test_conv2d_op.py 50.4 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
21
import paddle.fluid.core as core
L
liym27 已提交
22
import paddle.fluid as fluid
23 24
from op_test import OpTest
from paddle.fluid import Program, program_guard
25 26


L
liym27 已提交
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
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
46
    in_n, in_c, in_h, in_w = input.shape
L
liym27 已提交
47 48 49
    f_n, f_c, f_h, f_w = filter.shape
    out_n = in_n
    out_c = f_n
C
chengduoZH 已提交
50 51
    assert f_c * group == in_c
    assert np.mod(out_c, group) == 0
M
minqiyang 已提交
52
    sub_out_c = out_c // group
L
liym27 已提交
53
    sub_f_n = f_n // group
C
chengduoZH 已提交
54

C
chengduoZH 已提交
55 56
    stride, pad, dilation = conv_param['stride'], conv_param['pad'], conv_param[
        'dilation']
L
liym27 已提交
57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76

    # 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]
77
        input_data_shape = input.shape[2:4]
L
liym27 已提交
78 79 80 81 82 83 84 85 86 87 88 89
        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
90

武毅 已提交
91 92
    d_bolck_h = (dilation[0] * (f_h - 1) + 1)
    d_bolck_w = (dilation[1] * (f_w - 1) + 1)
C
chengduoZH 已提交
93

L
liym27 已提交
94 95
    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
96 97
                       mode='constant',
                       constant_values=0)
C
chengduoZH 已提交
98

L
liym27 已提交
99
    filter_dilation = np.zeros((f_n, f_c, d_bolck_h, d_bolck_w))
C
chengduoZH 已提交
100 101 102
    filter_dilation[:, :, 0:d_bolck_h:dilation[0], 0:d_bolck_w:dilation[
        1]] = filter

C
chengduoZH 已提交
103 104 105
    for i in range(out_h):
        for j in range(out_w):
            for g in range(group):
C
chengduoZH 已提交
106 107
                input_pad_masked = \
                    input_pad[:, g * f_c:(g + 1) * f_c,
C
chengduoZH 已提交
108 109
                    i * stride[0]:i * stride[0] + d_bolck_h,
                    j * stride[1]:j * stride[1] + d_bolck_w]
C
chengduoZH 已提交
110

L
liym27 已提交
111 112
                f_sub = filter_dilation[g * sub_f_n:(g + 1) * sub_f_n, :, :, :]
                # sub_f_n == sub_out_c
C
chengduoZH 已提交
113
                for k in range(sub_out_c):
L
liym27 已提交
114
                    # Multiplication of Corresponding Elements, then sum all
C
chengduoZH 已提交
115 116 117
                    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
118

L
liym27 已提交
119 120 121
    if channel_last:
        out = np.transpose(out, [0, 2, 3, 1])

122
    return out, in_n, out_h, out_w, out_c
C
chengduoZH 已提交
123 124


L
liym27 已提交
125 126 127 128 129 130
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
131 132
            self.dtype = np.float32 if core.is_compiled_with_rocm(
            ) else np.float64
L
liym27 已提交
133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156

    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(
157
                    place, ['Input'], 'Output', no_grad_set=set(['Filter']))
L
liym27 已提交
158 159 160 161 162

        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(
163
                    place, ['Filter'], 'Output', no_grad_set=set(['Input']))
L
liym27 已提交
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

    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
190 191
            self.dtype = np.float32 if core.is_compiled_with_rocm(
            ) else np.float64
L
liym27 已提交
192 193 194 195 196 197 198 199 200 201 202 203 204

        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


205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222
def create_test_cudnn_channel_last_fp16_class(parent, grad_check=True):
    @unittest.skipIf(not core.is_compiled_with_cuda(),
                     "core is not compiled with CUDA")
    class TestCudnnChannelLastFp16(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(
223
                    place, ['Input'], 'Output', no_grad_set=set(['Filter']))
224 225 226 227 228

        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(
229
                    place, ['Filter'], 'Output', no_grad_set=set(['Input']))
230 231 232 233 234 235 236 237 238 239 240 241 242

        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__, "CudnnChannelLastFp16")
    TestCudnnChannelLastFp16.__name__ = cls_name
    globals()[cls_name] = TestCudnnChannelLastFp16


L
liym27 已提交
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
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
271 272
            self.dtype = np.float32 if core.is_compiled_with_rocm(
            ) else np.float64
L
liym27 已提交
273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288

