test_ops.py 53.4 KB
Newer Older
1
import math
2
import os
3
from abc import ABC, abstractmethod
4
5
from functools import lru_cache
from typing import Tuple
6

7
import numpy as np
8
import pytest
9
import torch
10
import torch.fx
11
12
from common_utils import needs_cuda, cpu_and_gpu, assert_equal
from PIL import Image
13
from torch import nn, Tensor
14
from torch.autograd import gradcheck
15
from torch.nn.modules.utils import _pair
16
from torchvision import models, ops
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
from torchvision.models.feature_extraction import get_graph_node_names


class RoIOpTesterModuleWrapper(nn.Module):
    def __init__(self, obj):
        super().__init__()
        self.layer = obj
        self.n_inputs = 2

    def forward(self, a, b):
        self.layer(a, b)


class MultiScaleRoIAlignModuleWrapper(nn.Module):
    def __init__(self, obj):
        super().__init__()
        self.layer = obj
        self.n_inputs = 3

    def forward(self, a, b, c):
        self.layer(a, b, c)


class DeformConvModuleWrapper(nn.Module):
    def __init__(self, obj):
        super().__init__()
        self.layer = obj
        self.n_inputs = 3

    def forward(self, a, b, c):
        self.layer(a, b, c)


class StochasticDepthWrapper(nn.Module):
    def __init__(self, obj):
        super().__init__()
        self.layer = obj
        self.n_inputs = 1

    def forward(self, a):
        self.layer(a)
58
59


60
61
class RoIOpTester(ABC):
    dtype = torch.float64
62

63
64
    @pytest.mark.parametrize("device", cpu_and_gpu())
    @pytest.mark.parametrize("contiguous", (True, False))
65
    def test_forward(self, device, contiguous, x_dtype=None, rois_dtype=None, **kwargs):
66
67
        x_dtype = self.dtype if x_dtype is None else x_dtype
        rois_dtype = self.dtype if rois_dtype is None else rois_dtype
68
69
70
        pool_size = 5
        # n_channels % (pool_size ** 2) == 0 required for PS opeartions.
        n_channels = 2 * (pool_size ** 2)
71
        x = torch.rand(2, n_channels, 10, 10, dtype=x_dtype, device=device)
72
73
        if not contiguous:
            x = x.permute(0, 1, 3, 2)
74
75
76
77
78
        rois = torch.tensor(
            [[0, 0, 0, 9, 9], [0, 0, 5, 4, 9], [0, 5, 5, 9, 9], [1, 0, 0, 9, 9]],  # format is (xyxy)
            dtype=rois_dtype,
            device=device,
        )
79

80
        pool_h, pool_w = pool_size, pool_size
81
        y = self.fn(x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, **kwargs)
82
        # the following should be true whether we're running an autocast test or not.
83
        assert y.dtype == x.dtype
84
85
86
        gt_y = self.expected_fn(
            x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, device=device, dtype=self.dtype, **kwargs
        )
87

88
        tol = 1e-3 if (x_dtype is torch.half or rois_dtype is torch.half) else 1e-5
89
        torch.testing.assert_close(gt_y.to(y), y, rtol=tol, atol=tol)
90

91
92
93
94
95
96
97
98
99
    @pytest.mark.parametrize("device", cpu_and_gpu())
    def test_is_leaf_node(self, device):
        op_obj = self.make_obj(wrap=True).to(device=device)
        graph_node_names = get_graph_node_names(op_obj)

        assert len(graph_node_names) == 2
        assert len(graph_node_names[0]) == len(graph_node_names[1])
        assert len(graph_node_names[0]) == 1 + op_obj.n_inputs

100
    @pytest.mark.parametrize("seed", range(10))
101
102
    @pytest.mark.parametrize("device", cpu_and_gpu())
    @pytest.mark.parametrize("contiguous", (True, False))
103
104
    def test_backward(self, seed, device, contiguous):
        torch.random.manual_seed(seed)
105
106
107
108
        pool_size = 2
        x = torch.rand(1, 2 * (pool_size ** 2), 5, 5, dtype=self.dtype, device=device, requires_grad=True)
        if not contiguous:
            x = x.permute(0, 1, 3, 2)
109
110
111
        rois = torch.tensor(
            [[0, 0, 0, 4, 4], [0, 0, 2, 3, 4], [0, 2, 2, 4, 4]], dtype=self.dtype, device=device  # format is (xyxy)
        )
112

113
114
        def func(z):
            return self.fn(z, rois, pool_size, pool_size, spatial_scale=1, sampling_ratio=1)
115

116
        script_func = self.get_script_fn(rois, pool_size)
117

118
119
        gradcheck(func, (x,))
        gradcheck(script_func, (x,))
120

121
    @needs_cuda
122
123
    @pytest.mark.parametrize("x_dtype", (torch.float, torch.half))
    @pytest.mark.parametrize("rois_dtype", (torch.float, torch.half))
124
125
126
    def test_autocast(self, x_dtype, rois_dtype):
        with torch.cuda.amp.autocast():
            self.test_forward(torch.device("cuda"), contiguous=False, x_dtype=x_dtype, rois_dtype=rois_dtype)
127
128
129

    def _helper_boxes_shape(self, func):
        # test boxes as Tensor[N, 5]
130
        with pytest.raises(AssertionError):
131
132
133
134
135
            a = torch.linspace(1, 8 * 8, 8 * 8).reshape(1, 1, 8, 8)
            boxes = torch.tensor([[0, 0, 3, 3]], dtype=a.dtype)
            func(a, boxes, output_size=(2, 2))

        # test boxes as List[Tensor[N, 4]]
136
        with pytest.raises(AssertionError):
137
138
139
140
            a = torch.linspace(1, 8 * 8, 8 * 8).reshape(1, 1, 8, 8)
            boxes = torch.tensor([[0, 0, 3]], dtype=a.dtype)
            ops.roi_pool(a, [boxes], output_size=(2, 2))

141
    @abstractmethod
142
143
    def fn(*args, **kwargs):
        pass
144

145
146
147
148
    @abstractmethod
    def make_obj(*args, **kwargs):
        pass

149
    @abstractmethod
150
151
    def get_script_fn(*args, **kwargs):
        pass
152

153
    @abstractmethod
154
155
    def expected_fn(*args, **kwargs):
        pass
156

157

158
class TestRoiPool(RoIOpTester):
159
160
    def fn(self, x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, **kwargs):
        return ops.RoIPool((pool_h, pool_w), spatial_scale)(x, rois)
161

162
163
164
165
    def make_obj(self, pool_h=5, pool_w=5, spatial_scale=1, wrap=False):
        obj = ops.RoIPool((pool_h, pool_w), spatial_scale)
        return RoIOpTesterModuleWrapper(obj) if wrap else obj

166
    def get_script_fn(self, rois, pool_size):
Nicolas Hug's avatar
Nicolas Hug committed
167
168
        scriped = torch.jit.script(ops.roi_pool)
        return lambda x: scriped(x, rois, pool_size)
169

170
171
172
    def expected_fn(
        self, x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, device=None, dtype=torch.float64
    ):
173
174
        if device is None:
            device = torch.device("cpu")
175

176
177
        n_channels = x.size(1)
        y = torch.zeros(rois.size(0), n_channels, pool_h, pool_w, dtype=dtype, device=device)
178

179
180
        def get_slice(k, block):
            return slice(int(np.floor(k * block)), int(np.ceil((k + 1) * block)))
181

182
183
184
        for roi_idx, roi in enumerate(rois):
            batch_idx = int(roi[0])
            j_begin, i_begin, j_end, i_end = (int(round(x.item() * spatial_scale)) for x in roi[1:])
185
            roi_x = x[batch_idx, :, i_begin : i_end + 1, j_begin : j_end + 1]
186

187
188
189
            roi_h, roi_w = roi_x.shape[-2:]
            bin_h = roi_h / pool_h
            bin_w = roi_w / pool_w
190

191
192
193
194
195
196
            for i in range(0, pool_h):
                for j in range(0, pool_w):
                    bin_x = roi_x[:, get_slice(i, bin_h), get_slice(j, bin_w)]
                    if bin_x.numel() > 0:
                        y[roi_idx, :, i, j] = bin_x.reshape(n_channels, -1).max(dim=1)[0]
        return y
197

