test_conv2d_op.py 44.1 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15 16
from __future__ import print_function

17 18
import unittest
import numpy as np
D
dzhwinter 已提交
19

20
import paddle.fluid.core as core
L
liym27 已提交
21
import paddle.fluid as fluid
22 23
from op_test import OpTest
from paddle.fluid import Program, program_guard
24 25


L
liym27 已提交
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
def conv2d_forward_naive(input,
                         filter,
                         group,
                         conv_param,
                         padding_algorithm='EXPLICIT',
                         data_format='NCHW'):
    if padding_algorithm not in ["SAME", "VALID", "EXPLICIT"]:
        raise ValueError("Unknown Attr(padding_algorithm): '%s'. "
                         "It can only be 'SAME' or 'VALID'." %
                         str(padding_algorithm))

    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError("Unknown Attr(data_format): '%s' ."
                         "It can only be 'NCHW' or 'NHWC'." % str(data_format))

    channel_last = (data_format == "NHWC")
    if channel_last:
        input = np.transpose(input, [0, 3, 1, 2])

C
chengduoZH 已提交
45
    in_n, in_c, in_h, in_w = input.shape
L
liym27 已提交
46 47 48
    f_n, f_c, f_h, f_w = filter.shape
    out_n = in_n
    out_c = f_n
C
chengduoZH 已提交
49 50
    assert f_c * group == in_c
    assert np.mod(out_c, group) == 0
M
minqiyang 已提交
51
    sub_out_c = out_c // group
L
liym27 已提交
52
    sub_f_n = f_n // group
C
chengduoZH 已提交
53

C
chengduoZH 已提交
54 55
    stride, pad, dilation = conv_param['stride'], conv_param['pad'], conv_param[
        'dilation']
L
liym27 已提交
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93

    # update pad and dilation
    def _get_padding_with_SAME(input_shape, pool_size, pool_stride):
        padding = []
        for input_size, filter_size, stride_size in zip(input_shape, pool_size,
                                                        pool_stride):
            out_size = int((input_size + stride_size - 1) / stride_size)
            pad_sum = np.max((
                (out_size - 1) * stride_size + filter_size - input_size, 0))
            pad_0 = int(pad_sum / 2)
            pad_1 = int(pad_sum - pad_0)
            padding.append(pad_0)
            padding.append(pad_1)
        return padding

    ksize = filter.shape[2:4]
    if padding_algorithm == "VALID":
        pad = [0, 0, 0, 0]
    elif padding_algorithm == "SAME":
        dilation = [1, 1]
        input_data_shape = []
        if data_format == "NCHW":
            input_data_shape = input.shape[2:4]
        elif data_format == "NHWC":
            input_data_shape = input.shape[1:3]
        pad = _get_padding_with_SAME(input_data_shape, ksize, stride)

    pad_h_0, pad_h_1 = pad[0], pad[0]
    pad_w_0, pad_w_1 = pad[1], pad[1]
    if len(pad) == 4:
        pad_h_0, pad_h_1 = pad[0], pad[1]
        pad_w_0, pad_w_1 = pad[2], pad[3]

    out_h = 1 + (in_h + pad_h_0 + pad_h_1 - (dilation[0] *
                                             (f_h - 1) + 1)) // stride[0]
    out_w = 1 + (in_w + pad_w_0 + pad_w_1 - (dilation[1] *
                                             (f_w - 1) + 1)) // stride[1]
    out = np.zeros((out_n, out_c, out_h, out_w))
C
chengduoZH 已提交
94

武毅 已提交
95 96
    d_bolck_h = (dilation[0] * (f_h - 1) + 1)
    d_bolck_w = (dilation[1] * (f_w - 1) + 1)
C
chengduoZH 已提交
97

L
liym27 已提交
98 99
    input_pad = np.pad(input, ((0, 0), (0, 0), (pad_h_0, pad_h_1),
                               (pad_w_0, pad_w_1)),
C
chengduoZH 已提交
100 101
                       mode='constant',
                       constant_values=0)
C
chengduoZH 已提交
102

L
liym27 已提交
103
    filter_dilation = np.zeros((f_n, f_c, d_bolck_h, d_bolck_w))
C
chengduoZH 已提交
104 105 106
    filter_dilation[:, :, 0:d_bolck_h:dilation[0], 0:d_bolck_w:dilation[
        1]] = filter

C
chengduoZH 已提交
107 108 109
    for i in range(out_h):
        for j in range(out_w):
            for g in range(group):
C
chengduoZH 已提交
110 111
                input_pad_masked = \
                    input_pad[:, g * f_c:(g + 1) * f_c,
C
chengduoZH 已提交
112 113
                    i * stride[0]:i * stride[0] + d_bolck_h,
                    j * stride[1]:j * stride[1] + d_bolck_w]
C
chengduoZH 已提交
114