        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
289 290
            self.dtype = np.float32 if core.is_compiled_with_rocm(
            ) else np.float64
L
liym27 已提交
291 292 293 294 295 296 297 298 299 300

        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


C
cnn 已提交
301
class TestConv2DOp(OpTest):
302
    def setUp(self):
K
Kexin Zhao 已提交
303
        self.op_type = "conv2d"
304
        self.use_cudnn = False
305
        self.exhaustive_search = False
306
        self.use_cuda = False
307
        self.use_mkldnn = False
308
        self.fuse_relu_before_depthwise_conv = False
309
        self.data_format = "AnyLayout"
310
        self.dtype = np.float64
K
Kexin Zhao 已提交
311
        self.init_kernel_type()
C
chengduoZH 已提交
312
        self.init_group()
C
chengduoZH 已提交
313
        self.init_dilation()
C
chengduoZH 已提交
314
        self.init_test_case()
C
chengduoZH 已提交
315

C
chengduoZH 已提交
316 317 318 319 320
        conv2d_param = {
            'stride': self.stride,
            'pad': self.pad,
            'dilation': self.dilations
        }
321

K
Kexin Zhao 已提交
322
        input = np.random.random(self.input_size).astype(self.dtype)
G
guomingz 已提交
323
        if not self.has_cuda():
324 325 326 327 328 329 330 331
            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
332
        filter = np.random.uniform(-1, 1, self.filter_size).astype(self.dtype)
L
liym27 已提交
333

334
        output, _, _, _, _ = conv2d_forward_naive(input2, filter, self.groups,
335 336
                                                  conv2d_param)
        output = output.astype(self.dtype)
K
Kexin Zhao 已提交
337 338

        self.inputs = {
K
Kexin Zhao 已提交
339 340
            'Input': OpTest.np_dtype_to_fluid_dtype(input),
            'Filter': OpTest.np_dtype_to_fluid_dtype(filter)
K
Kexin Zhao 已提交
341
        }
H
hedaoyuan 已提交
342
        self.attrs = {
C
chengduoZH 已提交
343 344
            'strides': self.stride,
            'paddings': self.pad,
C
chengduoZH 已提交
345
            'groups': self.groups,
346
            'dilations': self.dilations,
347
            'use_cudnn': self.use_cudnn,
348
            'use_mkldnn': self.use_mkldnn,
349
            'data_format': self.data_format,
350 351
            'fuse_relu_before_depthwise_conv':
            self.fuse_relu_before_depthwise_conv,
352
            'exhaustive_search': self.exhaustive_search
H
hedaoyuan 已提交
353
        }
354 355
        self.outputs = {'Output': output}

G
guomingz 已提交
356
    def has_cuda(self):
357 358
        return core.is_compiled_with_cuda() and (self.use_cudnn or
                                                 self.use_cuda)
359

H
hedaoyuan 已提交
360
    def test_check_output(self):
G
guomingz 已提交
361
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
362 363 364
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
        self.check_output_with_place(
            place, atol=1e-5, check_dygraph=(self.use_mkldnn == False))
H
hedaoyuan 已提交
365

H
hedaoyuan 已提交
366
    def test_check_grad(self):
367 368
        if self.dtype == np.float16 or (hasattr(self, "no_need_check_grad") and
                                        self.no_need_check_grad == True):
K
Kexin Zhao 已提交
369
            return
G
guomingz 已提交
370
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
371
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
372
        self.check_grad_with_place(
373 374 375 376
            place, {'Input', 'Filter'},
            'Output',
            max_relative_error=0.02,
            check_dygraph=(self.use_mkldnn == False))
H
hedaoyuan 已提交
377

378
    def test_check_grad_no_filter(self):
379 380
        if self.dtype == np.float16 or (hasattr(self, "no_need_check_grad") and
                                        self.no_need_check_grad == True):
K
Kexin Zhao 已提交
381
            return
G
guomingz 已提交
382
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
383
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
384 385 386 387
        self.check_grad_with_place(
            place, ['Input'],
            'Output',
            max_relative_error=0.02,
388 389
            no_grad_set=set(['Filter']),
            check_dygraph=(self.use_mkldnn == False))
390 391

    def test_check_grad_no_input(self):
392 393
        if self.dtype == np.float16 or (hasattr(self, "no_need_check_grad") and
                                        self.no_need_check_grad == True):
K
Kexin Zhao 已提交
394
            return
G
guomingz 已提交
395
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
396
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
397 398 399
        self.check_grad_with_place(
            place, ['Filter'],
            'Output',
400 401
            no_grad_set=set(['Input']),
            check_dygraph=(self.use_mkldnn == False))
402