198
    def test_boxes_shape(self):
199
200
        self._helper_boxes_shape(ops.roi_pool)

201

202
class TestPSRoIPool(RoIOpTester):
203
204
    def fn(self, x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, **kwargs):
        return ops.PSRoIPool((pool_h, pool_w), 1)(x, rois)
205

206
207
208
209
    def make_obj(self, pool_h=5, pool_w=5, spatial_scale=1, wrap=False):
        obj = ops.PSRoIPool((pool_h, pool_w), spatial_scale)
        return RoIOpTesterModuleWrapper(obj) if wrap else obj

210
    def get_script_fn(self, rois, pool_size):
Nicolas Hug's avatar
Nicolas Hug committed
211
212
        scriped = torch.jit.script(ops.ps_roi_pool)
        return lambda x: scriped(x, rois, pool_size)
213

214
215
216
    def expected_fn(
        self, x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, device=None, dtype=torch.float64
    ):
217
218
219
        if device is None:
            device = torch.device("cpu")
        n_input_channels = x.size(1)
220
        assert n_input_channels % (pool_h * pool_w) == 0, "input channels must be divisible by ph * pw"
221
222
223
224
225
226
227
228
229
        n_output_channels = int(n_input_channels / (pool_h * pool_w))
        y = torch.zeros(rois.size(0), n_output_channels, pool_h, pool_w, dtype=dtype, device=device)

        def get_slice(k, block):
            return slice(int(np.floor(k * block)), int(np.ceil((k + 1) * block)))

        for roi_idx, roi in enumerate(rois):
            batch_idx = int(roi[0])
            j_begin, i_begin, j_end, i_end = (int(round(x.item() * spatial_scale)) for x in roi[1:])
230
            roi_x = x[batch_idx, :, i_begin : i_end + 1, j_begin : j_end + 1]
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245

            roi_height = max(i_end - i_begin, 1)
            roi_width = max(j_end - j_begin, 1)
            bin_h, bin_w = roi_height / float(pool_h), roi_width / float(pool_w)

            for i in range(0, pool_h):
                for j in range(0, pool_w):
                    bin_x = roi_x[:, get_slice(i, bin_h), get_slice(j, bin_w)]
                    if bin_x.numel() > 0:
                        area = bin_x.size(-2) * bin_x.size(-1)
                        for c_out in range(0, n_output_channels):
                            c_in = c_out * (pool_h * pool_w) + pool_w * i + j
                            t = torch.sum(bin_x[c_in, :, :])
                            y[roi_idx, c_out, i, j] = t / area
        return y
246

247
    def test_boxes_shape(self):
248
249
        self._helper_boxes_shape(ops.ps_roi_pool)

250

251
252
def bilinear_interpolate(data, y, x, snap_border=False):
    height, width = data.shape
253

254
255
256
257
258
    if snap_border:
        if -1 < y <= 0:
            y = 0
        elif height - 1 <= y < height:
            y = height - 1
259

260
261
262
263
        if -1 < x <= 0:
            x = 0
        elif width - 1 <= x < width:
            x = width - 1
264

265
266
267
268
    y_low = int(math.floor(y))
    x_low = int(math.floor(x))
    y_high = y_low + 1
    x_high = x_low + 1
269

270
271
    wy_h = y - y_low
    wx_h = x - x_low
272
    wy_l = 1 - wy_h
273
    wx_l = 1 - wx_h
274

275
    val = 0
276
277
278
279
    for wx, xp in zip((wx_l, wx_h), (x_low, x_high)):
        for wy, yp in zip((wy_l, wy_h), (y_low, y_high)):
            if 0 <= yp < height and 0 <= xp < width:
                val += wx * wy * data[yp, xp]
280
    return val
281
282


283
class TestRoIAlign(RoIOpTester):
AhnDW's avatar
AhnDW committed
284
    def fn(self, x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, aligned=False, **kwargs):
285
286
287
        return ops.RoIAlign(
            (pool_h, pool_w), spatial_scale=spatial_scale, sampling_ratio=sampling_ratio, aligned=aligned
        )(x, rois)
288

289
290
291
292
293
294
    def make_obj(self, pool_h=5, pool_w=5, spatial_scale=1, sampling_ratio=-1, aligned=False, wrap=False):
        obj = ops.RoIAlign(
            (pool_h, pool_w), spatial_scale=spatial_scale, sampling_ratio=sampling_ratio, aligned=aligned
        )
        return RoIOpTesterModuleWrapper(obj) if wrap else obj

295
    def get_script_fn(self, rois, pool_size):
Nicolas Hug's avatar
Nicolas Hug committed
296
297
        scriped = torch.jit.script(ops.roi_align)
        return lambda x: scriped(x, rois, pool_size)
298

299
300
301
302
303
304
305
306
307
308
309
310
    def expected_fn(
        self,
        in_data,
        rois,
        pool_h,
        pool_w,
        spatial_scale=1,
        sampling_ratio=-1,
        aligned=False,
        device=None,
        dtype=torch.float64,
    ):
311
312
        if device is None:
            device = torch.device("cpu")
313
314
315
        n_channels = in_data.size(1)
        out_data = torch.zeros(rois.size(0), n_channels, pool_h, pool_w, dtype=dtype, device=device)

316
        offset = 0.5 if aligned else 0.0
AhnDW's avatar
AhnDW committed
317

318
319
        for r, roi in enumerate(rois):
            batch_idx = int(roi[0])
AhnDW's avatar
AhnDW committed
320
            j_begin, i_begin, j_end, i_end = (x.item() * spatial_scale - offset for x in roi[1:])
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340

            roi_h = i_end - i_begin
            roi_w = j_end - j_begin
            bin_h = roi_h / pool_h
            bin_w = roi_w / pool_w

            for i in range(0, pool_h):
                start_h = i_begin + i * bin_h
                grid_h = sampling_ratio if sampling_ratio > 0 else int(np.ceil(bin_h))
                for j in range(0, pool_w):
                    start_w = j_begin + j * bin_w
                    grid_w = sampling_ratio if sampling_ratio > 0 else int(np.ceil(bin_w))

                    for channel in range(0, n_channels):

                        val = 0
                        for iy in range(0, grid_h):
                            y = start_h + (iy + 0.5) * bin_h / grid_h
                            for ix in range(0, grid_w):
                                x = start_w + (ix + 0.5) * bin_w / grid_w
341
                                val += bilinear_interpolate(in_data[batch_idx, channel, :, :], y, x, snap_border=True)
342
343
344
                        val /= grid_h * grid_w

                        out_data[r, channel, i, j] = val
345
346
        return out_data

347
    def test_boxes_shape(self):
348
349
        self._helper_boxes_shape(ops.roi_align)

350
351
352
    @pytest.mark.parametrize("aligned", (True, False))
    @pytest.mark.parametrize("device", cpu_and_gpu())
    @pytest.mark.parametrize("contiguous", (True, False))
353
    def test_forward(self, device, contiguous, aligned, x_dtype=None, rois_dtype=None):
354
355
356
        super().test_forward(
            device=device, contiguous=contiguous, x_dtype=x_dtype, rois_dtype=rois_dtype, aligned=aligned
        )
357

358
    @needs_cuda
359
360
361
    @pytest.mark.parametrize("aligned", (True, False))
    @pytest.mark.parametrize("x_dtype", (torch.float, torch.half))
    @pytest.mark.parametrize("rois_dtype", (torch.float, torch.half))
362
363
    def test_autocast(self, aligned, x_dtype, rois_dtype):
        with torch.cuda.amp.autocast():
364
365
366
            self.test_forward(
                torch.device("cuda"), contiguous=False, aligned=aligned, x_dtype=x_dtype, rois_dtype=rois_dtype
            )
367
368
369
370
371
372
373

    def _make_rois(self, img_size, num_imgs, dtype, num_rois=1000):
        rois = torch.randint(0, img_size // 2, size=(num_rois, 5)).to(dtype)
        rois[:, 0] = torch.randint(0, num_imgs, size=(num_rois,))  # set batch index
        rois[:, 3:] += rois[:, 1:3]  # make sure boxes aren't degenerate
        return rois

