros(n_repetitions, 3)
        data.uniform_(-math.pi, math.pi)

        data[:, 1].uniform_(-half_pi + tolerance, half_pi - tolerance)
        for convention in self._tait_bryan_conventions():
            matrices = euler_angles_to_matrix(data, convention)
            mdata = matrix_to_euler_angles(matrices, convention)
123
            self.assertClose(data, mdata)
facebook-github-bot's avatar
facebook-github-bot committed
124
125
126
127
128

        data[:, 1] += half_pi
        for convention in self._proper_euler_conventions():
            matrices = euler_angles_to_matrix(data, convention)
            mdata = matrix_to_euler_angles(matrices, convention)
129
            self.assertClose(data, mdata)
facebook-github-bot's avatar
facebook-github-bot committed
130
131
132
133
134
135
136

    def test_to_euler(self):
        """mtx -> euler -> mtx"""
        data = random_rotations(13, dtype=torch.float64)
        for convention in self._all_euler_angle_conventions():
            euler_angles = matrix_to_euler_angles(data, convention)
            mdata = euler_angles_to_matrix(euler_angles, convention)
137
            self.assertClose(data, mdata)
facebook-github-bot's avatar
facebook-github-bot committed
138
139
140

    def test_euler_grad_exists(self):
        """Euler angle calculations are differentiable."""
141
142
        rotation = random_rotation(dtype=torch.float64)
        rotation.requires_grad = True
facebook-github-bot's avatar
facebook-github-bot committed
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
        for convention in self._all_euler_angle_conventions():
            euler_angles = matrix_to_euler_angles(rotation, convention)
            mdata = euler_angles_to_matrix(euler_angles, convention)
            [g] = torch.autograd.grad(mdata.sum(), rotation)
            self.assertTrue(torch.isfinite(g).all())

    def test_quaternion_multiplication(self):
        """Quaternion and matrix multiplication are equivalent."""
        a = random_quaternions(15, torch.float64).reshape((3, 5, 4))
        b = random_quaternions(21, torch.float64).reshape((7, 3, 1, 4))
        ab = quaternion_multiply(a, b)
        self.assertEqual(ab.shape, (7, 3, 5, 4))
        a_matrix = quaternion_to_matrix(a)
        b_matrix = quaternion_to_matrix(b)
        ab_matrix = torch.matmul(a_matrix, b_matrix)
        ab_from_matrix = matrix_to_quaternion(ab_matrix)
159
        self._assert_quaternions_close(ab, ab_from_matrix)
facebook-github-bot's avatar
facebook-github-bot committed
160

161
162
163
164
165
166
167
168
169
170
171
172
    def test_matrix_to_quaternion_corner_case(self):
        """Check no bad gradients from sqrt(0)."""
        matrix = torch.eye(3, requires_grad=True)
        target = torch.Tensor([0.984808, 0, 0.174, 0])

        optimizer = torch.optim.Adam([matrix], lr=0.05)
        optimizer.zero_grad()
        q = matrix_to_quaternion(matrix)
        loss = torch.sum((q - target) ** 2)
        loss.backward()
        optimizer.step()

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
        self.assertClose(matrix, matrix, msg="Result has non-finite values")
        delta = 1e-2
        self.assertLess(
            matrix.trace(),
            3.0 - delta,
            msg="Identity initialisation unchanged by a gradient step",
        )

    def test_matrix_to_quaternion_by_pi(self):
        # We check that rotations by pi around each of the 26
        # nonzero vectors containing nothing but 0, 1 and -1
        # are mapped to the right quaternions.
        # This is representative across the directions.
        options = [0.0, -1.0, 1.0]
        axes = [
            torch.tensor(vec)
            for vec in itertools.islice(  # exclude [0, 0, 0]
                itertools.product(options, options, options), 1, None
            )
        ]

        axes = torch.nn.functional.normalize(torch.stack(axes), dim=-1)
        # Rotation by pi around unit vector x is given by
        # the matrix 2 x x^T - Id.
        R = 2 * torch.matmul(axes[..., None], axes[..., None, :]) - torch.eye(3)
        quats_hat = matrix_to_quaternion(R)
        R_hat = quaternion_to_matrix(quats_hat)
        self.assertClose(R, R_hat, atol=1e-3)
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
    def test_from_axis_angle(self):
        """axis_angle -> mtx -> axis_angle"""
        n_repetitions = 20
        data = torch.rand(n_repetitions, 3)
        matrices = axis_angle_to_matrix(data)
        mdata = matrix_to_axis_angle(matrices)
        self.assertClose(data, mdata, atol=2e-6)