L
liym27 已提交
115 116
                f_sub = filter_dilation[g * sub_f_n:(g + 1) * sub_f_n, :, :, :]
                # sub_f_n == sub_out_c
C
chengduoZH 已提交
117
                for k in range(sub_out_c):
L
liym27 已提交
118
                    # Multiplication of Corresponding Elements, then sum all
C
chengduoZH 已提交
119 120 121
                    out[:, g * sub_out_c + k, i, j] = \
                        np.sum(input_pad_masked * f_sub[k, :, :, :],
                               axis=(1, 2, 3))
C
chengduoZH 已提交
122

L
liym27 已提交
123 124 125
    if channel_last:
        out = np.transpose(out, [0, 2, 3, 1])

126
    return out, in_n, out_h, out_w, out_c
C
chengduoZH 已提交
127 128


L
liym27 已提交
129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
def create_test_cudnn_class(parent):
    @unittest.skipIf(not core.is_compiled_with_cuda(),
                     "core is not compiled with CUDA")
    class TestCUDNNCase(parent):
        def init_kernel_type(self):
            self.use_cudnn = True

    cls_name = "{0}_{1}".format(parent.__name__, "CUDNN")
    TestCUDNNCase.__name__ = cls_name
    globals()[cls_name] = TestCUDNNCase


def create_test_cudnn_fp16_class(parent, grad_check=True):
    @unittest.skipIf(not core.is_compiled_with_cuda(),
                     "core is not compiled with CUDA")
    class TestConv2DCUDNNFp16(parent):
        def init_kernel_type(self):
            self.use_cudnn = True
            self.dtype = np.float16

        def test_check_output(self):
            if core.is_compiled_with_cuda():
                place = core.CUDAPlace(0)
                if core.is_float16_supported(place):
                    self.check_output_with_place(place, atol=2e-2)

        def test_check_grad_no_filter(self):
            place = core.CUDAPlace(0)
            if core.is_float16_supported(place) and grad_check:
                self.check_grad_with_place(
                    place, ['Input'],
                    'Output',
                    max_relative_error=0.02,
                    no_grad_set=set(['Filter']))

        def test_check_grad_no_input(self):
            place = core.CUDAPlace(0)
            if core.is_float16_supported(place) and grad_check:
                self.check_grad_with_place(
                    place, ['Filter'],
                    'Output',
                    max_relative_error=0.02,
                    no_grad_set=set(['Input']))

    cls_name = "{0}_{1}".format(parent.__name__, "CUDNNFp16")
    TestConv2DCUDNNFp16.__name__ = cls_name
    globals()[cls_name] = TestConv2DCUDNNFp16


def create_test_channel_last_class(parent):
    class TestChannelLastCase(parent):
        def init_data_format(self):
            self.data_format = "NHWC"

        def init_test_case_2(self):
            N, C, H, W = self.input_size
            self.input_size = [N, H, W, C]

    cls_name = "{0}_{1}".format(parent.__name__, "ChannelLast")
    TestChannelLastCase.__name__ = cls_name
    globals()[cls_name] = TestChannelLastCase


def create_test_cudnn_channel_last_class(parent):
    @unittest.skipIf(not core.is_compiled_with_cuda(),
                     "core is not compiled with CUDA")
    class TestCudnnChannelLastCase(parent):
        def init_kernel_type(self):
            self.use_cudnn = True

        def init_data_format(self):
            self.data_format = "NHWC"

        def init_test_case_2(self):
            N, C, H, W = self.input_size
            self.input_size = [N, H, W, C]

    cls_name = "{0}_{1}".format(parent.__name__, "CudnnChannelLast")
    TestCudnnChannelLastCase.__name__ = cls_name
    globals()[cls_name] = TestCudnnChannelLastCase


def create_test_padding_SAME_class(parent):
    class TestPaddingSMAECase(parent):
        def init_paddings(self):
            self.pad = [0, 0]
            self.padding_algorithm = "SAME"

    cls_name = "{0}_{1}".format(parent.__name__, "PaddingSAMEOp")
    TestPaddingSMAECase.__name__ = cls_name
    globals()[cls_name] = TestPaddingSMAECase


def create_test_padding_VALID_class(parent):
    class TestPaddingVALIDCase(parent):
        def init_paddings(self):
            self.pad = [1, 1]
            self.padding_algorithm = "VALID"

    cls_name = "{0}_{1}".format(parent.__name__, "PaddingVALIDOp")
    TestPaddingVALIDCase.__name__ = cls_name
    globals()[cls_name] = TestPaddingVALIDCase


def create_test_cudnn_padding_SAME_class(parent):
    @unittest.skipIf(not core.is_compiled_with_cuda(),
                     "core is not compiled with CUDA")
    class TestCUDNNPaddingSMAECase(parent):
        def init_kernel_type(self):
            self.use_cudnn = True

