test_functional.py 32.8 KB
Newer Older
1 2 3
# -*- coding: utf-8 -*-
# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
#
4
# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
5 6 7 8 9
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
import itertools
10
import platform
11
from functools import partial
12 13 14

import numpy as np
import pytest
15
from utils import opr_test
16

17
import megengine.amp as amp
18
import megengine.core.ops.builtin as builtin
19 20
import megengine.core.tensor.dtype as dtype
import megengine.functional as F
21
import megengine.jit as jit
M
Megvii Engine Team 已提交
22
from megengine import Parameter, Tensor, is_cuda_available, tensor
23
from megengine.core._trace_option import use_symbolic_shape
24
from megengine.core.autodiff.grad import Grad
25
from megengine.core.tensor.utils import make_shape_tuple
26
from megengine.device import get_device_count
27 28


29 30 31 32 33 34 35 36 37 38 39 40 41
def test_where():
    maskv0 = np.array([[1, 0], [0, 1]], dtype=np.bool_)
    xv0 = np.array([[1, np.inf], [np.nan, 4]], dtype=np.float32)
    yv0 = np.array([[5, 6], [7, 8]], dtype=np.float32)

    maskv1 = np.array([[1, 0, 1], [1, 0, 0], [1, 1, 0]], dtype=np.bool_)
    xv1 = np.array([[1, np.inf, 2], [0, np.nan, 4], [1, 5, 7]], dtype=np.float32)
    yv1 = np.array([[5, 6, 9], [2, 7, 8], [2, 1, 9]], dtype=np.float32)

    cases = [
        {"input": [maskv0, xv0, yv0]},
        {"input": [maskv1, xv1, yv1]},
    ]
42
    opr_test(cases, F.where, ref_fn=np.where, test_trace=False)
43 44 45 46 47 48 49 50 51 52 53 54 55

    maskv2 = np.array([1, 1, 1], dtype=np.bool_)
    xv2 = np.array([1, 3, 2], dtype=np.float32)
    yv2 = np.array([5, 6, 9], dtype=np.float32)

    maskv3 = np.array([0, 0, 0], dtype=np.bool_)
    xv3 = np.array([1, 3, 2], dtype=np.float32)
    yv3 = np.array([5, 6, 9], dtype=np.float32)

    cases = [
        {"input": [maskv2, xv2, yv2]},
        {"input": [maskv3, xv3, yv3]},
    ]
56
    opr_test(cases, F.where, ref_fn=np.where, test_trace=False)
57 58


59
def test_dropout():
60 61 62 63 64 65 66 67
    # test training mode
    data = tensor(np.ones(10000000, dtype=np.float32))
    out = F.nn.dropout(data, 1.0 / 3.0, training=True)
    assert not out.numpy().all()

    # test eval mode
    out = F.nn.dropout(data, 1.0 / 3.0, training=False)
    assert out.numpy().all()
68 69


70 71 72 73 74
def test_matinv():
    shape1 = (5, 5)
    shape2 = (3, 9, 9)
    data1 = np.random.random(shape1).astype("float32")
    data2 = np.random.random(shape2).astype("float32")
M
Megvii Engine Team 已提交
75 76 77
    # make matrix diagonally dominant for numerical stability
    data1 += (np.eye(shape1[0]) * shape1[0]).astype("float32")
    data2 += np.broadcast_to((np.eye(shape2[1]) * shape2[1]).astype("float32"), shape2)
78 79 80 81 82 83 84 85 86

    cases = [
        {"input": data1},
        {"input": data2},
    ]

    opr_test(
        cases,
        F.matinv,
87
        compare_fn=lambda x, y: np.testing.assert_allclose(x.numpy(), y, rtol=1e-4),
88 89 90 91
        ref_fn=np.linalg.inv,
    )


92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
def test_matmul():
    shape1 = 3
    shape2 = 3
    shape3 = (3, 5)
    shape4 = (5, 6)
    data1 = np.random.random(shape1).astype("float32")
    data2 = np.random.random(shape2).astype("float32")
    data3 = np.random.random(shape3).astype("float32")
    data4 = np.random.random(shape4).astype("float32")

    cases = [
        {"input": [data1, data2]},
        {"input": [data2, data3]},
        {"input": [data3, data4]},
    ]
    opr_test(cases, F.matmul, ref_fn=np.matmul)

    batch_size = 10
110 111 112 113 114
    shape1 = (2,)
    shape2 = (batch_size, 2, 3)
    shape3 = (batch_size, 3, 4)
    shape4 = (batch_size, 10, 4, 2)
    shape5 = (batch_size, 10, 2, 4)
115 116 117
    data1 = np.random.random(shape1).astype("float32")
    data2 = np.random.random(shape2).astype("float32")
    data3 = np.random.random(shape3).astype("float32")
118 119
    data4 = np.random.random(shape4).astype("float32")
    data5 = np.random.random(shape5).astype("float32")
120

121 122 123 124 125 126
    cases = [
        {"input": [data1, data2]},
        {"input": [data2, data3]},
        {"input": [data3, data4]},
        {"input": [data4, data5]},
    ]
127
    opr_test(cases, F.matmul, ref_fn=np.matmul)
128

129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
    opr_test(
        [{"input": [data1, data4]}],
        F.matmul,
        ref_fn=lambda x, y: np.matmul(x, y.transpose(0, 1, 3, 2)),
        transpose_b=True,
    )

    opr_test(
        [{"input": [data3, data2]}],
        F.matmul,
        ref_fn=lambda x, y: np.matmul(x.transpose(0, 2, 1), y.transpose(0, 2, 1)),
        transpose_a=True,
        transpose_b=True,
    )