    def test_from_axis_angle_has_grad(self):
        n_repetitions = 20
        data = torch.rand(n_repetitions, 3, requires_grad=True)
        matrices = axis_angle_to_matrix(data)
        mdata = matrix_to_axis_angle(matrices)
        quats = axis_angle_to_quaternion(data)
        mdata2 = quaternion_to_axis_angle(quats)
        (grad,) = torch.autograd.grad(mdata.sum() + mdata2.sum(), data)
        self.assertTrue(torch.isfinite(grad).all())

    def test_to_axis_angle(self):
        """mtx -> axis_angle -> mtx"""
        data = random_rotations(13, dtype=torch.float64)
        euler_angles = matrix_to_axis_angle(data)
        mdata = axis_angle_to_matrix(euler_angles)
        self.assertClose(data, mdata)

facebook-github-bot's avatar
facebook-github-bot committed
227
228
    def test_quaternion_application(self):
        """Applying a quaternion is the same as applying the matrix."""
229
230
        quaternions = random_quaternions(3, torch.float64)
        quaternions.requires_grad = True
facebook-github-bot's avatar
facebook-github-bot committed
231
232
233
234
        matrices = quaternion_to_matrix(quaternions)
        points = torch.randn(3, 3, dtype=torch.float64, requires_grad=True)
        transform1 = quaternion_apply(quaternions, points)
        transform2 = torch.matmul(matrices, points[..., None])[..., 0]
235
        self.assertClose(transform1, transform2)
facebook-github-bot's avatar
facebook-github-bot committed
236
237
238
239

        [p, q] = torch.autograd.grad(transform1.sum(), [points, quaternions])
        self.assertTrue(torch.isfinite(p).all())
        self.assertTrue(torch.isfinite(q).all())
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

    def test_6d(self):
        """Converting to 6d and back"""
        r = random_rotations(13, dtype=torch.float64)

        # 6D representation is not unique,
        # but we implement it by taking the first two rows of the matrix
        r6d = matrix_to_rotation_6d(r)
        self.assertClose(r6d, r[:, :2, :].reshape(-1, 6))

        # going to 6D and back should not change the matrix
        r_hat = rotation_6d_to_matrix(r6d)
        self.assertClose(r_hat, r)

        # moving the second row R2 in the span of (R1, R2) should not matter
        r6d[:, 3:] += 2 * r6d[:, :3]
        r6d[:, :3] *= 3.0
        r_hat = rotation_6d_to_matrix(r6d)
        self.assertClose(r_hat, r)

        # check that we map anything to a valid rotation
        r6d = torch.rand(13, 6)
        r6d[:4, :] *= 3.0
        r6d[4:8, :] -= 0.5
        r = rotation_6d_to_matrix(r6d)
        self.assertClose(
            torch.matmul(r, r.permute(0, 2, 1)), torch.eye(3).expand_as(r), atol=1e-6
        )
268

269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
    @unittest.skipIf(LooseVersion(torch.__version__) < "1.9", "recent torchscript only")
    def test_scriptable(self):
        torch.jit.script(axis_angle_to_matrix)
        torch.jit.script(axis_angle_to_quaternion)
        torch.jit.script(euler_angles_to_matrix)
        torch.jit.script(matrix_to_axis_angle)
        torch.jit.script(matrix_to_euler_angles)
        torch.jit.script(matrix_to_quaternion)
        torch.jit.script(matrix_to_rotation_6d)
        torch.jit.script(quaternion_apply)
        torch.jit.script(quaternion_multiply)
        torch.jit.script(quaternion_to_matrix)
        torch.jit.script(quaternion_to_axis_angle)
        torch.jit.script(random_quaternions)
        torch.jit.script(random_rotation)
        torch.jit.script(random_rotations)
        torch.jit.script(random_quaternions)
        torch.jit.script(rotation_6d_to_matrix)

288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
    def _assert_quaternions_close(
        self,
        input: Union[torch.Tensor, np.ndarray],
        other: Union[torch.Tensor, np.ndarray],
        *,
        rtol: float = 1e-05,
        atol: float = 1e-08,
        equal_nan: bool = False,
        msg: Optional[str] = None,
    ):
        self.assertEqual(np.shape(input), np.shape(other))
        dot = (input * other).sum(-1)
        ones = torch.ones_like(dot)
        self.assertClose(
            dot.abs(), ones, rtol=rtol, atol=atol, equal_nan=equal_nan, msg=msg
        )