        def init_paddings(self):
            self.pad = [1, 1]
            self.padding_algorithm = "SAME"

    cls_name = "{0}_{1}".format(parent.__name__, "CudnnPaddingSAMEOp")
    TestCUDNNPaddingSMAECase.__name__ = cls_name
    globals()[cls_name] = TestCUDNNPaddingSMAECase


def create_test_cudnn_padding_VALID_class(parent):
    @unittest.skipIf(not core.is_compiled_with_cuda(),
                     "core is not compiled with CUDA")
    class TestCUDNNPaddingVALIDCase(parent):
        def init_kernel_type(self):
            self.use_cudnn = True

        def init_paddings(self):
            self.pad = [1, 1]
            self.padding_algorithm = "VALID"

    cls_name = "{0}_{1}".format(parent.__name__, "CudnnPaddingVALIDOp")
    TestCUDNNPaddingVALIDCase.__name__ = cls_name
    globals()[cls_name] = TestCUDNNPaddingVALIDCase


H
hedaoyuan 已提交
265
class TestConv2dOp(OpTest):
266
    def setUp(self):
K
Kexin Zhao 已提交
267
        self.op_type = "conv2d"
268
        self.use_cudnn = False
269
        self.exhaustive_search = False
270
        self.use_cuda = False
271
        self.use_mkldnn = False
272
        self.fuse_relu_before_depthwise_conv = False
273
        self.data_format = "AnyLayout"
K
Kexin Zhao 已提交
274
        self.dtype = np.float32
K
Kexin Zhao 已提交
275
        self.init_kernel_type()
C
chengduoZH 已提交
276
        self.init_group()
C
chengduoZH 已提交
277
        self.init_dilation()
C
chengduoZH 已提交
278
        self.init_test_case()
C
chengduoZH 已提交
279

C
chengduoZH 已提交
280 281 282 283 284
        conv2d_param = {
            'stride': self.stride,
            'pad': self.pad,
            'dilation': self.dilations
        }
285

K
Kexin Zhao 已提交
286
        input = np.random.random(self.input_size).astype(self.dtype)
G
guomingz 已提交
287
        if not self.has_cuda():
288 289 290 291 292 293 294 295
            self.fuse_relu_before_depthwise_conv = False
        if self.fuse_relu_before_depthwise_conv:
            input = input - 0.5
            input -= (input < 0) * 0.1
            input += (input >= 0) * 0.1
            input2 = np.maximum(input, 0.0)
        else:
            input2 = input
G
guomingz 已提交
296
        filter = np.random.uniform(-1, 1, self.filter_size).astype(self.dtype)
L
liym27 已提交
297

298
        output, _, _, _, _ = conv2d_forward_naive(input2, filter, self.groups,
299 300
                                                  conv2d_param)
        output = output.astype(self.dtype)
K
Kexin Zhao 已提交
301 302

        self.inputs = {
K
Kexin Zhao 已提交
303 304
            'Input': OpTest.np_dtype_to_fluid_dtype(input),
            'Filter': OpTest.np_dtype_to_fluid_dtype(filter)
K
Kexin Zhao 已提交
305
        }
H
hedaoyuan 已提交
306
        self.attrs = {
C
chengduoZH 已提交
307 308
            'strides': self.stride,
            'paddings': self.pad,
C
chengduoZH 已提交
309
            'groups': self.groups,
310
            'dilations': self.dilations,
311
            'use_cudnn': self.use_cudnn,
312
            'use_mkldnn': self.use_mkldnn,
313
            'data_format': self.data_format,
314 315
            'fuse_relu_before_depthwise_conv':
            self.fuse_relu_before_depthwise_conv,
316
            'exhaustive_search': self.exhaustive_search
H
hedaoyuan 已提交
317
        }
318 319
        self.outputs = {'Output': output}

G
guomingz 已提交
320
    def has_cuda(self):
321 322
        return core.is_compiled_with_cuda() and (self.use_cudnn or
                                                 self.use_cuda)
323

H
hedaoyuan 已提交
324
    def test_check_output(self):
G
guomingz 已提交
325
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
326
        self.check_output_with_place(place, atol=1e-5)
H
hedaoyuan 已提交
327

H
hedaoyuan 已提交
328
    def test_check_grad(self):
K
Kexin Zhao 已提交
329 330
        if self.dtype == np.float16:
            return
G
guomingz 已提交
331
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
332
        self.check_grad_with_place(
Y
Yu Yang 已提交
333
            place, {'Input', 'Filter'}, 'Output', max_relative_error=0.02)
H
hedaoyuan 已提交
334