144

145 146 147 148
def test_interpolate():
    def linear_interpolate():
        inp = tensor(np.arange(1, 3, dtype=np.float32).reshape(1, 1, 2))

149 150
        out = F.vision.interpolate(inp, scale_factor=2.0, mode="linear")
        out2 = F.vision.interpolate(inp, 4, mode="linear")
151

152
        np.testing.assert_allclose(
153 154
            out.numpy(), np.array([[[1.0, 1.25, 1.75, 2.0]]], dtype=np.float32)
        )
155
        np.testing.assert_allclose(
156 157 158 159 160 161
            out2.numpy(), np.array([[[1.0, 1.25, 1.75, 2.0]]], dtype=np.float32)
        )

    def many_batch_interpolate():
        inp = tensor(np.arange(1, 9, dtype=np.float32).reshape(2, 1, 2, 2))

162 163
        out = F.vision.interpolate(inp, [4, 4])
        out2 = F.vision.interpolate(inp, scale_factor=2.0)
164

165
        np.testing.assert_allclose(out.numpy(), out2.numpy())
166 167 168 169

    def assign_corner_interpolate():
        inp = tensor(np.arange(1, 5, dtype=np.float32).reshape(1, 1, 2, 2))

170 171
        out = F.vision.interpolate(inp, [4, 4], align_corners=True)
        out2 = F.vision.interpolate(inp, scale_factor=2.0, align_corners=True)
172

173
        np.testing.assert_allclose(out.numpy(), out2.numpy())
174 175 176 177 178

    def error_shape_linear_interpolate():
        inp = tensor(np.arange(1, 5, dtype=np.float32).reshape(1, 1, 2, 2))

        with pytest.raises(ValueError):
179
            F.vision.interpolate(inp, scale_factor=2.0, mode="linear")
180 181 182 183 184

    def inappropriate_scale_linear_interpolate():
        inp = tensor(np.arange(1, 3, dtype=np.float32).reshape(1, 1, 2))

        with pytest.raises(ValueError):
185
            F.vision.interpolate(inp, scale_factor=[2.0, 3.0], mode="linear")
186 187 188 189 190 191 192 193 194

    linear_interpolate()
    many_batch_interpolate()
    assign_corner_interpolate()
    error_shape_linear_interpolate()
    inappropriate_scale_linear_interpolate()


def _save_to(self, name="grad"):
195
    def callback(grad):
196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217
        setattr(self, name, grad)

    return callback


def _gen_roi_inp():
    inp_feat = np.random.randn(2, 32, 256, 256)
    rois = np.zeros((4, 5))
    rois[:, 0] = [0, 0, 1, 1]
    rois[:, 1:3] = np.random.rand(4, 2) * 100
    rois[:, 3:] = np.random.rand(4, 2) * 100 + 150

    inp_feat = tensor(inp_feat)
    rois = tensor(rois)
    return inp_feat, rois


def test_roi_align():
    inp_feat, rois = _gen_roi_inp()
    grad = Grad().wrt(inp_feat, callback=_save_to(inp_feat))

    output_shape = (7, 7)
218
    out_feat = F.vision.roi_align(
219 220 221 222 223 224 225 226
        inp_feat,
        rois,
        output_shape=output_shape,
        mode="average",
        spatial_scale=1.0 / 4,
        sample_points=2,
        aligned=True,
    )
227 228 229 230 231
    assert make_shape_tuple(out_feat.shape) == (
        rois.shape[0],
        inp_feat.shape[1],
        *output_shape,
    )
232 233

    grad(out_feat, tensor(F.ones_like(out_feat)))
234
    assert make_shape_tuple(inp_feat.grad.shape) == make_shape_tuple(inp_feat.shape)
235 236


237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336
def _gen_correlation(random=True, constant=1, image_shape=(2, 1, 160, 160)):
    if random:
        inp_feat1 = np.random.randn(
            image_shape[0], image_shape[1], image_shape[2], image_shape[3]
        )
        inp_feat2 = np.random.randn(
            image_shape[0], image_shape[1], image_shape[2], image_shape[3]
        )
    else:
        inp_feat1 = np.ones(image_shape) * constant
        inp_feat2 = np.ones(image_shape) * constant

    return tensor(inp_feat1), tensor(inp_feat2)


def test_correlation():
    ##test case 0 check the grad shape
    data1, data2 = _gen_correlation()
    grad = Grad().wrt(data1, callback=_save_to(data1))

    out_feat = F.vision.correlation(
        data1,
        data2,
        kernel_size=5,
        max_displacement=4,
        stride1=2,
        stride2=2,
        pad_size=2,
        is_multiply=True,
    )

    grad(out_feat, tensor(F.ones_like(out_feat)))
    assert make_shape_tuple(data1.grad.shape) == make_shape_tuple(data1.shape)

    ##test case 1 from https://github.com/NVIDIA/flownet2-pytorch/issues/194
    data1, data2 = _gen_correlation(random=False, image_shape=(1, 1, 3, 3))

    out_feat = F.vision.correlation(
        data1,
        data2,
        kernel_size=3,
        max_displacement=0,
        stride1=1,
        stride2=1,
        pad_size=0,
        is_multiply=True,
    )
    assert abs(out_feat.sum() - 1) < 1e-9