C
chengduoZH 已提交
403 404 405 406 407
    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
408
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
409 410
        self.filter_size = [6, f_c, 3, 3]

L
liym27 已提交
411 412 413
    def init_test_case_2(self):
        pass

C
chengduoZH 已提交
414 415 416
    def init_dilation(self):
        self.dilations = [1, 1]

C
chengduoZH 已提交
417
    def init_group(self):
H
hedaoyuan 已提交
418 419
        self.groups = 1

K
Kexin Zhao 已提交
420 421
    def init_kernel_type(self):
        pass
武毅 已提交
422

H
hedaoyuan 已提交
423

C
cnn 已提交
424
class TestWithPad(TestConv2DOp):
C
chengduoZH 已提交
425 426 427 428 429
    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
430
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
431 432 433
        self.filter_size = [6, f_c, 3, 3]


C
cnn 已提交
434
class TestWithStride(TestConv2DOp):
C
chengduoZH 已提交
435 436 437 438 439
    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
440
        f_c = self.input_size[1] // self.groups
C
chengduoZH 已提交
441 442 443
        self.filter_size = [6, f_c, 3, 3]


C
cnn 已提交
444
class TestWithGroup(TestConv2DOp):
Z
zhupengyang 已提交
445 446 447 448 449 450 451 452
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.group = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [18, f_c, 3, 3]
H
hedaoyuan 已提交
453

武毅 已提交
454

C
cnn 已提交
455
class TestWith1x1(TestConv2DOp):
C
chengduoZH 已提交
456 457 458 459 460
    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
461
        f_c = self.input_size[1] // self.groups
Z
zhupengyang 已提交
462
        self.filter_size = [120, f_c, 1, 1]
C
chengduoZH 已提交
463 464 465 466 467

    def init_group(self):
        self.groups = 3


C
cnn 已提交
468
class TestWithDepthWise3x3(TestConv2DOp):
469 470 471 472 473 474
    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
Z
zhupengyang 已提交
475
        self.filter_size = [12, f_c, 3, 3]
476 477 478 479 480 481 482 483

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

    def init_group(self):
        self.groups = 4


C
cnn 已提交
484
class TestWithDepthWise5x5(TestConv2DOp):
485 486 487 488 489 490 491 492 493 494 495 496
    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


C
cnn 已提交
497
class TestWithDepthWise7x7(TestConv2DOp):
498 499 500 501 502 503 504 505 506 507 508 509
    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
cnn 已提交
510
class TestWithDilation(TestConv2DOp):
C
chengduoZH 已提交
511 512 513 514 515
    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
516
        f_c = self.input_size[1] // self.groups
Z
zhupengyang 已提交
517
        self.filter_size = [12, f_c, 3, 3]
C
chengduoZH 已提交
518

C
chengduoZH 已提交
519 520
    def init_dilation(self):
        self.dilations = [2, 2]
C
chengduoZH 已提交
521

C
chengduoZH 已提交
522
    def init_group(self):
C
chengduoZH 已提交
523
        self.groups = 3
武毅 已提交
524

C
chengduoZH 已提交
525

C
cnn 已提交
526
class TestWithInput1x1Filter1x1(TestConv2DOp):
527 528 529
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 1]
Z
zhupengyang 已提交
530
        self.input_size = [100, 3, 1, 1]  # NCHW
531
        assert np.mod(self.input_size[1], self.groups) == 0
M
minqiyang 已提交
532
        f_c = self.input_size[1] // self.groups
Z
zhupengyang 已提交
533
        self.filter_size = [120, f_c, 1, 1]
534 535 536 537 538

    def init_group(self):
        self.groups = 3


C
cnn 已提交
539
#----------------Conv2DCUDNN----------------
C
chengduoZH 已提交
540

C
cnn 已提交
541
create_test_cudnn_class(TestConv2DOp)
C
chengduo 已提交
542 543 544 545 546
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
547

C
cnn 已提交
548
#----------------Conv2DCUDNN fp16----------------
C
chengduo 已提交
549

C
cnn 已提交
550
create_test_cudnn_fp16_class(TestConv2DOp, grad_check=False)
C
chengduo 已提交
551 552 553 554 555
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
556

L
liym27 已提交
557
#----------------TestDepthwiseConv -----
K
Kexin Zhao 已提交
558 559