335
    def test_check_grad_no_filter(self):
K
Kexin Zhao 已提交
336 337
        if self.dtype == np.float16:
            return
G
guomingz 已提交
338
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
339 340 341 342 343
        self.check_grad_with_place(
            place, ['Input'],
            'Output',
            max_relative_error=0.02,
            no_grad_set=set(['Filter']))
344 345

    def test_check_grad_no_input(self):
K
Kexin Zhao 已提交
346 347
        if self.dtype == np.float16:
            return
G
guomingz 已提交
348
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
349 350 351 352 353
        self.check_grad_with_place(
            place, ['Filter'],
            'Output',
            max_relative_error=0.02,
            no_grad_set=set(['Input']))
354

C
chengduoZH 已提交
355 356 357 358 359
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
360
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
361 362
        self.filter_size = [6, f_c, 3, 3]

L
liym27 已提交
363 364 365
    def init_test_case_2(self):
        pass

C
chengduoZH 已提交
366 367 368
    def init_dilation(self):
        self.dilations = [1, 1]

C
chengduoZH 已提交
369
    def init_group(self):
H
hedaoyuan 已提交
370 371
        self.groups = 1

K
Kexin Zhao 已提交
372 373
    def init_kernel_type(self):
        pass
武毅 已提交
374

H
hedaoyuan 已提交
375

C
chengduoZH 已提交
376 377 378 379 380 381
class TestWithPad(TestConv2dOp):
    def init_test_case(self):
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
382
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
383 384 385 386 387 388 389 390 391
        self.filter_size = [6, f_c, 3, 3]


class TestWithStride(TestConv2dOp):
    def init_test_case(self):
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 6, 6]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
392
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
393 394 395
        self.filter_size = [6, f_c, 3, 3]


H
hedaoyuan 已提交
396
class TestWithGroup(TestConv2dOp):
C
chengduoZH 已提交
397
    def init_group(self):
H
hedaoyuan 已提交
398 399
        self.groups = 3

武毅 已提交
400

C
chengduoZH 已提交
401 402 403 404 405 406
class TestWith1x1(TestConv2dOp):
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
407
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
408 409 410 411 412 413
        self.filter_size = [6, f_c, 1, 1]

    def init_group(self):
        self.groups = 3


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
class TestWithDepthWise3x3(TestConv2dOp):
    def init_test_case(self):
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [3, 4, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [8, f_c, 3, 3]

    def init_dilation(self):
        self.dilations = [2, 2]

    def init_group(self):
        self.groups = 4


class TestWithDepthWise5x5(TestConv2dOp):
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
        self.input_size = [2, 4, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [8, f_c, 5, 5]

    def init_group(self):
        self.groups = 4


class TestWithDepthWise7x7(TestConv2dOp):
    def init_test_case(self):
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 8, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [16, f_c, 7, 7]

    def init_group(self):
        self.groups = 8


C
chengduoZH 已提交
456 457 458 459 460 461
class TestWithDilation(TestConv2dOp):
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
        self.input_size = [2, 3, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
462
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
463
        self.filter_size = [6, f_c, 3, 3]
C
chengduoZH 已提交
464

C
chengduoZH 已提交
465 466
    def init_dilation(self):
        self.dilations = [2, 2]
C
chengduoZH 已提交
467

C
chengduoZH 已提交
468
    def init_group(self):
C
chengduoZH 已提交
469
        self.groups = 3
武毅 已提交
470

C
chengduoZH 已提交
471

472 473 474 475 476 477
class TestWithInput1x1Filter1x1(TestConv2dOp):
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
        self.input_size = [2, 3, 1, 1]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
478
        f_c = self.input_size[1] // self.groups
479 480 481 482 483 484
        self.filter_size = [6, f_c, 1, 1]

    def init_group(self):
        self.groups = 3


485
#----------------Conv2dCUDNN----------------
C
chengduoZH 已提交
486

C
chengduo 已提交
487 488 489 490 491 492
create_test_cudnn_class(TestConv2dOp)
create_test_cudnn_class(TestWithPad)
create_test_cudnn_class(TestWithStride)
create_test_cudnn_class(TestWithGroup)
create_test_cudnn_class(TestWith1x1)
create_test_cudnn_class(TestWithInput1x1Filter1x1)
K
Kexin Zhao 已提交
493

L
liym27 已提交
494
#----------------Conv2dCUDNN fp16----------------
C
chengduo 已提交
495

C
chengduo 已提交
496 497 498 499 500 501
create_test_cudnn_fp16_class(TestConv2dOp, grad_check=False)
create_test_cudnn_fp16_class(TestWithPad, grad_check=False)
create_test_cudnn_fp16_class(TestWithStride, grad_check=False)
create_test_cudnn_fp16_class(TestWithGroup, grad_check=False)
create_test_cudnn_fp16_class(TestWith1x1, grad_check=False)
create_test_cudnn_fp16_class(TestWithInput1x1Filter1x1, grad_check=False)
C
chengduo 已提交
502