374
375
376
    @pytest.mark.parametrize("aligned", (True, False))
    @pytest.mark.parametrize("scale, zero_point", ((1, 0), (2, 10), (0.1, 50)))
    @pytest.mark.parametrize("qdtype", (torch.qint8, torch.quint8, torch.qint32))
377
    def test_qroialign(self, aligned, scale, zero_point, qdtype):
378
379
380
381
382
383
384
        """Make sure quantized version of RoIAlign is close to float version"""
        pool_size = 5
        img_size = 10
        n_channels = 2
        num_imgs = 1
        dtype = torch.float

385
386
387
388
389
390
391
        x = torch.randint(50, 100, size=(num_imgs, n_channels, img_size, img_size)).to(dtype)
        qx = torch.quantize_per_tensor(x, scale=scale, zero_point=zero_point, dtype=qdtype)

        rois = self._make_rois(img_size, num_imgs, dtype)
        qrois = torch.quantize_per_tensor(rois, scale=scale, zero_point=zero_point, dtype=qdtype)

        x, rois = qx.dequantize(), qrois.dequantize()  # we want to pass the same inputs
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
        y = ops.roi_align(
            x,
            rois,
            output_size=pool_size,
            spatial_scale=1,
            sampling_ratio=-1,
            aligned=aligned,
        )
        qy = ops.roi_align(
            qx,
            qrois,
            output_size=pool_size,
            spatial_scale=1,
            sampling_ratio=-1,
            aligned=aligned,
        )

        # The output qy is itself a quantized tensor and there might have been a loss of info when it was
        # quantized. For a fair comparison we need to quantize y as well
        quantized_float_y = torch.quantize_per_tensor(y, scale=scale, zero_point=zero_point, dtype=qdtype)

        try:
            # Ideally, we would assert this, which passes with (scale, zero) == (1, 0)
            assert (qy == quantized_float_y).all()
        except AssertionError:
            # But because the computation aren't exactly the same between the 2 RoIAlign procedures, some
            # rounding error may lead to a difference of 2 in the output.
            # For example with (scale, zero) = (2, 10), 45.00000... will be quantized to 44
            # but 45.00000001 will be rounded to 46. We make sure below that:
            # - such discrepancies between qy and quantized_float_y are very rare (less then 5%)
            # - any difference between qy and quantized_float_y is == scale
            diff_idx = torch.where(qy != quantized_float_y)
            num_diff = diff_idx[0].numel()
426
            assert num_diff / qy.numel() < 0.05
427
428
429
430
431
432
433

            abs_diff = torch.abs(qy[diff_idx].dequantize() - quantized_float_y[diff_idx].dequantize())
            t_scale = torch.full_like(abs_diff, fill_value=scale)
            torch.testing.assert_close(abs_diff, t_scale, rtol=1e-5, atol=1e-5)

    def test_qroi_align_multiple_images(self):
        dtype = torch.float
434
435
        x = torch.randint(50, 100, size=(2, 3, 10, 10)).to(dtype)
        qx = torch.quantize_per_tensor(x, scale=1, zero_point=0, dtype=torch.qint8)
436
        rois = self._make_rois(img_size=10, num_imgs=2, dtype=dtype, num_rois=10)
437
        qrois = torch.quantize_per_tensor(rois, scale=1, zero_point=0, dtype=torch.qint8)
438
439
        with pytest.raises(RuntimeError, match="Only one image per batch is allowed"):
            ops.roi_align(qx, qrois, output_size=5)
440

441

442
class TestPSRoIAlign(RoIOpTester):
443
    def fn(self, x, rois, pool_h, pool_w, spatial_scale=1, sampling_ratio=-1, **kwargs):
444
        return ops.PSRoIAlign((pool_h, pool_w), spatial_scale=spatial_scale, sampling_ratio=sampling_ratio)(x, rois)
445

446
447
448
449
    def make_obj(self, pool_h=5, pool_w=5, spatial_scale=1, sampling_ratio=-1, wrap=False):
        obj = ops.PSRoIAlign((pool_h, pool_w), spatial_scale=spatial_scale, sampling_ratio=sampling_ratio)
        return RoIOpTesterModuleWrapper(obj) if wrap else obj

450
    def get_script_fn(self, rois, pool_size):
Nicolas Hug's avatar
Nicolas Hug committed
451
452
        scriped = torch.jit.script(ops.ps_roi_align)
        return lambda x: scriped(x, rois, pool_size)
453

454
455
456
    def expected_fn(
        self, in_data, rois, pool_h, pool_w, device, spatial_scale=1, sampling_ratio=-1, dtype=torch.float64
    ):
457
458
        if device is None:
            device = torch.device("cpu")
459
        n_input_channels = in_data.size(1)
460
        assert n_input_channels % (pool_h * pool_w) == 0, "input channels must be divisible by ph * pw"
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
        n_output_channels = int(n_input_channels / (pool_h * pool_w))
        out_data = torch.zeros(rois.size(0), n_output_channels, pool_h, pool_w, dtype=dtype, device=device)

        for r, roi in enumerate(rois):
            batch_idx = int(roi[0])
            j_begin, i_begin, j_end, i_end = (x.item() * spatial_scale - 0.5 for x in roi[1:])

            roi_h = i_end - i_begin
            roi_w = j_end - j_begin
            bin_h = roi_h / pool_h
            bin_w = roi_w / pool_w

            for i in range(0, pool_h):
                start_h = i_begin + i * bin_h
                grid_h = sampling_ratio if sampling_ratio > 0 else int(np.ceil(bin_h))
                for j in range(0, pool_w):
                    start_w = j_begin + j * bin_w
                    grid_w = sampling_ratio if sampling_ratio > 0 else int(np.ceil(bin_w))
                    for c_out in range(0, n_output_channels):
                        c_in = c_out * (pool_h * pool_w) + pool_w * i + j

                        val = 0
                        for iy in range(0, grid_h):
                            y = start_h + (iy + 0.5) * bin_h / grid_h
                            for ix in range(0, grid_w):
                                x = start_w + (ix + 0.5) * bin_w / grid_w
487
                                val += bilinear_interpolate(in_data[batch_idx, c_in, :, :], y, x, snap_border=True)
488
489
490
491
                        val /= grid_h * grid_w

                        out_data[r, c_out, i, j] = val
        return out_data
492

493
    def test_boxes_shape(self):
494
495
        self._helper_boxes_shape(ops.ps_roi_align)

496

497
class TestMultiScaleRoIAlign:
498
499
500
501
502
503
    def make_obj(self, fmap_names=None, output_size=(7, 7), sampling_ratio=2, wrap=False):
        if fmap_names is None:
            fmap_names = ["0"]
        obj = ops.poolers.MultiScaleRoIAlign(fmap_names, output_size, sampling_ratio)
        return MultiScaleRoIAlignModuleWrapper(obj) if wrap else obj

504
    def test_msroialign_repr(self):
505
        fmap_names = ["0"]
506
507
508
        output_size = (7, 7)
        sampling_ratio = 2
        # Pass mock feature map names
509
        t = self.make_obj(fmap_names, output_size, sampling_ratio, wrap=False)
510
511

        # Check integrity of object __repr__ attribute
512
513
514
515
        expected_string = (
            f"MultiScaleRoIAlign(featmap_names={fmap_names}, output_size={output_size}, "
            f"sampling_ratio={sampling_ratio})"
        )
516
        assert repr(t) == expected_string
517

518
519
520
521
522
523
524
525
526
    @pytest.mark.parametrize("device", cpu_and_gpu())
    def test_is_leaf_node(self, device):
        op_obj = self.make_obj(wrap=True).to(device=device)
        graph_node_names = get_graph_node_names(op_obj)

        assert len(graph_node_names) == 2
        assert len(graph_node_names[0]) == len(graph_node_names[1])
        assert len(graph_node_names[0]) == 1 + op_obj.n_inputs