    ##test case 2 check same image subduction
    data1, data2 = _gen_correlation(random=False, image_shape=(1, 1, 3, 3))

    out_feat = F.vision.correlation(
        data1,
        data2,
        kernel_size=3,
        max_displacement=0,
        stride1=1,
        stride2=1,
        pad_size=0,
        is_multiply=False,
    )
    assert out_feat.sum() < 1e-9

    ##test case 3 check same image subduction
    data1, data2 = _gen_correlation(random=False, image_shape=(1, 1, 3, 3))

    out_feat = F.vision.correlation(
        data1,
        data2,
        kernel_size=3,
        max_displacement=0,
        stride1=1,
        stride2=1,
        pad_size=0,
        is_multiply=False,
    )
    assert out_feat.sum() < 1e-9

    ##test case 4 check correlation
    data1, _ = _gen_correlation(
        random=False, image_shape=(1, 1, 220, 220), constant=2.0
    )
    _, data2 = _gen_correlation(
        random=False, image_shape=(1, 1, 220, 220), constant=1.0
    )

    out_feat = F.vision.correlation(
        data1,
        data2,
        kernel_size=3,
        max_displacement=2,
        stride1=1,
        stride2=2,
        pad_size=0,
        is_multiply=False,
    )
    assert abs(out_feat.mean() - 1) < 1e-9


337 338 339 340
def test_roi_pooling():
    inp_feat, rois = _gen_roi_inp()
    grad = Grad().wrt(inp_feat, callback=_save_to(inp_feat))
    output_shape = (7, 7)
341
    out_feat = F.vision.roi_pooling(
342 343
        inp_feat, rois, output_shape=output_shape, mode="max", scale=1.0 / 4,
    )
344 345 346 347 348
    assert make_shape_tuple(out_feat.shape) == (
        rois.shape[0],
        inp_feat.shape[1],
        *output_shape,
    )
349 350

    grad(out_feat, tensor(F.ones_like(out_feat)))
351
    assert make_shape_tuple(inp_feat.grad.shape) == make_shape_tuple(inp_feat.shape)
352 353


354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413
def test_adaptive_avg_pool2d():
    inp = tensor(np.arange(0, 16, dtype=np.float32).reshape(1, 1, 4, 4))
    oshp = (2, 2)
    grad = Grad().wrt(inp, callback=_save_to(inp))
    outp = F.adaptive_avg_pool2d(inp, oshp,)
    assert make_shape_tuple(outp.shape) == (inp.shape[0], inp.shape[1], *oshp,)
    np.testing.assert_equal(
        outp.numpy(), np.array([[[[2.5, 4.5], [10.5, 12.5]]]], dtype=np.float32)
    )

    grad(outp, tensor(F.ones_like(outp)))
    assert make_shape_tuple(inp.grad.shape) == make_shape_tuple(inp.shape)
    np.testing.assert_equal(
        inp.grad.numpy(),
        np.array(
            [
                [
                    [
                        [0.25, 0.25, 0.25, 0.25],
                        [0.25, 0.25, 0.25, 0.25],
                        [0.25, 0.25, 0.25, 0.25],
                        [0.25, 0.25, 0.25, 0.25],
                    ]
                ]
            ],
            dtype=np.float32,
        ),
    )


def test_adaptive_max_pool2d():
    inp = tensor(np.arange(0, 16, dtype=np.float32).reshape(1, 1, 4, 4))
    oshp = (2, 2)
    grad = Grad().wrt(inp, callback=_save_to(inp))
    outp = F.adaptive_max_pool2d(inp, oshp,)
    assert make_shape_tuple(outp.shape) == (inp.shape[0], inp.shape[1], *oshp,)
    np.testing.assert_equal(
        outp.numpy(), np.array([[[[5, 7], [13, 15]]]], dtype=np.float32)
    )

    grad(outp, tensor(F.ones_like(outp)))
    assert make_shape_tuple(inp.grad.shape) == make_shape_tuple(inp.shape)
    np.testing.assert_equal(
        inp.grad.numpy(),
        np.array(
            [
                [
                    [
                        [0.0, 0.0, 0.0, 0.0],
                        [0.0, 1.0, 0.0, 1.0],
                        [0.0, 0.0, 0.0, 0.0],
                        [0.0, 1.0, 0.0, 1.0],
                    ]
                ]
            ],
            dtype=np.float32,
        ),
    )


414 415 416 417
def test_one_hot():
    def onehot_low_dimension():
        inp = tensor(np.arange(1, 4, dtype=np.int32))
        out = F.one_hot(inp, num_classes=4)
418

419
        np.testing.assert_allclose(
420 421
            out.numpy(), np.eye(4, dtype=np.int32)[np.arange(1, 4, dtype=np.int32)]
        )
422

423 424 425 426 427
    def onehot_high_dimension():
        arr = np.array(
            [[3, 2, 4, 4, 2, 4, 0, 4, 4, 1], [4, 1, 1, 3, 2, 2, 4, 2, 4, 3]],
            dtype=np.int32,
        )
428

429 430
        inp = tensor(arr)
        out = F.one_hot(inp, 10)
431

432
        np.testing.assert_allclose(out.numpy(), np.eye(10, dtype=np.int32)[arr])
433