L
liym27 已提交
503
#----------------TestDepthwiseConv -----
K
Kexin Zhao 已提交
504 505


506 507
class TestDepthwiseConv(TestConv2dOp):
    def init_test_case(self):
508
        self.use_cuda = True
509 510 511 512 513
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
514
        f_c = self.input_size[1] // self.groups
515
        self.filter_size = [3, f_c, 3, 3]
516
        self.op_type = "depthwise_conv2d"
517 518 519 520


class TestDepthwiseConv2(TestConv2dOp):
    def init_test_case(self):
521 522 523 524 525 526 527 528 529 530 531 532 533 534
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [3, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


class TestDepthwiseConv3(TestConv2dOp):
    def init_test_case(self):
        self.use_cuda = True
535 536 537 538 539
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
540
        f_c = self.input_size[1] // self.groups
541
        self.filter_size = [6, f_c, 3, 3]
542
        self.op_type = "depthwise_conv2d"
543 544


545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572
class TestDepthwiseConvWithDilation(TestConv2dOp):
    def init_test_case(self):
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


class TestDepthwiseConvWithDilation2(TestConv2dOp):
    def init_test_case(self):
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644
class TestDepthwiseConvandFuse(TestConv2dOp):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [3, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


class TestDepthwiseConv2andFuse(TestConv2dOp):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [3, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


class TestDepthwiseConv3andFuse(TestConv2dOp):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


class TestDepthwiseConvWithDilationandFuse(TestConv2dOp):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


class TestDepthwiseConvWithDilation2andFuse(TestConv2dOp):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"


645 646 647 648 649 650
class TestCUDNNExhaustiveSearch(TestConv2dOp):
    def init_kernel_type(self):
        self.use_cudnn = True
        self.exhaustive_search = True


651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
class TestConv2dOpError(OpTest):
    def test_errors(self):
        with program_guard(Program(), Program()):

            def test_Variable():
                # the input of conv2d must be Variable.
                x1 = fluid.create_lod_tensor(
                    np.array([-1, 3, 5, 5]), [[1, 1, 1, 1]], fluid.CPUPlace())
                fluid.layers.conv2d(x1, 1, 1)

            self.assertRaises(TypeError, test_Variable)

            def test_dtype():
                # the input dtype of conv2d must be float16 or float32 or float64
                # float16 only can be set on GPU place
                x2 = fluid.layers.data(
                    name='x2', shape=[3, 4, 5, 6], dtype="int32")
                fluid.layers.conv2d(x2, 1, 1)

            self.assertRaises(TypeError, test_dtype)


673 674
# Please Don't remove the following code.
# Currently, CI use cudnn V5.0 which not support dilation conv.
675
# class TestCUDNNWithDilation(TestWithDilation):
C
chengduoZH 已提交
676 677 678
#     def init_op_type(self):
#         self.op_type = "conv_cudnn"

L
liym27 已提交
679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 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 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 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377
# ---- test asymmetric padding ----


class TestConv2dOp_v2(OpTest):
    def setUp(self):
        self.op_type = "conv2d"
        self.use_cudnn = False
        self.exhaustive_search = False
        self.use_cuda = False
        self.use_mkldnn = False
        self.fuse_relu_before_depthwise_conv = False
        self.dtype = np.float32
        self.init_kernel_type()
        self.init_group()
        self.init_dilation()
        self.init_data_format()
        self.init_test_case()

        self.init_paddings()
        self.init_test_case_2()

        conv2d_param = {
            'stride': self.stride,
            'pad': self.pad,
            'dilation': self.dilations
        }

        input = np.random.random(self.input_size).astype(self.dtype)
        if not self.has_cuda():
            self.fuse_relu_before_depthwise_conv = False
        if self.fuse_relu_before_depthwise_conv:
            input = input - 0.5
            input -= (input < 0) * 0.1
            input += (input >= 0) * 0.1
            input2 = np.maximum(input, 0.0)
        else:
            input2 = input
        filter = np.random.uniform(-1, 1, self.filter_size).astype(self.dtype)
        output, _, _, _, _ = conv2d_forward_naive(
            input2, filter, self.groups, conv2d_param, self.padding_algorithm,
            self.data_format)
        output = output.astype(self.dtype)

        self.inputs = {
            'Input': OpTest.np_dtype_to_fluid_dtype(input),
            'Filter': OpTest.np_dtype_to_fluid_dtype(filter)
        }
        self.attrs = {
            'strides': self.stride,
            'paddings': self.pad,
            'padding_algorithm': self.padding_algorithm,
            'groups': self.groups,
            'dilations': self.dilations,
            'use_cudnn': self.use_cudnn,
            'use_mkldnn': self.use_mkldnn,
            'data_format': self.data_format,
            'fuse_relu_before_depthwise_conv':
            self.fuse_relu_before_depthwise_conv,
            'exhaustive_search': self.exhaustive_search
        }
        self.outputs = {'Output': output}