C
cnn 已提交
560
class TestDepthwiseConv(TestConv2DOp):
561
    def init_test_case(self):
562
        self.use_cuda = True
563 564 565 566 567
        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
568
        f_c = self.input_size[1] // self.groups
Z
zhupengyang 已提交
569
        self.filter_size = [12, f_c, 3, 3]
570
        self.op_type = "depthwise_conv2d"
571 572


C
cnn 已提交
573
class TestDepthwiseConv2(TestConv2DOp):
574
    def init_test_case(self):
575 576 577 578 579 580 581
        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
Z
zhupengyang 已提交
582
        self.filter_size = [12, f_c, 3, 3]
583 584 585
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
586
class TestDepthwiseConv3(TestConv2DOp):
587 588
    def init_test_case(self):
        self.use_cuda = True
589 590 591 592 593
        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
594
        f_c = self.input_size[1] // self.groups
Z
zhupengyang 已提交
595
        self.filter_size = [24, f_c, 3, 3]
596
        self.op_type = "depthwise_conv2d"
597 598


C
cnn 已提交
599
class TestDepthwiseConvWithDilation(TestConv2DOp):
600 601 602 603 604 605 606 607 608
    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
Z
zhupengyang 已提交
609
        self.filter_size = [24, f_c, 3, 3]
610 611 612
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
613
class TestDepthwiseConvWithDilation2(TestConv2DOp):
614 615 616 617 618 619 620 621 622
    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
Z
zhupengyang 已提交
623
        self.filter_size = [24, f_c, 3, 3]
624 625 626
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
627
class TestDepthwiseConvandFuse(TestConv2DOp):
628 629 630 631 632 633 634 635 636
    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
Z
zhupengyang 已提交
637
        self.filter_size = [12, f_c, 3, 3]
638 639 640
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
641
class TestDepthwiseConv2andFuse(TestConv2DOp):
642 643 644 645 646 647 648 649 650
    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
Z
zhupengyang 已提交
651
        self.filter_size = [12, f_c, 3, 3]
652 653 654
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
655
class TestDepthwiseConv3andFuse(TestConv2DOp):
656 657 658 659 660 661 662 663 664
    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
Z
zhupengyang 已提交
665
        self.filter_size = [24, f_c, 3, 3]
666 667 668
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
669
class TestDepthwiseConvWithDilationandFuse(TestConv2DOp):
670 671 672 673 674 675 676 677 678 679
    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
Z
zhupengyang 已提交
680
        self.filter_size = [24, f_c, 3, 3]
681 682 683
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
684
class TestDepthwiseConvWithDilation2andFuse(TestConv2DOp):
685 686 687 688 689 690 691 692 693 694
    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
Z
zhupengyang 已提交
695
        self.filter_size = [24, f_c, 3, 3]
696 697 698
        self.op_type = "depthwise_conv2d"


C
cnn 已提交
699
class TestCUDNNExhaustiveSearch(TestConv2DOp):
700 701 702
    def init_kernel_type(self):
        self.use_cudnn = True
        self.exhaustive_search = True
703
        self.dtype = np.float32 if core.is_compiled_with_rocm() else np.float64
704 705


C
cnn 已提交
706
class TestConv2DOpError(unittest.TestCase):
707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727
    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)


728 729
# Please Don't remove the following code.
# Currently, CI use cudnn V5.0 which not support dilation conv.
730
# class TestCUDNNWithDilation(TestWithDilation):
C
chengduoZH 已提交
731 732 733
#     def init_op_type(self):
#         self.op_type = "conv_cudnn"

L
liym27 已提交
734 735 736
# ---- test asymmetric padding ----


C
cnn 已提交
737
class TestConv2DOp_v2(OpTest):
L
liym27 已提交
738 739 740 741 742 743 744
    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
745
        self.dtype = np.float64
L
liym27 已提交
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
        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):
800
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
L
liym27 已提交
801
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
802 803
        self.check_output_with_place(
            place, atol=1e-5, check_dygraph=(self.use_mkldnn == False))
L
liym27 已提交
804 805

    def test_check_grad(self):
806
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
L
liym27 已提交
807 808 809 810
        if self.dtype == np.float16:
            return
        place = core.CUDAPlace(0) if self.has_cuda() else core.CPUPlace()
        self.check_grad_with_place(
811 812 813 814
            place, {'Input', 'Filter'},
            'Output',
            max_relative_error=0.02,
            check_dygraph=(self.use_mkldnn == False))
L
liym27 已提交
815 816

