test_tensor.py 12.6 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
#
# 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.
9
import os
10 11
import platform

12 13
import numpy as np
import pytest
14
from utils import opr_test
15 16

import megengine.functional as F
M
Megvii Engine Team 已提交
17
from megengine import tensor
18
from megengine.core._trace_option import use_symbolic_shape
19
from megengine.core.tensor.utils import astensor1d
20
from megengine.distributed.helper import get_device_count_by_fork
21 22 23 24


def test_eye():
    dtype = np.float32
25
    cases = [{"input": [10, 20]}, {"input": [30]}]
26
    for case in cases:
27
        np.testing.assert_allclose(
28 29 30
            F.eye(case["input"], dtype=dtype).numpy(),
            np.eye(*case["input"]).astype(dtype),
        )
31 32 33 34 35 36 37 38
        np.testing.assert_allclose(
            F.eye(*case["input"], dtype=dtype).numpy(),
            np.eye(*case["input"]).astype(dtype),
        )
        np.testing.assert_allclose(
            F.eye(tensor(case["input"]), dtype=dtype).numpy(),
            np.eye(*case["input"]).astype(dtype),
        )
39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55


def test_concat():
    def get_data_shape(length: int):
        return (length, 2, 3)

    data1 = np.random.random(get_data_shape(5)).astype("float32")
    data2 = np.random.random(get_data_shape(6)).astype("float32")
    data3 = np.random.random(get_data_shape(7)).astype("float32")

    def run(data1, data2):
        return F.concat([data1, data2])

    cases = [{"input": [data1, data2]}, {"input": [data1, data3]}]
    opr_test(cases, run, ref_fn=lambda x, y: np.concatenate([x, y]))


56 57 58 59 60 61 62 63
def test_concat_device():
    data1 = tensor(np.random.random((3, 2, 2)).astype("float32"), device="cpu0")
    data2 = tensor(np.random.random((2, 2, 2)).astype("float32"), device="cpu1")

    out = F.concat([data1, data2], device="cpu0")
    assert str(out.device).split(":")[0] == "cpu0"


64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79
def test_stack():
    data1 = np.random.random((3, 2, 2)).astype("float32")
    data2 = np.random.random((3, 2, 2)).astype("float32")
    data3 = np.random.random((3, 2, 2)).astype("float32")

    cases = [{"input": [data1, data2]}, {"input": [data1, data3]}]
    for ai in range(3):

        def run(data1, data2):
            return F.stack([data1, data2], axis=ai)

        opr_test(cases, run, ref_fn=lambda x, y: np.stack([x, y], axis=ai))


def test_split():
    data = np.random.random((2, 3, 4, 5)).astype(np.float32)
80 81 82 83
    inp = tensor(data)

    mge_out0 = F.split(inp, 2, axis=3)
    mge_out1 = F.split(inp, [3], axis=3)
84 85 86

    np_out = np.split(data, [3, 5], axis=3)

87 88 89 90
    assert len(mge_out0) == 2
    assert len(mge_out1) == 2

    np.testing.assert_equal(mge_out0[0].numpy(), np_out[0])
91 92
    np.testing.assert_equal(mge_out1[0].numpy(), np_out[0])

93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
    np.testing.assert_equal(mge_out0[1].numpy(), np_out[1])
    np.testing.assert_equal(mge_out1[1].numpy(), np_out[1])

    try:
        F.split(inp, 4)
        assert False
    except ValueError as e:
        pass

    try:
        F.split(inp, [3, 3, 5], axis=3)
        assert False
    except ValueError as e:
        assert str(e) == "Invalid nsplits_or_secions: [3, 3, 5]"

108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124

def test_reshape():
    x = np.arange(6, dtype="float32")
    xx = tensor(x)
    y = x.reshape(1, 2, 3)

    for shape in [
        (1, 2, 3),
        (1, -1, 3),
        (1, tensor(-1), 3),
        np.array([1, -1, 3], dtype="int32"),
        tensor([1, -1, 3]),
    ]:
        yy = F.reshape(xx, shape)
        np.testing.assert_equal(yy.numpy(), y)