    def has_cuda(self):
        return core.is_compiled_with_cuda() and (self.use_cudnn or
                                                 self.use_cuda)

    def test_check_output(self):
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
        self.check_output_with_place(place, atol=1e-5)

    def test_check_grad(self):
        if self.dtype == np.float16:
            return
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
        self.check_grad_with_place(
            place, {'Input', 'Filter'}, 'Output', max_relative_error=0.02)

    def test_check_grad_no_filter(self):
        if self.dtype == np.float16:
            return
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
        self.check_grad_with_place(
            place, ['Input'],
            'Output',
            max_relative_error=0.02,
            no_grad_set=set(['Filter']))

    def test_check_grad_no_input(self):
        if self.dtype == np.float16:
            return
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
        self.check_grad_with_place(
            place, ['Filter'],
            'Output',
            max_relative_error=0.02,
            no_grad_set=set(['Input']))

    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]

    def init_dilation(self):
        self.dilations = [1, 1]

    def init_group(self):
        self.groups = 1

    def init_kernel_type(self):
        pass

    def init_paddings(self):
        self.pad = [0, 0]
        self.padding_algorithm = "EXPLICIT"

    def init_data_format(self):
        self.data_format = "NCHW"

    def init_test_case_2(self):
        pass


class TestConv2dOp_AsyPadding(TestConv2dOp_v2):
    def init_paddings(self):
        self.pad = [0, 0, 1, 2]
        self.padding_algorithm = "EXPLICIT"


class TestWithPad_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]

    def init_paddings(self):
        self.pad = [2, 1, 3, 2]
        self.padding_algorithm = "EXPLICIT"


class TestWithStride_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [2, 2]
        self.input_size = [2, 3, 6, 6]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]

    def init_paddings(self):
        self.pad = [2, 1, 3, 2]
        self.padding_algorithm = "EXPLICIT"


class TestWithGroup_AsyPadding(TestConv2dOp_v2):
    def init_group(self):
        self.groups = 3


class TestWith1x1_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 1, 1]

    def init_group(self):
        self.groups = 3

    def init_paddings(self):
        self.pad = [2, 2, 4, 0]
        self.padding_algorithm = "EXPLICIT"


class TestWithDepthWise3x3_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [1, 1]
        self.input_size = [3, 4, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [8, f_c, 3, 3]

    def init_dilation(self):
        self.dilations = [2, 2]

    def init_group(self):
        self.groups = 4

    def init_paddings(self):
        self.pad = [1, 3, 2, 1]
        self.padding_algorithm = "EXPLICIT"


class TestWithDepthWise5x5_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [1, 1]
        self.input_size = [2, 4, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [8, f_c, 5, 5]

    def init_group(self):
        self.groups = 4

    def init_paddings(self):
        self.pad = [0, 1, 1, 0]
        self.padding_algorithm = "EXPLICIT"


class TestWithDepthWise7x7_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [2, 2]
        self.input_size = [2, 8, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [16, f_c, 7, 7]

    def init_group(self):
        self.groups = 8

    def init_paddings(self):
        self.pad = [1, 3, 4, 1]
        self.padding_algorithm = "EXPLICIT"


class TestWithDilation_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [1, 1]
        self.input_size = [2, 3, 10, 10]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]

    def init_dilation(self):
        self.dilations = [2, 2]

    def init_group(self):
        self.groups = 3

    def init_paddings(self):
        self.pad = [0, 1, 3, 0]
        self.padding_algorithm = "EXPLICIT"


class TestWithInput1x1Filter1x1_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.stride = [1, 1]
        self.input_size = [2, 3, 1, 1]  # NCHW
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 1, 1]

    def init_group(self):
        self.groups = 3

    def init_paddings(self):
        self.pad = [0, 3, 4, 0]
        self.padding_algorithm = "EXPLICIT"


create_test_cudnn_class(TestConv2dOp_AsyPadding)
create_test_cudnn_class(TestWithPad_AsyPadding)
create_test_cudnn_class(TestWithStride_AsyPadding)
create_test_cudnn_class(TestWithGroup_AsyPadding)
create_test_cudnn_class(TestWith1x1_AsyPadding)
create_test_cudnn_class(TestWithInput1x1Filter1x1_AsyPadding)