434 435
    onehot_low_dimension()
    onehot_high_dimension()
436 437


438
def test_interpolate_fastpath():
439 440 441 442 443
    # check shape
    test_cases = [
        [(1, 1, 10, 10), (5, 5)],
        [(1, 3, 10, 10), (20, 20)],
        [(10, 1, 10, 10), (1, 1)],
M
Megvii Engine Team 已提交
444
        # [(10, 10, 1, 1), (10, 10)], # FIXME, it causes random CI failure
445 446 447
    ]
    for inp_shape, target_shape in test_cases:
        x = tensor(np.random.randn(*inp_shape), dtype=np.float32)
448
        out = F.vision.interpolate(x, target_shape, mode="bilinear")
449 450 451 452 453
        assert out.shape[0] == x.shape[0] and out.shape[1] == x.shape[1]
        assert out.shape[2] == target_shape[0] and out.shape[3] == target_shape[1]

    # check value
    x = tensor(np.ones((3, 3, 10, 10)), dtype=np.float32)
454
    out = F.vision.interpolate(x, (15, 5), mode="bilinear")
455 456 457 458
    np.testing.assert_equal(out.numpy(), np.ones((3, 3, 15, 5)).astype(np.float32))

    np_x = np.arange(32)
    x = tensor(np_x).astype(np.float32).reshape(1, 1, 32, 1)
459
    out = F.vision.interpolate(x, (1, 1), mode="bilinear")
460 461 462
    np.testing.assert_equal(out.item(), np_x.mean())


463 464
@pytest.mark.parametrize("dt", [np.float32, np.int8, np.uint8, np.float16])
def test_warp_perspective(dt):
465
    inp_shape = (1, 1, 4, 4)
466
    x = tensor(np.arange(16, dtype=dt).reshape(inp_shape))
467 468 469 470 471 472 473
    M_shape = (1, 3, 3)
    # M defines a translation: dst(1, 1, h, w) = rst(1, 1, h+1, w+1)
    M = tensor(
        np.array(
            [[1.0, 0.0, 1.0], [0.0, 1.0, 1.0], [0.0, 0.0, 1.0]], dtype=np.float32
        ).reshape(M_shape)
    )
474
    outp = F.vision.warp_perspective(x, M, (2, 2))
475
    np.testing.assert_equal(outp.numpy(), np.array([[[[5, 6], [9, 10]]]], dtype=dt))
476 477


478 479
@pytest.mark.parametrize("dt", [np.float32, np.int8, np.uint8, np.float16])
def test_warp_perspective_mat_idx(dt):
480
    inp_shape = (2, 1, 4, 4)
481
    x = tensor(np.arange(32, dtype=dt).reshape(inp_shape))
482 483 484 485 486 487 488 489 490 491 492 493 494
    M_shape = (1, 3, 3)
    # M defines a translation: dst(1, 1, h, w) = rst(1, 1, h+1, w+1)
    M = tensor(
        np.array(
            [[1.0, 0.0, 1.0], [0.0, 1.0, 1.0], [0.0, 0.0, 1.0]], dtype=np.float32
        ).reshape(M_shape)
    )
    M = F.concat([M,] * 4, 0)
    outp = F.vision.warp_perspective(x, M, (2, 2), mat_idx=[0, 1, 1, 0])
    np.testing.assert_equal(
        outp.numpy(),
        np.array(
            [
495 496 497 498
                [[[5, 6], [9, 10]]],
                [[[21, 22], [25, 26]]],
                [[[21, 22], [25, 26]]],
                [[[5, 6], [9, 10]]],
499
            ],
500
            dtype=dt,
501 502 503 504
        ),
    )


505 506 507 508
def test_warp_affine():
    inp_shape = (1, 3, 3, 3)
    x = tensor(np.arange(27, dtype=np.float32).reshape(inp_shape))
    weightv = [[[1.26666667, 0.6, -83.33333333], [-0.33333333, 1, 66.66666667]]]
509
    outp = F.vision.warp_affine(x, tensor(weightv), (2, 2), border_mode="wrap")
510 511 512 513 514 515 516 517 518 519 520 521 522
    res = np.array(
        [
            [
                [[7.875, 8.875, 9.875], [8.90625, 9.90625, 10.90625]],
                [[18.75, 19.75, 20.75], [14.90625, 15.90625, 16.90625]],
            ]
        ],
        dtype=np.float32,
    )
    if not is_cuda_available():
        np.testing.assert_almost_equal(outp.numpy(), res, 5)


523 524 525 526 527 528 529 530 531
def test_remap():
    inp_shape = (1, 1, 4, 4)
    inp = tensor(np.arange(16, dtype=np.float32).reshape(inp_shape))
    map_xy_shape = (1, 2, 2, 2)
    map_xy = tensor(
        np.array(
            [[[1.0, 0.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 1.0]]], dtype=np.float32
        ).reshape(map_xy_shape)
    )
532
    outp = F.vision.remap(inp, map_xy)
533 534 535 536 537
    np.testing.assert_equal(
        outp.numpy(), np.array([[[[1.0, 4.0], [4.0, 4.0]]]], dtype=np.float32)
    )


538 539 540 541 542 543 544 545 546 547
def test_binary_cross_entropy():
    data1_shape = (2, 2)
    label1_shape = (2, 2)
    data2_shape = (2, 3)
    label2_shape = (2, 3)

    def sigmoid(x):
        return 1 / (1 + np.exp(-x))

    def compare_fn(x, y):
548
        np.testing.assert_allclose(x.numpy(), y, atol=5e-4)
549 550

    np.random.seed(123)
551
    data1 = np.random.uniform(size=data1_shape).astype(np.float32)
552 553 554 555
    label1 = np.random.uniform(size=label1_shape).astype(np.float32)
    expect1 = np.array([0.6361], dtype=np.float32)

    np.random.seed(123)
556
    data2 = np.random.uniform(size=data2_shape).astype(np.float32)
557 558 559 560 561 562 563
    label2 = np.random.uniform(size=label2_shape).astype(np.float32)
    expect2 = np.array([0.6750], dtype=np.float32)

    cases = [
        {"input": [data1, label1], "output": expect1,},
        {"input": [data2, label2], "output": expect2,},
    ]
564
    opr_test(cases, F.nn.binary_cross_entropy, compare_fn=compare_fn)
565