125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
def test_reshape_shape_inference():
    x_shape_known = tensor([1, 2, 3, 4], dtype="float32")
    x_shape_unknown = F.broadcast_to(tensor([1.0]), shape=tensor([1, 1, 1, 1]).sum())
    tshp_unknown = astensor1d((tensor([2]), tensor([2])), x_shape_known)
    tshp_known = astensor1d((2, 2), x_shape_known)
    tshp_known_unspec = astensor1d((2, -1), x_shape_known)

    def check_shape(output, target):
        source = output.shape
        if isinstance(source, tensor):
            source = source.numpy()
        np.testing.assert_equal(source, target)

    def func(x, target_shape):
        return x.reshape(target_shape)

    cases = [
        {"input": [x_shape_known, tshp_unknown], "output": [(2, 2),]},
        {"input": [x_shape_unknown, tshp_unknown], "output": [(2, 2),]},
        {"input": [x_shape_known, tshp_known], "output": [(2, 2),]},
        {"input": [x_shape_known, tshp_known_unspec], "output": [(2, 2),]},
        {"input": [x_shape_unknown, tshp_known], "output": [(2, 2),]},
        {"input": [x_shape_unknown, tshp_known_unspec], "output": [(2, 2),]},
    ]
    opr_test(cases, func, compare_fn=check_shape, test_trace=True)


152 153 154 155 156 157
def test_squeeze():
    x = np.arange(6, dtype="float32").reshape(1, 2, 3, 1)
    xx = tensor(x)

    for axis in [None, 3, -4, (3, -4)]:
        y = np.squeeze(x, axis)
158
        yy = F.squeeze(xx, axis)
159 160 161 162 163 164 165 166 167
        np.testing.assert_equal(y, yy.numpy())


def test_expand_dims():
    x = np.arange(6, dtype="float32").reshape(2, 3)
    xx = tensor(x)

    for axis in [2, -3, (3, -4), (1, -4)]:
        y = np.expand_dims(x, axis)
168
        yy = F.expand_dims(xx, axis)
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
        np.testing.assert_equal(y, yy.numpy())


def test_elemwise_dtype_promotion():
    x = np.random.rand(2, 3).astype("float32")
    y = np.random.rand(1, 3).astype("float16")
    xx = tensor(x)
    yy = tensor(y)
    z = xx * yy
    np.testing.assert_equal(z.numpy(), x * y)

    z = xx + y
    np.testing.assert_equal(z.numpy(), x + y)

    z = x - yy
    np.testing.assert_equal(z.numpy(), x - y)


def test_linspace():
    cases = [
        {"input": [1, 9, 9]},
        {"input": [3, 10, 8]},
    ]
    opr_test(
        cases,
        F.linspace,
        ref_fn=lambda start, end, step: np.linspace(start, end, step, dtype=np.float32),
    )

    cases = [
        {"input": [9, 1, 9]},
        {"input": [10, 3, 8]},
    ]
    opr_test(
        cases,
        F.linspace,
        ref_fn=lambda start, end, step: np.linspace(start, end, step, dtype=np.float32),
206 207 208 209 210 211 212 213 214 215
    )

    cases = [
        {"input": [1, tensor(9), 9]},
        {"input": [tensor(1), 9, tensor(9)]},
    ]
    opr_test(
        cases,
        F.linspace,
        ref_fn=lambda start, end, step: np.linspace(1, 9, 9, dtype=np.float32),
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
    )


def test_arange():
    cases = [
        {"input": [1, 9, 1]},
        {"input": [2, 10, 2]},
    ]
    opr_test(
        cases,
        F.arange,
        ref_fn=lambda start, end, step: np.arange(start, end, step, dtype=np.float32),
    )

    cases = [
        {"input": [9, 1, -1]},
        {"input": [10, 2, -2]},
    ]
    opr_test(
        cases,
        F.arange,
        ref_fn=lambda start, end, step: np.arange(start, end, step, dtype=np.float32),
    )

    cases = [
        {"input": [9.3, 1.2, -0.5]},
        {"input": [10.3, 2.1, -1.7]},
    ]
    opr_test(
        cases,
        F.arange,
        ref_fn=lambda start, end, step: np.arange(start, end, step, dtype=np.float32),
    )


def test_round():
    data1_shape = (15,)
    data2_shape = (25,)
    data1 = np.random.random(data1_shape).astype(np.float32)
    data2 = np.random.random(data2_shape).astype(np.float32)