class TestDepthwiseConv_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.use_cuda = True
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [3, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [1, 1, 0, 1]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConv2_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.use_cuda = True
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [3, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [0, 1, 0, 2]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConv3_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.use_cuda = True
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [1, 1, 0, 0]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConvWithDilation_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [1, 1, 2, 1]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConvWithDilation2_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [0, 1, 1, 0]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConvandFuse_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [3, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [2, 1, 2, 3]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConv2andFuse_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [3, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [1, 1, 1, 2]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConv3andFuse_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [1, 2, 0, 2]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConvWithDilationandFuse_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [2, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [2, 1, 1, 0]
        self.padding_algorithm = "EXPLICIT"


class TestDepthwiseConvWithDilation2andFuse_AsyPadding(TestConv2dOp_v2):
    def init_test_case(self):
        self.fuse_relu_before_depthwise_conv = True
        self.use_cuda = True
        self.pad = [1, 1]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.groups = 3
        self.dilations = [2, 2]
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [6, f_c, 3, 3]
        self.op_type = "depthwise_conv2d"

    def init_paddings(self):
        self.pad = [1, 3, 1, 3]
        self.padding_algorithm = "EXPLICIT"


#---------- test SAME VALID -----------
create_test_padding_SAME_class(TestConv2dOp_AsyPadding)
create_test_padding_SAME_class(TestWithPad_AsyPadding)
create_test_padding_SAME_class(TestWithStride_AsyPadding)
create_test_padding_SAME_class(TestWithGroup_AsyPadding)
create_test_padding_SAME_class(TestWithInput1x1Filter1x1_AsyPadding)

create_test_padding_VALID_class(TestConv2dOp_AsyPadding)
create_test_padding_VALID_class(TestWithPad_AsyPadding)
create_test_padding_VALID_class(TestWithStride_AsyPadding)
create_test_padding_VALID_class(TestWithGroup_AsyPadding)
create_test_padding_VALID_class(TestWithInput1x1Filter1x1_AsyPadding)

create_test_cudnn_padding_SAME_class(TestConv2dOp_AsyPadding)
create_test_cudnn_padding_SAME_class(TestWithPad_AsyPadding)
create_test_cudnn_padding_SAME_class(TestWithStride_AsyPadding)
create_test_cudnn_padding_SAME_class(TestWithGroup_AsyPadding)
create_test_cudnn_padding_SAME_class(TestWithInput1x1Filter1x1_AsyPadding)

create_test_cudnn_padding_VALID_class(TestConv2dOp_AsyPadding)
create_test_cudnn_padding_VALID_class(TestWithPad_AsyPadding)
create_test_cudnn_padding_VALID_class(TestWithStride_AsyPadding)
create_test_cudnn_padding_VALID_class(TestWithGroup_AsyPadding)
create_test_cudnn_padding_VALID_class(TestWithInput1x1Filter1x1_AsyPadding)

# depthwise conv2d

create_test_padding_SAME_class(TestDepthwiseConv_AsyPadding)
create_test_padding_SAME_class(TestDepthwiseConvWithDilation_AsyPadding)
create_test_padding_SAME_class(TestDepthwiseConvandFuse_AsyPadding)
create_test_padding_SAME_class(TestDepthwiseConvWithDilationandFuse_AsyPadding)

create_test_padding_VALID_class(TestDepthwiseConv_AsyPadding)
create_test_padding_VALID_class(TestDepthwiseConvWithDilation_AsyPadding)
create_test_padding_VALID_class(TestDepthwiseConvandFuse_AsyPadding)
create_test_padding_VALID_class(TestDepthwiseConvWithDilationandFuse_AsyPadding)

# ------------ test channel last ---------
create_test_channel_last_class(TestConv2dOp_AsyPadding)
create_test_channel_last_class(TestWithPad_AsyPadding)
create_test_channel_last_class(TestWithGroup_AsyPadding)
create_test_channel_last_class(TestWith1x1_AsyPadding)
create_test_channel_last_class(TestWithInput1x1Filter1x1_AsyPadding)

create_test_channel_last_class(TestDepthwiseConv_AsyPadding)
create_test_channel_last_class(TestDepthwiseConvWithDilation2_AsyPadding)
create_test_channel_last_class(TestDepthwiseConvandFuse_AsyPadding)
create_test_channel_last_class(TestDepthwiseConvWithDilationandFuse_AsyPadding)

create_test_cudnn_channel_last_class(TestConv2dOp_AsyPadding)
create_test_cudnn_channel_last_class(TestWithPad_AsyPadding)
create_test_cudnn_channel_last_class(TestWithStride_AsyPadding)
create_test_cudnn_channel_last_class(TestWithGroup_AsyPadding)
create_test_cudnn_channel_last_class(TestWithDilation_AsyPadding)