566 567 568 569 570
    cases = [
        {"input": [sigmoid(data1), label1], "output": expect1,},
        {"input": [sigmoid(data2), label2], "output": expect2,},
    ]
    opr_test(
571 572 573
        cases,
        partial(F.nn.binary_cross_entropy, with_logits=False),
        compare_fn=compare_fn,
574 575
    )

576 577 578 579 580 581 582 583 584 585 586

def test_hinge_loss():
    np.random.seed(123)
    # case with L1 norm
    cases = []
    for shape in [(2, 2), (2, 3)]:
        data = np.random.uniform(size=shape).astype(np.float32)
        label = 2 * np.random.randint(0, 1, size=shape).astype(np.float32) - 1
        expect = np.clip(0, np.inf, 1 - data * label).sum(axis=1).mean()
        cases.append({"input": [data, label], "output": expect})

587
    opr_test(cases, F.nn.hinge_loss)
588 589 590 591 592 593 594 595 596 597

    # cases with L2 norm
    cases = []
    for shape in [(2, 2), (2, 3)]:
        data = np.random.uniform(size=shape).astype(np.float32)
        label = 2 * np.random.randint(0, 1, size=shape).astype(np.float32) - 1
        expect = ((np.clip(0, np.inf, 1 - data * label) ** 2).sum(axis=1)).mean()
        cases.append({"input": [data, label], "output": expect})

    def hinge_loss_with_l2_norm(pred, label):
598
        return F.nn.hinge_loss(pred, label, "L2")
599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614

    opr_test(cases, hinge_loss_with_l2_norm)


def test_nms():
    x = np.array(
        [
            [0, 0, 100, 100],
            [10, 10, 100, 100],
            [50, 50, 100, 100],
            [100, 100, 150, 150],
        ],
        dtype=np.float32,
    )
    inp = tensor(x)
    scores = tensor([0.5, 0.8, 0.9, 0.6], dtype=np.float32)
615
    result = F.vision.nms(inp, scores=scores, iou_thresh=0.5)
616 617 618
    np.testing.assert_equal(result.numpy(), np.array([2, 1, 3], dtype=np.int32))


619
@pytest.mark.skipif(
620
    get_device_count("gpu") > 0, reason="cuda does not support nchw int8"
621
)
622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643
def test_conv_bias():
    inp_scale = 1.5
    w_scale = 2.5
    outp_scale = 1.5
    inp_dtype = dtype.qint8(inp_scale)
    w_dtype = dtype.qint8(w_scale)
    b_dtype = dtype.qint32(inp_scale * w_scale)
    out_dtype = dtype.qint8(outp_scale)

    def run(
        N,
        IC,
        OC,
        IH,
        IW,
        KH,
        KW,
        PH,
        PW,
        SH,
        SW,
        has_bias=True,
644
        nonlinear_mode="identity",
645 646
    ):
        inp_v = np.random.normal(size=(N, IC, IH, IW))
647
        w_v = np.random.normal(size=(OC, IC, KH, KW))
648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668
        b_v = np.random.normal(size=(1, OC, 1, 1))
        inp_scale = dtype.get_scale(inp_dtype)
        w_scale = dtype.get_scale(w_dtype)
        b_scale = dtype.get_scale(b_dtype)

        inpv = dtype.convert_to_qint8(inp_v * inp_scale, inp_dtype)
        wv = dtype.convert_to_qint8(w_v * w_scale, w_dtype)
        bv = dtype.convert_to_qint32(b_v * b_scale, b_dtype)

        inp_int8 = tensor(inpv, dtype=inp_dtype)
        w_int8 = Parameter(wv, dtype=w_dtype)
        b_int32 = Parameter(bv, dtype=b_dtype)

        inp_fp32 = inp_int8.astype("float32")
        w_fp32 = w_int8.astype("float32")
        b_fp32 = b_int32.astype("float32")

        def convert_to_nchw4(var):
            var = F.reshape(
                var, (var.shape[0], var.shape[1] // 4, 4, var.shape[2], var.shape[3])
            )
669
            var = F.transpose(var, (0, 1, 3, 4, 2))
670 671 672 673 674 675
            return var

        def run_conv2d(inp, w, b):
            O = F.conv2d(
                inp, w, b if has_bias else None, stride=(SH, SW), padding=(PH, PW),
            )
676
            if nonlinear_mode == "relu":
677 678 679 680 681 682 683 684 685 686
                return F.relu(O)
            else:
                return O

        def run_conv_bias(inp, w, b, format="NCHW"):
            b = b if has_bias else Parameter(np.zeros_like(b.numpy()))
            if format == "NCHW4":
                inp = convert_to_nchw4(inp)
                w = convert_to_nchw4(w)
                b = convert_to_nchw4(b)
687
            return F.quantized.conv_bias_activation(
688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704
                inp,
                w,
                b,
                stride=(SH, SW),
                padding=(PH, PW),
                dtype=out_dtype,
                nonlinear_mode=nonlinear_mode,
            )