527

528
529
class TestNMS:
    def _reference_nms(self, boxes, scores, iou_threshold):
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
        """
        Args:
            box_scores (N, 5): boxes in corner-form and probabilities.
            iou_threshold: intersection over union threshold.
        Returns:
             picked: a list of indexes of the kept boxes
        """
        picked = []
        _, indexes = scores.sort(descending=True)
        while len(indexes) > 0:
            current = indexes[0]
            picked.append(current.item())
            if len(indexes) == 1:
                break
            current_box = boxes[current, :]
            indexes = indexes[1:]
            rest_boxes = boxes[indexes, :]
            iou = ops.box_iou(rest_boxes, current_box.unsqueeze(0)).squeeze(1)
            indexes = indexes[iou <= iou_threshold]

        return torch.as_tensor(picked)

552
553
554
555
556
    def _create_tensors_with_iou(self, N, iou_thresh):
        # force last box to have a pre-defined iou with the first box
        # let b0 be [x0, y0, x1, y1], and b1 be [x0, y0, x1 + d, y1],
        # then, in order to satisfy ops.iou(b0, b1) == iou_thresh,
        # we need to have d = (x1 - x0) * (1 - iou_thresh) / iou_thresh
557
558
559
        # Adjust the threshold upward a bit with the intent of creating
        # at least one box that exceeds (barely) the threshold and so
        # should be suppressed.
560
        boxes = torch.rand(N, 4) * 100
561
562
563
        boxes[:, 2:] += boxes[:, :2]
        boxes[-1, :] = boxes[0, :]
        x0, y0, x1, y1 = boxes[-1].tolist()
564
        iou_thresh += 1e-5
565
        boxes[-1, 2] += (x1 - x0) * (1 - iou_thresh) / iou_thresh
566
567
568
        scores = torch.rand(N)
        return boxes, scores

569
    @pytest.mark.parametrize("iou", (0.2, 0.5, 0.8))
570
571
572
    @pytest.mark.parametrize("seed", range(10))
    def test_nms_ref(self, iou, seed):
        torch.random.manual_seed(seed)
573
        err_msg = "NMS incompatible between CPU and reference implementation for IoU={}"
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
        boxes, scores = self._create_tensors_with_iou(1000, iou)
        keep_ref = self._reference_nms(boxes, scores, iou)
        keep = ops.nms(boxes, scores, iou)
        assert torch.allclose(keep, keep_ref), err_msg.format(iou)

    def test_nms_input_errors(self):
        with pytest.raises(RuntimeError):
            ops.nms(torch.rand(4), torch.rand(3), 0.5)
        with pytest.raises(RuntimeError):
            ops.nms(torch.rand(3, 5), torch.rand(3), 0.5)
        with pytest.raises(RuntimeError):
            ops.nms(torch.rand(3, 4), torch.rand(3, 2), 0.5)
        with pytest.raises(RuntimeError):
            ops.nms(torch.rand(3, 4), torch.rand(4), 0.5)

589
    @pytest.mark.parametrize("iou", (0.2, 0.5, 0.8))
590
591
    @pytest.mark.parametrize("scale, zero_point", ((1, 0), (2, 50), (3, 10)))
    def test_qnms(self, iou, scale, zero_point):
592
593
594
        # Note: we compare qnms vs nms instead of qnms vs reference implementation.
        # This is because with the int convertion, the trick used in _create_tensors_with_iou
        # doesn't really work (in fact, nms vs reference implem will also fail with ints)
595
        err_msg = "NMS and QNMS give different results for IoU={}"
596
597
        boxes, scores = self._create_tensors_with_iou(1000, iou)
        scores *= 100  # otherwise most scores would be 0 or 1 after int convertion
598

599
600
        qboxes = torch.quantize_per_tensor(boxes, scale=scale, zero_point=zero_point, dtype=torch.quint8)
        qscores = torch.quantize_per_tensor(scores, scale=scale, zero_point=zero_point, dtype=torch.quint8)
601

602
603
        boxes = qboxes.dequantize()
        scores = qscores.dequantize()
604

605
606
        keep = ops.nms(boxes, scores, iou)
        qkeep = ops.nms(qboxes, qscores, iou)
607

608
        assert torch.allclose(qkeep, keep), err_msg.format(iou)
609

610
    @needs_cuda
611
    @pytest.mark.parametrize("iou", (0.2, 0.5, 0.8))
612
    def test_nms_cuda(self, iou, dtype=torch.float64):
613
        tol = 1e-3 if dtype is torch.half else 1e-5
614
        err_msg = "NMS incompatible between CPU and CUDA for IoU={}"
615

616
617
618
619
620
621
622
623
624
625
626
627
        boxes, scores = self._create_tensors_with_iou(1000, iou)
        r_cpu = ops.nms(boxes, scores, iou)
        r_cuda = ops.nms(boxes.cuda(), scores.cuda(), iou)

        is_eq = torch.allclose(r_cpu, r_cuda.cpu())
        if not is_eq:
            # if the indices are not the same, ensure that it's because the scores
            # are duplicate
            is_eq = torch.allclose(scores[r_cpu], scores[r_cuda.cpu()], rtol=tol, atol=tol)
        assert is_eq, err_msg.format(iou)

    @needs_cuda
628
    @pytest.mark.parametrize("iou", (0.2, 0.5, 0.8))
629
630
631
632
633
634
    @pytest.mark.parametrize("dtype", (torch.float, torch.half))
    def test_autocast(self, iou, dtype):
        with torch.cuda.amp.autocast():
            self.test_nms_cuda(iou=iou, dtype=dtype)

    @needs_cuda
635
    def test_nms_cuda_float16(self):
636
637
638
639
640
641
642
        boxes = torch.tensor(
            [
                [285.3538, 185.5758, 1193.5110, 851.4551],
                [285.1472, 188.7374, 1192.4984, 851.0669],
                [279.2440, 197.9812, 1189.4746, 849.2019],
            ]
        ).cuda()
643
644
645
646
647
        scores = torch.tensor([0.6370, 0.7569, 0.3966]).cuda()

        iou_thres = 0.2
        keep32 = ops.nms(boxes, scores, iou_thres)
        keep16 = ops.nms(boxes.to(torch.float16), scores.to(torch.float16), iou_thres)
648
        assert_equal(keep32, keep16)
649

650
651
    @pytest.mark.parametrize("seed", range(10))
    def test_batched_nms_implementations(self, seed):
652
        """Make sure that both implementations of batched_nms yield identical results"""
653
        torch.random.manual_seed(seed)
654
655

        num_boxes = 1000
656
        iou_threshold = 0.9
657
658
659
660
661
662
663
664
665
666

        boxes = torch.cat((torch.rand(num_boxes, 2), torch.rand(num_boxes, 2) + 10), dim=1)
        assert max(boxes[:, 0]) < min(boxes[:, 2])  # x1 < x2
        assert max(boxes[:, 1]) < min(boxes[:, 3])  # y1 < y2

        scores = torch.rand(num_boxes)
        idxs = torch.randint(0, 4, size=(num_boxes,))
        keep_vanilla = ops.boxes._batched_nms_vanilla(boxes, scores, idxs, iou_threshold)
        keep_trick = ops.boxes._batched_nms_coordinate_trick(boxes, scores, idxs, iou_threshold)

667
668
669
        torch.testing.assert_close(
            keep_vanilla, keep_trick, msg="The vanilla and the trick implementation yield different nms outputs."
        )
670
671
672

        # Also make sure an empty tensor is returned if boxes is empty
        empty = torch.empty((0,), dtype=torch.int64)
673
        torch.testing.assert_close(empty, ops.batched_nms(empty, None, None, None))
674