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


261 262 263 264 265 266 267
def test_flatten():
    data0_shape = (2, 3, 4, 5)
    data1_shape = (4, 5, 6, 7)
    data0 = np.random.random(data0_shape).astype(np.float32)
    data1 = np.random.random(data1_shape).astype(np.float32)

    def compare_fn(x, y):
268
        assert x.shape[0] == y
269 270 271 272

    output0 = (2 * 3 * 4 * 5,)
    output1 = (4 * 5 * 6 * 7,)
    cases = [
273 274
        {"input": data0, "output": output0},
        {"input": data1, "output": output1},
275 276 277 278 279 280
    ]
    opr_test(cases, F.flatten, compare_fn=compare_fn)

    output0 = (2, 3 * 4 * 5)
    output1 = (4, 5 * 6 * 7)
    cases = [
281 282
        {"input": data0, "output": output0},
        {"input": data1, "output": output1},
283 284 285 286 287 288
    ]
    opr_test(cases, F.flatten, compare_fn=compare_fn, start_axis=1)

    output0 = (2, 3, 4 * 5)
    output1 = (4, 5, 6 * 7)
    cases = [
289 290
        {"input": data0, "output": output0},
        {"input": data1, "output": output1},
291 292 293 294 295 296
    ]
    opr_test(cases, F.flatten, compare_fn=compare_fn, start_axis=2)

    output0 = (2, 3 * 4, 5)
    output1 = (4, 5 * 6, 7)
    cases = [
297 298
        {"input": data0, "output": output0},
        {"input": data1, "output": output1},
299 300 301 302
    ]
    opr_test(cases, F.flatten, compare_fn=compare_fn, start_axis=1, end_axis=2)


303 304 305 306 307
def test_broadcast():
    input1_shape = (20, 30)
    output1_shape = (30, 20, 30)
    data1 = np.random.random(input1_shape).astype(np.float32)

308
    input2_shape = (10, 1)
309 310 311
    output2_shape = (20, 10, 20)
    data2 = np.random.random(input2_shape).astype(np.float32)

312 313 314 315
    input3_shape = (10, 10)
    output3_shape = (10, 10)
    data3 = np.random.random(input3_shape).astype(np.float32)

316
    def compare_fn(x, y):
317
        assert x.shape[0] == y
318 319 320 321

    cases = [
        {"input": [data1, output1_shape], "output": output1_shape},
        {"input": [data2, output2_shape], "output": output2_shape},
322
        {"input": [data3, output3_shape], "output": output3_shape},
323
    ]
324
    opr_test(cases, F.broadcast_to, compare_fn=compare_fn)
325

326
    x = F.ones((2, 1, 3))
327
    with pytest.raises(RuntimeError):
328
        F.broadcast_to(x, (2, 3, 4))
329

330
    with pytest.raises(RuntimeError):
331
        F.broadcast_to(x, (4, 1, 3))
332

333
    with pytest.raises(RuntimeError):
334
        F.broadcast_to(x, (1, 3))
335

336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376

def test_utils_astensor1d():
    reference = tensor(0)

    # literal
    x = [1, 2, 3]
    for dtype in [None, "float32"]:
        xx = astensor1d(x, reference, dtype=dtype)
        assert type(xx) is tensor
        np.testing.assert_equal(xx.numpy(), x)