# --------- test python API ---------------
class TestConv2dAPI(OpTest):
    def test_api(self):

        input_NHWC = fluid.layers.data(
            name="input_NHWC",
            shape=[2, 5, 5, 3],
            append_batch_size=False,
            dtype="float32")

        input_NCHW = fluid.layers.data(
            name="input_NCHW",
            shape=[2, 3, 5, 5],
            append_batch_size=False,
            dtype="float32")

        fluid.layers.conv2d(
            input=input_NHWC,
            num_filters=3,
            filter_size=[3, 3],
            stride=[1, 1],
            padding=0,
            dilation=[1, 1],
            groups=1,
            data_format="NCHW")

        fluid.layers.conv2d(
            input=input_NCHW,
            num_filters=3,
            filter_size=[3, 3],
            stride=[1, 1],
            padding=[1, 2, 1, 0],
            dilation=[1, 1],
            groups=1,
            data_format="NCHW")

        fluid.layers.conv2d(
            input=input_NCHW,
            num_filters=3,
            filter_size=[3, 3],
            stride=[1, 1],
            padding=[[0, 0], [0, 0], [1, 1], [1, 1]],
            dilation=[1, 1],
            groups=1,
            data_format="NCHW")

        fluid.layers.conv2d(
            input=input_NHWC,
            num_filters=3,
            filter_size=[3, 3],
            stride=[1, 1],
            padding=[[0, 0], [1, 1], [1, 1], [0, 0]],
            dilation=[1, 1],
            groups=1,
            data_format="NHWC")

        fluid.layers.conv2d(
            input=input_NCHW,
            num_filters=3,
            filter_size=[3, 3],
            stride=[1, 1],
            padding="SAME",
            dilation=[1, 1],
            groups=1,
            data_format="NCHW")

        fluid.layers.conv2d(
            input=input_NCHW,
            num_filters=3,
            filter_size=[3, 3],
            stride=[1, 1],
            padding="VALID",
            dilation=[1, 1],
            groups=1,
            data_format="NCHW")


class TestConv2dAPI_Error(OpTest):
    def test_api(self):
        input = fluid.layers.data(
            name="input",
            shape=[2, 5, 5, 5],
            append_batch_size=False,
            dtype="float32")

        # ValueError: cudnn
        def run_1():
            fluid.layers.conv2d(
                input=input,
                num_filters=3,
                filter_size=[3, 3],
                stride=[1, 1],
                padding=0,
                dilation=[1, 1],
                groups=1,
                use_cudnn=[0],
                data_format="NCHW")

        self.assertRaises(ValueError, run_1)

        # ValueError: data_format
        def run_2():
            fluid.layers.conv2d(
                input=input,
                num_filters=3,
                filter_size=[3, 3],
                stride=[1, 1],
                padding=0,
                dilation=[1, 1],
                groups=1,
                use_cudnn=False,
                data_format="NCHWC")

        self.assertRaises(ValueError, run_2)

        # ValueError: padding
        def run_3():
            fluid.layers.conv2d(
                input=input,
                num_filters=3,
                filter_size=[3, 3],
                stride=[1, 1],
                padding="SAMEE",
                dilation=[1, 1],
                groups=1,
                use_cudnn=False,
                data_format="NCHW")

        self.assertRaises(ValueError, run_3)

        def run_4():
            fluid.layers.conv2d(
                input=input,
                num_filters=3,
                filter_size=[3, 3],
                stride=[1, 1],
                padding=[[0, 1], [0, 1], [0, 1], [0, 1]],
                dilation=[1, 1],
                groups=1,
                use_cudnn=False,
                data_format="NCHW")

        self.assertRaises(ValueError, run_4)

        def run_5():
            fluid.layers.conv2d(
                input=input,
                num_filters=3,
                filter_size=[3, 3],
                stride=[1, 1],
                padding=[[0, 1], [0, 1], [0, 1], [0, 1]],
                dilation=[1, 1],
                groups=1,
                use_cudnn=False,
                data_format="NHWC")

        self.assertRaises(ValueError, run_5)

        # ValueError: channel dimmention
        x = fluid.layers.data(
            name="x",
            shape=[2, 5, 5, -1],
            append_batch_size=False,
            dtype="float32")

        def run_6():
            fluid.layers.conv2d(
                input=x,
                num_filters=3,
                filter_size=[3, 3],
                stride=[1, 1],
                padding=0,
                dilation=[1, 1],
                groups=1,
                use_cudnn=False,
                data_format="NHWC")

        self.assertRaises(ValueError, run_6)

        # ValueError: groups
        def run_7():
            fluid.layers.conv2d(
                input=input,
                num_filters=3,
                filter_size=[3, 3],
                stride=[1, 1],
                padding=0,
                dilation=[1, 1],
                groups=3,
                use_cudnn=False,
                data_format="NHWC")

        self.assertRaises(ValueError, run_7)


1378 1379
if __name__ == '__main__':
    unittest.main()