675

676
677
678
class TestDeformConv:
    dtype = torch.float64

679
    def expected_fn(self, x, weight, offset, mask, bias, stride=1, padding=0, dilation=1):
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
        stride_h, stride_w = _pair(stride)
        pad_h, pad_w = _pair(padding)
        dil_h, dil_w = _pair(dilation)
        weight_h, weight_w = weight.shape[-2:]

        n_batches, n_in_channels, in_h, in_w = x.shape
        n_out_channels = weight.shape[0]

        out_h = (in_h + 2 * pad_h - (dil_h * (weight_h - 1) + 1)) // stride_h + 1
        out_w = (in_w + 2 * pad_w - (dil_w * (weight_w - 1) + 1)) // stride_w + 1

        n_offset_grps = offset.shape[1] // (2 * weight_h * weight_w)
        in_c_per_offset_grp = n_in_channels // n_offset_grps

        n_weight_grps = n_in_channels // weight.shape[1]
        in_c_per_weight_grp = weight.shape[1]
        out_c_per_weight_grp = n_out_channels // n_weight_grps

        out = torch.zeros(n_batches, n_out_channels, out_h, out_w, device=x.device, dtype=x.dtype)
        for b in range(n_batches):
            for c_out in range(n_out_channels):
                for i in range(out_h):
                    for j in range(out_w):
                        for di in range(weight_h):
                            for dj in range(weight_w):
                                for c in range(in_c_per_weight_grp):
                                    weight_grp = c_out // out_c_per_weight_grp
                                    c_in = weight_grp * in_c_per_weight_grp + c

                                    offset_grp = c_in // in_c_per_offset_grp
710
711
                                    mask_idx = offset_grp * (weight_h * weight_w) + di * weight_w + dj
                                    offset_idx = 2 * mask_idx
712
713
714
715

                                    pi = stride_h * i - pad_h + dil_h * di + offset[b, offset_idx, i, j]
                                    pj = stride_w * j - pad_w + dil_w * dj + offset[b, offset_idx + 1, i, j]

716
717
718
719
                                    mask_value = 1.0
                                    if mask is not None:
                                        mask_value = mask[b, mask_idx, i, j]

720
721
722
723
724
                                    out[b, c_out, i, j] += (
                                        mask_value
                                        * weight[c_out, c, di, dj]
                                        * bilinear_interpolate(x[b, c_in, :, :], pi, pj)
                                    )
725
726
727
        out += bias.view(1, n_out_channels, 1, 1)
        return out

728
    @lru_cache(maxsize=None)
729
    def get_fn_args(self, device, contiguous, batch_sz, dtype):
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
        n_in_channels = 6
        n_out_channels = 2
        n_weight_grps = 2
        n_offset_grps = 3

        stride = (2, 1)
        pad = (1, 0)
        dilation = (2, 1)

        stride_h, stride_w = stride
        pad_h, pad_w = pad
        dil_h, dil_w = dilation
        weight_h, weight_w = (3, 2)
        in_h, in_w = (5, 4)

        out_h = (in_h + 2 * pad_h - (dil_h * (weight_h - 1) + 1)) // stride_h + 1
        out_w = (in_w + 2 * pad_w - (dil_w * (weight_w - 1) + 1)) // stride_w + 1

748
        x = torch.rand(batch_sz, n_in_channels, in_h, in_w, device=device, dtype=dtype, requires_grad=True)
749

750
751
752
753
754
755
756
757
758
        offset = torch.randn(
            batch_sz,
            n_offset_grps * 2 * weight_h * weight_w,
            out_h,
            out_w,
            device=device,
            dtype=dtype,
            requires_grad=True,
        )
759

760
761
762
        mask = torch.randn(
            batch_sz, n_offset_grps * weight_h * weight_w, out_h, out_w, device=device, dtype=dtype, requires_grad=True
        )
763

764
765
766
767
768
769
770
771
772
        weight = torch.randn(
            n_out_channels,
            n_in_channels // n_weight_grps,
            weight_h,
            weight_w,
            device=device,
            dtype=dtype,
            requires_grad=True,
        )
773

774
        bias = torch.randn(n_out_channels, device=device, dtype=dtype, requires_grad=True)
775
776
777
778

        if not contiguous:
            x = x.permute(0, 1, 3, 2).contiguous().permute(0, 1, 3, 2)
            offset = offset.permute(1, 3, 0, 2).contiguous().permute(2, 0, 3, 1)
779
            mask = mask.permute(1, 3, 0, 2).contiguous().permute(2, 0, 3, 1)
780
781
            weight = weight.permute(3, 2, 0, 1).contiguous().permute(2, 3, 1, 0)

782
        return x, weight, offset, mask, bias, stride, pad, dilation
783

784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
    def make_obj(self, in_channels=6, out_channels=2, kernel_size=(3, 2), groups=2, wrap=False):
        obj = ops.DeformConv2d(
            in_channels, out_channels, kernel_size, stride=(2, 1), padding=(1, 0), dilation=(2, 1), groups=groups
        )
        return DeformConvModuleWrapper(obj) if wrap else obj

    @pytest.mark.parametrize("device", cpu_and_gpu())
    def test_is_leaf_node(self, device):
        op_obj = self.make_obj(wrap=True).to(device=device)
        graph_node_names = get_graph_node_names(op_obj)

        assert len(graph_node_names) == 2
        assert len(graph_node_names[0]) == len(graph_node_names[1])
        assert len(graph_node_names[0]) == 1 + op_obj.n_inputs

799
800
801
    @pytest.mark.parametrize("device", cpu_and_gpu())
    @pytest.mark.parametrize("contiguous", (True, False))
    @pytest.mark.parametrize("batch_sz", (0, 33))
802
803
    def test_forward(self, device, contiguous, batch_sz, dtype=None):
        dtype = dtype or self.dtype
804
        x, _, offset, mask, _, stride, padding, dilation = self.get_fn_args(device, contiguous, batch_sz, dtype)
805
806
807
808
        in_channels = 6
        out_channels = 2
        kernel_size = (3, 2)
        groups = 2
Nicolas Hug's avatar
Nicolas Hug committed
809
        tol = 2e-3 if dtype is torch.half else 1e-5
810

811
812
813
        layer = self.make_obj(in_channels, out_channels, kernel_size, groups, wrap=False).to(
            device=x.device, dtype=dtype
        )
814
        res = layer(x, offset, mask)
815
816
817

        weight = layer.weight.data
        bias = layer.bias.data
818
819
        expected = self.expected_fn(x, weight, offset, mask, bias, stride=stride, padding=padding, dilation=dilation)

820
        torch.testing.assert_close(
821
            res.to(expected), expected, rtol=tol, atol=tol, msg=f"\nres:\n{res}\nexpected:\n{expected}"
822
        )
823
824
825
826

        # no modulation test
        res = layer(x, offset)
        expected = self.expected_fn(x, weight, offset, None, bias, stride=stride, padding=padding, dilation=dilation)
827

828
        torch.testing.assert_close(
829
            res.to(expected), expected, rtol=tol, atol=tol, msg=f"\nres:\n{res}\nexpected:\n{expected}"
830
        )
831

832
833
834
835
836
    def test_wrong_sizes(self):
        in_channels = 6
        out_channels = 2
        kernel_size = (3, 2)
        groups = 2
837
838
839
840
841
842
        x, _, offset, mask, _, stride, padding, dilation = self.get_fn_args(
            "cpu", contiguous=True, batch_sz=10, dtype=self.dtype
        )
        layer = ops.DeformConv2d(
            in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups
        )
843
        with pytest.raises(RuntimeError, match="the shape of the offset"):
844
            wrong_offset = torch.rand_like(offset[:, :2])
845
            layer(x, wrong_offset)
846

847
        with pytest.raises(RuntimeError, match=r"mask.shape\[1\] is not valid"):
848
            wrong_mask = torch.rand_like(mask[:, :2])
849
            layer(x, offset, wrong_mask)
850

851
852
853
    @pytest.mark.parametrize("device", cpu_and_gpu())
    @pytest.mark.parametrize("contiguous", (True, False))
    @pytest.mark.parametrize("batch_sz", (0, 33))
854
    def test_backward(self, device, contiguous, batch_sz):
855
856
857
        x, weight, offset, mask, bias, stride, padding, dilation = self.get_fn_args(
            device, contiguous, batch_sz, self.dtype
        )
858
859

        def func(x_, offset_, mask_, weight_, bias_):
860
861
862
            return ops.deform_conv2d(
                x_, offset_, weight_, bias_, stride=stride, padding=padding, dilation=dilation, mask=mask_
            )
863