        format = "NCHW4" if is_cuda_available() else "NCHW"

        expected = run_conv2d(inp_fp32, w_fp32, b_fp32)
        expected = expected.astype(out_dtype).astype("float32")
        result = run_conv_bias(inp_int8, w_int8, b_int32, format=format).astype(
            "float32"
        )
        if format == "NCHW4":
705
            result = F.transpose(result, (0, 1, 4, 2, 3))
706 707
        expected = F.flatten(expected)
        result = F.flatten(result)
708
        np.testing.assert_allclose(result.numpy(), expected.numpy(), atol=outp_scale)
709 710 711 712 713 714 715 716 717

    run(1, 4, 4, 24, 33, 1, 1, 2, 3, 1, 1, False)
    run(10, 12, 24, 46, 46, 1, 1, 2, 1, 3, 1, False)
    run(10, 36, 8, 46, 26, 2, 2, 2, 1, 1, 2, False)

    run(1, 4, 4, 24, 33, 1, 1, 2, 3, 1, 1)
    run(10, 12, 24, 46, 46, 1, 1, 2, 1, 3, 1)
    run(10, 36, 8, 46, 26, 2, 2, 2, 1, 1, 2)

718 719
    run(10, 36, 8, 46, 26, 2, 2, 2, 1, 1, 2, False, "relu")
    run(10, 36, 8, 46, 26, 2, 2, 2, 1, 1, 2, True, "relu")
720 721


722
@pytest.mark.skipif(get_device_count("gpu") > 0, reason="no int8 algorithm on cuda")
723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772
def test_batch_conv_bias():
    inp_scale = 1.5
    w_scale = 2.5
    outp_scale = 1.5
    inp_dtype = dtype.qint8(inp_scale)
    w_dtype = dtype.qint8(w_scale)
    b_dtype = dtype.qint32(inp_scale * w_scale)
    out_dtype = dtype.qint8(outp_scale)

    def run(
        N, IC, OC, IH, IW, KH, KW, PH, PW, SH, SW, has_bias=True,
    ):
        inp_v = np.random.normal(size=(N, IC, IH, IW))
        w_v = np.random.normal(size=(N, OC, IC, KH, KW))
        b_v = np.random.normal(size=(1, OC, 1, 1))
        inp_scale = dtype.get_scale(inp_dtype)
        w_scale = dtype.get_scale(w_dtype)
        b_scale = dtype.get_scale(b_dtype)

        inpv = dtype.convert_to_qint8(inp_v * inp_scale, inp_dtype)
        wv = dtype.convert_to_qint8(w_v * w_scale, w_dtype)
        bv = dtype.convert_to_qint32(b_v * b_scale, b_dtype)

        inp_int8 = tensor(inpv, dtype=inp_dtype)
        w_int8 = Parameter(wv, dtype=w_dtype)
        b_int32 = Parameter(bv, dtype=b_dtype)

        inp_fp32 = inp_int8.astype("float32")
        w_fp32 = w_int8.astype("float32")
        b_fp32 = b_int32.astype("float32")

        def run_batch_conv_bias(inp, w, b):
            b = b if has_bias else Parameter(np.zeros_like(b.numpy()))
            result = F.quantized.batch_conv_bias_activation(
                inp, w, b, stride=(SH, SW), padding=(PH, PW), dtype=out_dtype,
            )
            return result.astype("float32")

        expected = F.conv2d(inp_fp32, w_fp32[0], b_fp32 if has_bias else None)[0]
        expected = expected.astype(out_dtype).astype("float32")
        expected = F.flatten(expected)

        result = run_batch_conv_bias(inp_int8, w_int8, b_int32)
        result = F.flatten(result)

        np.testing.assert_allclose(result.numpy(), expected.numpy(), atol=outp_scale)

    run(1, 4, 4, 5, 5, 3, 3, 0, 0, 1, 1, True)


773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793
def test_conv2d_io16c32():
    amp.enabled = True
    inp = tensor(np.random.randn(1, 3, 224, 224), dtype=np.float32)
    weight = tensor(np.random.randn(64, 3, 7, 7), dtype=np.float32)
    out = F.conv2d(inp, weight, None, (2, 2), (3, 3), (1, 1), 1)
    amp.enabled = False
    expected = F.conv2d(
        inp.astype("float16"),
        weight.astype("float16"),
        None,
        (2, 2),
        (3, 3),
        (1, 1),
        1,
        compute_mode="float32",
    )
    assert out.dtype == np.float16
    assert expected.dtype == np.float16
    np.testing.assert_allclose(out.numpy(), expected.numpy())


794
def test_conv2d_zero_stride_numpy_array():
795 796 797 798 799 800 801 802
    inp = np.random.randn(3, 224, 224).astype(np.float32)
    inp = inp[np.newaxis, :]

    inp = tensor(inp, dtype=np.float32)
    weight = tensor(np.random.randn(16, 3, 3, 3), dtype=np.float32)
    out = F.conv2d(inp, weight, None, (2, 2), (3, 3), (1, 1), 1)


803 804 805 806 807 808 809 810 811 812
def test_conv3d_zero_stride_numpy_array():
    inp = np.random.randn(3, 224, 224, 224).astype(np.float32)
    inp = inp[np.newaxis, :]

    inp = tensor(inp, dtype=np.float32)
    weight = tensor(np.random.randn(16, 3, 3, 3, 3), dtype=np.float32)
    out = F.conv3d(inp, weight, None, (2, 2, 2), (3, 3, 3), (1, 1, 1), 1)
    out.numpy()