    # numpy array
    x = np.asarray([1, 2, 3], dtype="int32")
    for dtype in [None, "float32"]:
        xx = astensor1d(x, reference, dtype=dtype)
        assert type(xx) is tensor
        np.testing.assert_equal(xx.numpy(), x.astype(dtype) if dtype else x)

    # tensor
    x = tensor([1, 2, 3], dtype="int32")
    for dtype in [None, "float32"]:
        xx = astensor1d(x, reference, dtype=dtype)
        assert type(xx) is tensor
        np.testing.assert_equal(xx.numpy(), x.numpy())

    # mixed
    x = [1, tensor(2), 3]
    for dtype in [None, "float32"]:
        xx = astensor1d(x, reference, dtype=dtype)
        assert type(xx) is tensor
        np.testing.assert_equal(xx.numpy(), [1, 2, 3])


def test_device():
    x = tensor([1, 2, 3], dtype="float32")

    y1 = F.eye(x.shape, dtype="float32")
    y2 = F.eye(x.shape, dtype="float32", device=None)
    np.testing.assert_almost_equal(y1.numpy(), y2.numpy())

    y3 = F.eye(x.shape, dtype="float32", device="xpux")
377
    y4 = F.eye(x.shape, dtype="float32", device=x.device)
378 379 380 381 382
    np.testing.assert_almost_equal(y3.numpy(), y4.numpy())

    y5 = F.full((3, 2), 4, device=x.device)
    y6 = F.full((3, 2), 4, device="xpux")
    np.testing.assert_almost_equal(y5.numpy(), y6.numpy())
383 384


385 386
def test_identity():
    x = tensor(np.random.random((5, 10)).astype(np.float32))
M
Megvii Engine Team 已提交
387
    y = F.copy(x)
388 389 390
    np.testing.assert_equal(y.numpy(), x)


391 392 393 394 395
def copy_test(dst, src):
    data = np.random.random((2, 3)).astype(np.float32)
    x = tensor(data, device=src)
    y = F.copy(x, dst)
    assert np.allclose(data, y.numpy())
396 397
    z = x.to(dst)
    assert np.allclose(data, z.numpy())
398 399


400
@pytest.mark.require_ngpu(1)
401 402 403 404
def test_copy_h2d():
    copy_test("cpu0", "gpu0")


405
@pytest.mark.require_ngpu(1)
406 407 408 409
def test_copy_d2h():
    copy_test("gpu0", "cpu0")


410
@pytest.mark.require_ngpu(2)
411 412 413
def test_copy_d2d():
    copy_test("gpu0", "gpu1")
    copy_test("gpu0:0", "gpu0:1")
414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463


@pytest.mark.parametrize(
    "shape, repeats, axis",
    [
        ((2,), 2, 0),
        ((2, 3, 4, 5), 3, 0),
        ((2, 3, 4, 5), 4, 3),
        ((2,), 2, None),
        ((2, 3, 4, 5), 3, None),
        ((), 1, None),
        ((), 10, None),
    ],
)
def test_repeat(shape, repeats, axis):
    def repeat_func(inp):
        return F.repeat(inp=inp, repeats=repeats, axis=axis)

    if shape != ():
        cases = [
            {"input": np.random.randn(*shape).astype("float32")},
        ]
    else:
        cases = [{"input": np.array(1.23)}]

    opr_test(
        cases, repeat_func, ref_fn=lambda inp: np.repeat(inp, repeats, axis),
    )


@pytest.mark.parametrize(
    "shape, reps",
    [
        ((2,), (2,)),
        ((2, 3, 4, 5), (1, 1, 1, 1)),
        ((2, 3, 4, 5), (1, 2, 3, 4)),
        ((2, 3, 4, 5), (2, 2, 2, 2, 2, 2, 2)),
    ],
)
def test_tile(shape, reps):
    def tile_func(inp):
        return F.tile(inp=inp, reps=reps)

    cases = [
        {"input": np.random.randn(*shape).astype("float32")},
    ]

    opr_test(
        cases, tile_func, ref_fn=lambda inp: np.tile(inp, reps),
    )