864
        gradcheck(func, (x, offset, mask, weight, bias), nondet_tol=1e-5, fast_mode=True)
865
866

        def func_no_mask(x_, offset_, weight_, bias_):
867
868
869
            return ops.deform_conv2d(
                x_, offset_, weight_, bias_, stride=stride, padding=padding, dilation=dilation, mask=None
            )
870

871
        gradcheck(func_no_mask, (x, offset, weight, bias), nondet_tol=1e-5, fast_mode=True)
872
873
874
875

        @torch.jit.script
        def script_func(x_, offset_, mask_, weight_, bias_, stride_, pad_, dilation_):
            # type:(Tensor, Tensor, Tensor, Tensor, Tensor, Tuple[int, int], Tuple[int, int], Tuple[int, int])->Tensor
876
877
878
879
880
881
882
883
884
885
            return ops.deform_conv2d(
                x_, offset_, weight_, bias_, stride=stride_, padding=pad_, dilation=dilation_, mask=mask_
            )

        gradcheck(
            lambda z, off, msk, wei, bi: script_func(z, off, msk, wei, bi, stride, padding, dilation),
            (x, offset, mask, weight, bias),
            nondet_tol=1e-5,
            fast_mode=True,
        )
886
887

        @torch.jit.script
888
889
        def script_func_no_mask(x_, offset_, weight_, bias_, stride_, pad_, dilation_):
            # type:(Tensor, Tensor, Tensor, Tensor, Tuple[int, int], Tuple[int, int], Tuple[int, int])->Tensor
890
891
892
893
894
895
896
897
898
899
            return ops.deform_conv2d(
                x_, offset_, weight_, bias_, stride=stride_, padding=pad_, dilation=dilation_, mask=None
            )

        gradcheck(
            lambda z, off, wei, bi: script_func_no_mask(z, off, wei, bi, stride, padding, dilation),
            (x, offset, weight, bias),
            nondet_tol=1e-5,
            fast_mode=True,
        )
900

901
    @needs_cuda
902
    @pytest.mark.parametrize("contiguous", (True, False))
903
    def test_compare_cpu_cuda_grads(self, contiguous):
904
905
906
        # Test from https://github.com/pytorch/vision/issues/2598
        # Run on CUDA only

907
908
        # compare grads computed on CUDA with grads computed on CPU
        true_cpu_grads = None
909

910
911
912
913
        init_weight = torch.randn(9, 9, 3, 3, requires_grad=True)
        img = torch.randn(8, 9, 1000, 110)
        offset = torch.rand(8, 2 * 3 * 3, 1000, 110)
        mask = torch.rand(8, 3 * 3, 1000, 110)
914

915
916
917
918
919
920
921
        if not contiguous:
            img = img.permute(0, 1, 3, 2).contiguous().permute(0, 1, 3, 2)
            offset = offset.permute(1, 3, 0, 2).contiguous().permute(2, 0, 3, 1)
            mask = mask.permute(1, 3, 0, 2).contiguous().permute(2, 0, 3, 1)
            weight = init_weight.permute(3, 2, 0, 1).contiguous().permute(2, 3, 1, 0)
        else:
            weight = init_weight
922

923
        for d in ["cpu", "cuda"]:
924

925
926
927
928
929
930
931
932
933
934
935
            out = ops.deform_conv2d(img.to(d), offset.to(d), weight.to(d), padding=1, mask=mask.to(d))
            out.mean().backward()
            if true_cpu_grads is None:
                true_cpu_grads = init_weight.grad
                assert true_cpu_grads is not None
            else:
                assert init_weight.grad is not None
                res_grads = init_weight.grad.to("cpu")
                torch.testing.assert_close(true_cpu_grads, res_grads)

    @needs_cuda
936
937
    @pytest.mark.parametrize("batch_sz", (0, 33))
    @pytest.mark.parametrize("dtype", (torch.float, torch.half))
938
939
940
941
    def test_autocast(self, batch_sz, dtype):
        with torch.cuda.amp.autocast():
            self.test_forward(torch.device("cuda"), contiguous=False, batch_sz=batch_sz, dtype=dtype)

942
943
944
945
    def test_forward_scriptability(self):
        # Non-regression test for https://github.com/pytorch/vision/issues/4078
        torch.jit.script(ops.DeformConv2d(in_channels=8, out_channels=8, kernel_size=3))

946
947

class TestFrozenBNT:
948
949
    def test_frozenbatchnorm2d_repr(self):
        num_features = 32
950
951
        eps = 1e-5
        t = ops.misc.FrozenBatchNorm2d(num_features, eps=eps)
952
953

        # Check integrity of object __repr__ attribute
954
        expected_string = f"FrozenBatchNorm2d({num_features}, eps={eps})"
955
        assert repr(t) == expected_string
956

957
958
959
    @pytest.mark.parametrize("seed", range(10))
    def test_frozenbatchnorm2d_eps(self, seed):
        torch.random.manual_seed(seed)
960
961
        sample_size = (4, 32, 28, 28)
        x = torch.rand(sample_size)
962
963
964
965
966
967
968
        state_dict = dict(
            weight=torch.rand(sample_size[1]),
            bias=torch.rand(sample_size[1]),
            running_mean=torch.rand(sample_size[1]),
            running_var=torch.rand(sample_size[1]),
            num_batches_tracked=torch.tensor(100),
        )
969

970
        # Check that default eps is equal to the one of BN
971
972
        fbn = ops.misc.FrozenBatchNorm2d(sample_size[1])
        fbn.load_state_dict(state_dict, strict=False)
973
        bn = torch.nn.BatchNorm2d(sample_size[1]).eval()
974
975
        bn.load_state_dict(state_dict)
        # Difference is expected to fall in an acceptable range
976
        torch.testing.assert_close(fbn(x), bn(x), rtol=1e-5, atol=1e-6)
977
978
979
980
981
982

        # Check computation for eps > 0
        fbn = ops.misc.FrozenBatchNorm2d(sample_size[1], eps=1e-5)
        fbn.load_state_dict(state_dict, strict=False)
        bn = torch.nn.BatchNorm2d(sample_size[1], eps=1e-5).eval()
        bn.load_state_dict(state_dict)
983
        torch.testing.assert_close(fbn(x), bn(x), rtol=1e-5, atol=1e-6)
984
985
986
987

    def test_frozenbatchnorm2d_n_arg(self):
        """Ensure a warning is thrown when passing `n` kwarg
        (remove this when support of `n` is dropped)"""
988
989
        with pytest.warns(DeprecationWarning):
            ops.misc.FrozenBatchNorm2d(32, eps=1e-5, n=32)
990

991

992
class TestBoxConversion:
993
994
995
    def _get_box_sequences():
        # Define here the argument type of `boxes` supported by region pooling operations
        box_tensor = torch.tensor([[0, 0, 0, 100, 100], [1, 0, 0, 100, 100]], dtype=torch.float)
996
997
998
999
        box_list = [
            torch.tensor([[0, 0, 100, 100]], dtype=torch.float),
            torch.tensor([[0, 0, 100, 100]], dtype=torch.float),
        ]
1000
1001
1002
        box_tuple = tuple(box_list)
        return box_tensor, box_list, box_tuple

1003
    @pytest.mark.parametrize("box_sequence", _get_box_sequences())
1004
    def test_check_roi_boxes_shape(self, box_sequence):
1005
        # Ensure common sequences of tensors are supported
1006
        ops._utils.check_roi_boxes_shape(box_sequence)
1007

1008
    @pytest.mark.parametrize("box_sequence", _get_box_sequences())
1009
    def test_convert_boxes_to_roi_format(self, box_sequence):
1010
1011
        # Ensure common sequences of tensors yield the same result
        ref_tensor = None
1012
1013
1014
1015
        if ref_tensor is None:
            ref_tensor = box_sequence
        else:
            assert_equal(ref_tensor, ops._utils.convert_boxes_to_roi_format(box_sequence))
1016
1017


1018
class TestBox:
1019
    def test_bbox_same(self):
1020
1021
1022
        box_tensor = torch.tensor(
            [[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 30, 35], [23, 35, 93, 95]], dtype=torch.float
        )
1023