813
def test_conv1d():
814 815
    inp = tensor(np.ones((2, 2, 4), dtype=np.float32))
    weight = tensor(np.ones((3, 2, 2), dtype=np.float32))
816 817 818 819 820 821 822 823 824
    out = F.conv1d(inp, weight, None, 2, 0, 1, 1)
    np.testing.assert_equal(
        out.numpy(),
        np.array(
            [[[4, 4], [4, 4], [4, 4]], [[4, 4], [4, 4], [4, 4]]], dtype=np.float32
        ),
    )


825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846
def test_batchnorm2d_io16c32():
    amp.enabled = True
    inp = tensor(np.random.randn(1, 3, 224, 224), dtype=np.float32)
    weight = tensor(np.ones((1, 3, 1, 1)), dtype=np.float32)
    bias = tensor(np.zeros((1, 3, 1, 1)), dtype=np.float32)

    out = F.batch_norm(inp, weight=weight, bias=bias, training=True, inplace=False)

    amp.enabled = False
    expected = F.batch_norm(
        inp.astype("float16"),
        weight=weight,
        bias=bias,
        training=True,
        inplace=False,
        compute_mode="float32",
    )
    assert out.dtype == np.float16
    assert expected.dtype == np.float16
    np.testing.assert_allclose(out.numpy(), expected.numpy())


847
def test_conv3d():
848 849
    inp = tensor(np.ones((2, 2, 4, 4, 4), dtype=np.float32))
    weight = tensor(np.ones((3, 2, 2, 2, 2), dtype=np.float32))
850 851 852 853 854 855 856
    out = F.conv3d(inp, weight, None, 2, 0, 1, 1)
    print(out.numpy().shape)
    np.testing.assert_equal(
        out.numpy(), np.ones((2, 3, 2, 2, 2), dtype=np.float32) * 16
    )


857 858 859 860 861 862 863 864
def test_condtake():
    x = np.array([[1, 2, 3], [4, 5, 6]])
    y = np.array([[True, False, True], [False, True, True]])
    xx = tensor(x)
    yy = tensor(y)
    val, idx = F.cond_take(yy, xx)
    np.testing.assert_equal(val.numpy(), x[y])
    np.testing.assert_equal(idx.numpy(), np.where(y.reshape(-1))[0])
865 866


867 868
@pytest.mark.parametrize("is_symbolic", [None, False, True])
def test_condtake(is_symbolic):
869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895
    shapes = [
        (3, 3, 3),
        (0,),
        (3, 0, 3),
    ]

    def fn(mask, data):
        return F.cond_take(mask, data)

    if is_symbolic is not None:
        fn = jit.trace(symbolic=is_symbolic)(fn)

    for shp in shapes:
        x_np = np.random.randn(*shp).astype("float32")
        mask_np = x_np > 0
        x = tensor(x_np)
        mask = tensor(mask_np)
        ref_out = x_np[mask_np]
        ref_idx = mask_np.flatten().nonzero()[0]
        for i in range(3):
            out, idx = fn(mask, x)
            np.testing.assert_equal(out.numpy(), ref_out)
            np.testing.assert_equal(idx.numpy(), ref_idx)
            if is_symbolic is None:
                break


896 897 898 899 900 901 902 903 904 905 906 907 908 909 910
def test_condtake_is_same():
    op1 = builtin.CondTake()
    op2 = builtin.CondTake()
    assert op1 == op2


def test_nms_is_same():
    op1 = builtin.NMSKeep(0.7, 100)
    op2 = builtin.NMSKeep(0.7, 100)
    op3 = builtin.NMSKeep(0.8, 100)
    op4 = builtin.NMSKeep(0.7, 200)
    assert op1 == op2
    assert op1 != op3
    assert op1 != op4
    assert op3 != op4
911 912


913 914 915 916 917 918 919 920 921 922 923 924 925 926 927
def test_argmxx_on_inf():
    def run_argmax():
        x = F.zeros((100, 100))
        x[:] = -float("inf")
        idxs = F.argmax(x, axis=0)
        return idxs

    def run_argmin():
        x = F.zeros((100, 100))
        x[:] = float("inf")
        idxs = F.argmin(x, axis=0)
        return idxs

    assert all(run_argmax() >= 0)
    assert all(run_argmin() >= 0)
928 929


930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956
def test_deformable_psroi_pooling():
    inp = np.random.random((1, 256, 64, 64)).astype("float32")
    rois = np.random.random((1, 5)).astype("float32")
    trans = np.random.random((24, 2, 7, 7)).astype("float32")

    pooled_h = 7
    pooled_w = 7
    sample_per_part = 4
    no_trans = False
    part_size = 7
    spatial_scale = 1.0 / 64
    trans_std = 0.1

    y = F.deformable_psroi_pooling(
        tensor(inp),
        tensor(rois),
        tensor(trans),
        no_trans,
        part_size,
        pooled_h,
        pooled_w,
        sample_per_part,
        spatial_scale,
        trans_std,
    )


957 958 959 960 961 962 963
def test_cvt_color():
    def rgb2gray(rgb):
        return np.dot(rgb[..., :3], [0.299, 0.587, 0.114])

    inp = np.random.randn(3, 3, 3, 3).astype(np.float32)
    out = np.expand_dims(rgb2gray(inp), 3).astype(np.float32)
    x = tensor(inp)
964
    y = F.vision.cvt_color(x, mode="RGB2GRAY")
965
    np.testing.assert_allclose(y.numpy(), out, atol=1e-5)
966 967 968 969 970 971 972


@pytest.mark.parametrize("val", [2, [2,], [2, 3]])
def test_ones(val):
    shp = tensor(val)
    np_shp = np.array(val)
    np.testing.assert_equal(F.ones(shp), np.ones(np_shp))
973 974 975 976 977 978 979 980 981 982 983 984 985 986 987


def test_assert_equal():
    shape = (2, 3, 4, 5)
    x = F.ones(shape, dtype=np.float32)
    y = F.zeros(shape, dtype=np.float32) + 1.00001
    z = F.utils._assert_equal(x, y)


def test_assert_not_equal():
    shape = (2, 3, 4, 5)
    x = F.ones(shape, dtype=np.float32)
    y = F.zeros(shape, dtype=np.float32) + 1.1
    with pytest.raises(RuntimeError):
        z = F.utils._assert_equal(x, y)
988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003


def test_neg_axis():
    x = tensor(np.random.normal(0, 1, (32, 5)))

    y = F.argmax(x, axis=-1)
    yy = F.argmax(x, axis=1)
    np.testing.assert_equal(y.numpy(), yy.numpy())

    y = F.argmax(x, axis=(-1, -2))
    yy = F.argmax(x, axis=(0, 1))
    np.testing.assert_equal(y.numpy(), yy.numpy())

    y = F.argmin(x, axis=(-1, -2))
    yy = F.argmin(x, axis=(0, 1))
    np.testing.assert_equal(y.numpy(), yy.numpy())
1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028


def test_sliding_window():
    N, C, H, W = 2, 3, 7, 8
    inp = np.random.normal(size=(N, C, H, W))
    ph, pw = 1, 2
    sh, sw = 2, 1
    wh, ww = 3, 2
    dh, dw = 1, 3
    s = lambda i, p, s, d, w: (i + p * 2 - (w - 1) * d - 1) // s + 1
    inp_pad = np.zeros((N, C, H + ph * 2, W + pw * 2))
    inp_pad[:, :, ph : H + ph, pw : W + pw] = inp
    gt_out = np.empty(
        (N, C, s(H, ph, sh, dh, wh), s(W, pw, sw, dw, ww), wh, ww), dtype=np.float32
    )
    for n, c, oh, ow in itertools.product(*map(range, gt_out.shape[:4])):
        ih, iw = oh * sh, ow * sw
        gt_out[n, c, oh, ow, :] = inp_pad[
            n, c, ih : ih + (wh - 1) * dh + 1 : dh, iw : iw + (ww - 1) * dw + 1 : dw
        ]

    out = F.sliding_window(
        tensor(inp), (wh, ww), padding=(ph, pw), stride=(sh, sw), dilation=(dh, dw)
    )
    np.testing.assert_equal(gt_out, out.numpy())
1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064


def test_sliding_window_transpose():
    N, C, H, W = 2, 3, 7, 8
    ph, pw = 1, 2
    sh, sw = 2, 1
    wh, ww = 3, 2
    dh, dw = 1, 3
    s = lambda i, p, s, d, w: (i + p * 2 - (w - 1) * d - 1) // s + 1
    inp = np.random.normal(
        size=(N, C, s(H, ph, sh, dh, wh), s(W, pw, sw, dw, ww), wh, ww)
    ).astype(np.float32)
    gt_out = np.zeros((N, C, H, W), dtype=np.float32)

    for n, c in itertools.product(*map(range, inp.shape[:2])):
        oh = 0
        for ih in range(-ph, H + ph - dh * (wh - 1), sh):
            ow = 0
            for iw in range(-pw, W + pw - dw * (ww - 1), sw):
                for kh, kw in itertools.product(*map(range, inp.shape[-2:])):
                    ih2 = ih + dh * kh
                    iw2 = iw + dw * kw
                    if ih2 >= 0 and ih2 < H and iw2 >= 0 and iw2 < W:
                        gt_out[n, c, ih2, iw2] += inp[n, c, oh, ow, kh, kw]
                ow += 1
            oh += 1

    out = F.sliding_window_transpose(
        tensor(inp),
        (H, W),
        (wh, ww),
        padding=(ph, pw),
        stride=(sh, sw),
        dilation=(dh, dw),
    )
    np.testing.assert_equal(gt_out, out.numpy())
1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083


def test_pad():
    src = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=np.float32)
    dst = np.pad(src, ((2, 2), (2, 2)), "constant")
    res = F.nn.pad(tensor(src), ((2, 2), (2, 2)), "CONSTANT")
    np.testing.assert_allclose(res, dst, atol=1e-5)

    dst = np.pad(src, ((2, 2), (2, 2)), "constant", constant_values=3)
    res = F.nn.pad(tensor(src), ((2, 2), (2, 2)), "CONSTANT", constant_value=3)
    np.testing.assert_allclose(res, dst, atol=1e-5)

    dst = np.pad(src, ((2, 2), (2, 2)), "edge")
    res = F.nn.pad(tensor(src), ((2, 2), (2, 2)), "EDGE")
    np.testing.assert_allclose(res, dst, atol=1e-5)

    dst = np.pad(src, ((2, 2), (2, 2)), "reflect")
    res = F.nn.pad(tensor(src), ((2, 2), (2, 2)), "REFLECT")
    np.testing.assert_allclose(res, dst, atol=1e-5)