1024
        exp_xyxy = torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 30, 35], [23, 35, 93, 95]], dtype=torch.float)
1025

1026
1027
1028
1029
        assert exp_xyxy.size() == torch.Size([4, 4])
        assert_equal(ops.box_convert(box_tensor, in_fmt="xyxy", out_fmt="xyxy"), exp_xyxy)
        assert_equal(ops.box_convert(box_tensor, in_fmt="xywh", out_fmt="xywh"), exp_xyxy)
        assert_equal(ops.box_convert(box_tensor, in_fmt="cxcywh", out_fmt="cxcywh"), exp_xyxy)
1030
1031
1032
1033

    def test_bbox_xyxy_xywh(self):
        # Simple test convert boxes to xywh and back. Make sure they are same.
        # box_tensor is in x1 y1 x2 y2 format.
1034
1035
1036
1037
        box_tensor = torch.tensor(
            [[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 30, 35], [23, 35, 93, 95]], dtype=torch.float
        )
        exp_xywh = torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 20, 20], [23, 35, 70, 60]], dtype=torch.float)
1038

1039
        assert exp_xywh.size() == torch.Size([4, 4])
1040
        box_xywh = ops.box_convert(box_tensor, in_fmt="xyxy", out_fmt="xywh")
1041
        assert_equal(box_xywh, exp_xywh)
1042
1043
1044

        # Reverse conversion
        box_xyxy = ops.box_convert(box_xywh, in_fmt="xywh", out_fmt="xyxy")
1045
        assert_equal(box_xyxy, box_tensor)
1046
1047
1048
1049

    def test_bbox_xyxy_cxcywh(self):
        # Simple test convert boxes to xywh and back. Make sure they are same.
        # box_tensor is in x1 y1 x2 y2 format.
1050
1051
1052
1053
1054
1055
        box_tensor = torch.tensor(
            [[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 30, 35], [23, 35, 93, 95]], dtype=torch.float
        )
        exp_cxcywh = torch.tensor(
            [[50, 50, 100, 100], [0, 0, 0, 0], [20, 25, 20, 20], [58, 65, 70, 60]], dtype=torch.float
        )
1056

1057
        assert exp_cxcywh.size() == torch.Size([4, 4])
1058
        box_cxcywh = ops.box_convert(box_tensor, in_fmt="xyxy", out_fmt="cxcywh")
1059
        assert_equal(box_cxcywh, exp_cxcywh)
1060
1061
1062

        # Reverse conversion
        box_xyxy = ops.box_convert(box_cxcywh, in_fmt="cxcywh", out_fmt="xyxy")
1063
        assert_equal(box_xyxy, box_tensor)
1064
1065

    def test_bbox_xywh_cxcywh(self):
1066
1067
1068
        box_tensor = torch.tensor(
            [[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 20, 20], [23, 35, 70, 60]], dtype=torch.float
        )
1069
1070

        # This is wrong
1071
1072
1073
        exp_cxcywh = torch.tensor(
            [[50, 50, 100, 100], [0, 0, 0, 0], [20, 25, 20, 20], [58, 65, 70, 60]], dtype=torch.float
        )
1074

1075
        assert exp_cxcywh.size() == torch.Size([4, 4])
1076
        box_cxcywh = ops.box_convert(box_tensor, in_fmt="xywh", out_fmt="cxcywh")
1077
        assert_equal(box_cxcywh, exp_cxcywh)
1078
1079
1080

        # Reverse conversion
        box_xywh = ops.box_convert(box_cxcywh, in_fmt="cxcywh", out_fmt="xywh")
1081
        assert_equal(box_xywh, box_tensor)
1082

1083
1084
    @pytest.mark.parametrize("inv_infmt", ["xwyh", "cxwyh"])
    @pytest.mark.parametrize("inv_outfmt", ["xwcx", "xhwcy"])
1085
    def test_bbox_invalid(self, inv_infmt, inv_outfmt):
1086
1087
1088
        box_tensor = torch.tensor(
            [[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 20, 20], [23, 35, 70, 60]], dtype=torch.float
        )
1089

1090
1091
        with pytest.raises(ValueError):
            ops.box_convert(box_tensor, inv_infmt, inv_outfmt)
1092
1093

    def test_bbox_convert_jit(self):
1094
1095
1096
        box_tensor = torch.tensor(
            [[0, 0, 100, 100], [0, 0, 0, 0], [10, 15, 30, 35], [23, 35, 93, 95]], dtype=torch.float
        )
1097

1098
1099
        scripted_fn = torch.jit.script(ops.box_convert)
        TOLERANCE = 1e-3
1100

1101
        box_xywh = ops.box_convert(box_tensor, in_fmt="xyxy", out_fmt="xywh")
1102
        scripted_xywh = scripted_fn(box_tensor, "xyxy", "xywh")
1103
        torch.testing.assert_close(scripted_xywh, box_xywh, rtol=0.0, atol=TOLERANCE)
1104

1105
        box_cxcywh = ops.box_convert(box_tensor, in_fmt="xyxy", out_fmt="cxcywh")
1106
        scripted_cxcywh = scripted_fn(box_tensor, "xyxy", "cxcywh")
1107
        torch.testing.assert_close(scripted_cxcywh, box_cxcywh, rtol=0.0, atol=TOLERANCE)
1108
1109


1110
class TestBoxArea:
Aditya Oke's avatar
Aditya Oke committed
1111
    def test_box_area(self):
1112
1113
        def area_check(box, expected, tolerance=1e-4):
            out = ops.box_area(box)
1114
            torch.testing.assert_close(out, expected, rtol=0.0, check_dtype=False, atol=tolerance)
1115
1116
1117
1118
1119
1120
1121
1122
1123

        # Check for int boxes
        for dtype in [torch.int8, torch.int16, torch.int32, torch.int64]:
            box_tensor = torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0]], dtype=dtype)
            expected = torch.tensor([10000, 0])
            area_check(box_tensor, expected)

        # Check for float32 and float64 boxes
        for dtype in [torch.float32, torch.float64]:
1124
1125
1126
1127
1128
1129
1130
1131
            box_tensor = torch.tensor(
                [
                    [285.3538, 185.5758, 1193.5110, 851.4551],
                    [285.1472, 188.7374, 1192.4984, 851.0669],
                    [279.2440, 197.9812, 1189.4746, 849.2019],
                ],
                dtype=dtype,
            )
1132
1133
1134
1135
            expected = torch.tensor([604723.0806, 600965.4666, 592761.0085], dtype=torch.float64)
            area_check(box_tensor, expected, tolerance=0.05)

        # Check for float16 box
1136
1137
1138
1139
        box_tensor = torch.tensor(
            [[285.25, 185.625, 1194.0, 851.5], [285.25, 188.75, 1192.0, 851.0], [279.25, 198.0, 1189.0, 849.0]],
            dtype=torch.float16,
        )
1140
1141
        expected = torch.tensor([605113.875, 600495.1875, 592247.25])
        area_check(box_tensor, expected)
Aditya Oke's avatar
Aditya Oke committed
1142

Aditya Oke's avatar
Aditya Oke committed
1143
1144
1145
1146
1147
1148
1149
1150
    def test_box_area_jit(self):
        box_tensor = torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0]], dtype=torch.float)
        TOLERANCE = 1e-3
        expected = ops.box_area(box_tensor)
        scripted_fn = torch.jit.script(ops.box_area)
        scripted_area = scripted_fn(box_tensor)
        torch.testing.assert_close(scripted_area, expected, rtol=0.0, atol=TOLERANCE)

Aditya Oke's avatar
Aditya Oke committed
1151

1152
class TestBoxIou:
Aditya Oke's avatar
Aditya Oke committed
1153
    def test_iou(self):
1154
1155
        def iou_check(box, expected, tolerance=1e-4):
            out = ops.box_iou(box, box)
1156
            torch.testing.assert_close(out, expected, rtol=0.0, check_dtype=False, atol=tolerance)
1157
1158
1159
1160
1161
1162
1163
1164
1165

        # Check for int boxes
        for dtype in [torch.int16, torch.int32, torch.int64]:
            box = torch.tensor([[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]], dtype=dtype)
            expected = torch.tensor([[1.0, 0.25, 0.0], [0.25, 1.0, 0.0], [0.0, 0.0, 1.0]])
            iou_check(box, expected)

        # Check for float boxes
        for dtype in [torch.float16, torch.float32, torch.float64]:
1166
1167
1168
1169
1170
1171
1172
1173
            box_tensor = torch.tensor(
                [
                    [285.3538, 185.5758, 1193.5110, 851.4551],
                    [285.1472, 188.7374, 1192.4984, 851.0669],
                    [279.2440, 197.9812, 1189.4746, 849.2019],
                ],
                dtype=dtype,
            )
1174
1175
            expected = torch.tensor([[1.0, 0.9933, 0.9673], [0.9933, 1.0, 0.9737], [0.9673, 0.9737, 1.0]])
            iou_check(box_tensor, expected, tolerance=0.002 if dtype == torch.float16 else 1e-4)
Aditya Oke's avatar
Aditya Oke committed
1176

Aditya Oke's avatar
Aditya Oke committed
1177
1178
1179
1180
1181
1182
1183
1184
    def test_iou_jit(self):
        box_tensor = torch.tensor([[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]], dtype=torch.float)
        TOLERANCE = 1e-3
        expected = ops.box_iou(box_tensor, box_tensor)
        scripted_fn = torch.jit.script(ops.box_iou)
        scripted_iou = scripted_fn(box_tensor, box_tensor)
        torch.testing.assert_close(scripted_iou, expected, rtol=0.0, atol=TOLERANCE)

Aditya Oke's avatar
Aditya Oke committed
1185

1186
class TestGenBoxIou:
Aditya Oke's avatar
Aditya Oke committed
1187
    def test_gen_iou(self):
1188
1189
        def gen_iou_check(box, expected, tolerance=1e-4):
            out = ops.generalized_box_iou(box, box)
1190
            torch.testing.assert_close(out, expected, rtol=0.0, check_dtype=False, atol=tolerance)
1191
1192
1193
1194
1195
1196
1197
1198
1199

        # Check for int boxes
        for dtype in [torch.int16, torch.int32, torch.int64]:
            box = torch.tensor([[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]], dtype=dtype)
            expected = torch.tensor([[1.0, 0.25, -0.7778], [0.25, 1.0, -0.8611], [-0.7778, -0.8611, 1.0]])
            gen_iou_check(box, expected)

        # Check for float boxes
        for dtype in [torch.float16, torch.float32, torch.float64]:
1200
1201
1202
1203
1204
1205
1206
1207
            box_tensor = torch.tensor(
                [
                    [285.3538, 185.5758, 1193.5110, 851.4551],
                    [285.1472, 188.7374, 1192.4984, 851.0669],
                    [279.2440, 197.9812, 1189.4746, 849.2019],
                ],
                dtype=dtype,
            )
1208
1209
            expected = torch.tensor([[1.0, 0.9933, 0.9673], [0.9933, 1.0, 0.9737], [0.9673, 0.9737, 1.0]])
            gen_iou_check(box_tensor, expected, tolerance=0.002 if dtype == torch.float16 else 1e-3)
Aditya Oke's avatar
Aditya Oke committed
1210

Aditya Oke's avatar
Aditya Oke committed
1211
1212
1213
1214
1215
1216
1217
1218
    def test_giou_jit(self):
        box_tensor = torch.tensor([[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]], dtype=torch.float)
        TOLERANCE = 1e-3
        expected = ops.generalized_box_iou(box_tensor, box_tensor)
        scripted_fn = torch.jit.script(ops.generalized_box_iou)
        scripted_iou = scripted_fn(box_tensor, box_tensor)
        torch.testing.assert_close(scripted_iou, expected, rtol=0.0, atol=TOLERANCE)

Aditya Oke's avatar
Aditya Oke committed
1219

1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
class TestMasksToBoxes:
    def test_masks_box(self):
        def masks_box_check(masks, expected, tolerance=1e-4):
            out = ops.masks_to_boxes(masks)
            assert out.dtype == torch.float
            torch.testing.assert_close(out, expected, rtol=0.0, check_dtype=False, atol=tolerance)

        # Check for int type boxes.
        def _get_image():
            assets_directory = os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets")
            mask_path = os.path.join(assets_directory, "masks.tiff")
            image = Image.open(mask_path)
            return image

        def _create_masks(image, masks):
            for index in range(image.n_frames):
                image.seek(index)
                frame = np.array(image)
                masks[index] = torch.tensor(frame)

            return masks

1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
        expected = torch.tensor(
            [
                [127, 2, 165, 40],
                [2, 50, 44, 92],
                [56, 63, 98, 100],
                [139, 68, 175, 104],
                [160, 112, 198, 145],
                [49, 138, 99, 182],
                [108, 148, 152, 213],
            ],
            dtype=torch.float,
        )
1254
1255
1256
1257
1258
1259
1260
1261

        image = _get_image()
        for dtype in [torch.float16, torch.float32, torch.float64]:
            masks = torch.zeros((image.n_frames, image.height, image.width), dtype=dtype)
            masks = _create_masks(image, masks)
            masks_box_check(masks, expected)


1262
class TestStochasticDepth:
1263
    @pytest.mark.parametrize("seed", range(10))
1264
1265
    @pytest.mark.parametrize("p", [0.2, 0.5, 0.8])
    @pytest.mark.parametrize("mode", ["batch", "row"])
1266
1267
    def test_stochastic_depth_random(self, seed, mode, p):
        torch.manual_seed(seed)
1268
1269
1270
        stats = pytest.importorskip("scipy.stats")
        batch_size = 5
        x = torch.ones(size=(batch_size, 3, 4, 4))
1271
        layer = ops.StochasticDepth(p=p, mode=mode)
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
        layer.__repr__()

        trials = 250
        num_samples = 0
        counts = 0
        for _ in range(trials):
            out = layer(x)
            non_zero_count = out.sum(dim=(1, 2, 3)).nonzero().size(0)
            if mode == "batch":
                if non_zero_count == 0:
                    counts += 1
                num_samples += 1
            elif mode == "row":
                counts += batch_size - non_zero_count
                num_samples += batch_size

        p_value = stats.binom_test(counts, num_samples, p=p)
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
        assert p_value > 0.01

    @pytest.mark.parametrize("seed", range(10))
    @pytest.mark.parametrize("p", (0, 1))
    @pytest.mark.parametrize("mode", ["batch", "row"])
    def test_stochastic_depth(self, seed, mode, p):
        torch.manual_seed(seed)
        batch_size = 5
        x = torch.ones(size=(batch_size, 3, 4, 4))
        layer = ops.StochasticDepth(p=p, mode=mode)

        out = layer(x)
        if p == 0:
            assert out.equal(x)
        elif p == 1:
            assert out.equal(torch.zeros_like(x))
1305

1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
    def make_obj(self, p, mode, wrap=False):
        obj = ops.StochasticDepth(p, mode)
        return StochasticDepthWrapper(obj) if wrap else obj

    @pytest.mark.parametrize("p", (0, 1))
    @pytest.mark.parametrize("mode", ["batch", "row"])
    def test_is_leaf_node(self, p, mode):
        op_obj = self.make_obj(p, mode, wrap=True)
        graph_node_names = get_graph_node_names(op_obj)

        assert len(graph_node_names) == 2
        assert len(graph_node_names[0]) == len(graph_node_names[1])
        assert len(graph_node_names[0]) == 1 + op_obj.n_inputs

1320

1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
class TestUtils:
    @pytest.mark.parametrize("norm_layer", [None, nn.BatchNorm2d, nn.LayerNorm])
    def test_split_normalization_params(self, norm_layer):
        model = models.mobilenet_v3_large(norm_layer=norm_layer)
        params = ops._utils.split_normalization_params(model, None if norm_layer is None else [norm_layer])

        assert len(params[0]) == 92
        assert len(params[1]) == 82


1331
if __name__ == "__main__":
1332
    pytest.main([__file__])