    def test_check_grad_no_filter(self):
817
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
L
liym27 已提交
818 819 820 821 822 823 824
        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,
825 826
            no_grad_set=set(['Filter']),
            check_dygraph=(self.use_mkldnn == False))
L
liym27 已提交
827 828

    def test_check_grad_no_input(self):
829
        # TODO(wangzhongpu): support mkldnn op in dygraph mode
L
liym27 已提交
830 831 832 833 834 835
        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',
836 837
            no_grad_set=set(['Input']),
            check_dygraph=(self.use_mkldnn == False))
L
liym27 已提交
838 839 840

    def init_test_case(self):
        self.pad = [0, 0]
841
        self.stride = [1, 2]
L
liym27 已提交
842 843 844
        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
845
        self.filter_size = [6, f_c, 4, 3]
L
liym27 已提交
846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866

    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


C
cnn 已提交
867
class TestConv2DOp_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
868 869 870 871 872
    def init_paddings(self):
        self.pad = [0, 0, 1, 2]
        self.padding_algorithm = "EXPLICIT"


C
cnn 已提交
873
class TestWithPad_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
874 875 876 877 878 879 880 881 882 883 884 885
    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"


C
cnn 已提交
886
class TestWithStride_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
887 888 889 890 891 892 893 894 895 896 897 898
    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"


C
cnn 已提交
899
class TestWithGroup_AsyPadding(TestConv2DOp_v2):
Z
zhupengyang 已提交
900 901 902 903 904 905 906 907
    def init_test_case(self):
        self.pad = [0, 0]
        self.stride = [1, 2]
        self.input_size = [2, 3, 5, 5]  # NCHW
        self.group = 3
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
        self.filter_size = [24, f_c, 4, 3]
L
liym27 已提交
908 909


C
cnn 已提交
910
class TestWith1x1_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
911 912 913 914 915
    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
Z
zhupengyang 已提交
916
        self.filter_size = [120, f_c, 1, 1]
L
liym27 已提交
917 918 919 920 921 922 923 924 925

    def init_group(self):
        self.groups = 3

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


C
cnn 已提交
926
class TestWithDepthWise3x3_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
927 928 929 930 931
    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
Z
zhupengyang 已提交
932
        self.filter_size = [16, f_c, 3, 3]
L
liym27 已提交
933 934 935 936 937 938 939 940 941 942 943 944

    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"


C
cnn 已提交
945
class TestWithDepthWise5x5_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960
    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"


C
cnn 已提交
961
class TestWithDepthWise7x7_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
962 963 964 965 966 967 968 969 970 971 972 973 974 975 976
    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"


C
cnn 已提交
977
class TestWithDilation_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
978 979 980 981 982
    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
Z
zhupengyang 已提交
983
        self.filter_size = [24, f_c, 3, 3]
L
liym27 已提交
984 985 986 987 988 989 990 991 992 993 994 995

    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"


C
cnn 已提交
996
class TestWithInput1x1Filter1x1_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
997 998
    def init_test_case(self):
        self.stride = [1, 1]
Z
zhupengyang 已提交
999
        self.input_size = [40, 3, 1, 1]  # NCHW
L
liym27 已提交
1000 1001
        assert np.mod(self.input_size[1], self.groups) == 0
        f_c = self.input_size[1] // self.groups
Z
zhupengyang 已提交
1002
        self.filter_size = [120, f_c, 1, 1]
L
liym27 已提交
1003 1004 1005 1006 1007 1008 1009 1010 1011

    def init_group(self):
        self.groups = 3

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


C
cnn 已提交
1012
create_test_cudnn_class(TestConv2DOp_AsyPadding)
L
liym27 已提交
1013 1014 1015 1016 1017 1018 1019
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)


C
cnn 已提交
1020
class TestDepthwiseConv_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1021 1022 1023 1024 1025 1026 1027
    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
Z
zhupengyang 已提交
1028
        self.filter_size = [12, f_c, 3, 3]
L
liym27 已提交
1029 1030 1031 1032 1033 1034 1035
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1036
class TestDepthwiseConv2_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1037 1038 1039 1040 1041 1042 1043
    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
Z
zhupengyang 已提交
1044
        self.filter_size = [12, f_c, 3, 3]
L
liym27 已提交
1045 1046 1047 1048 1049 1050 1051
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1052
class TestDepthwiseConv3_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1053 1054 1055 1056 1057 1058 1059
    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
Z
zhupengyang 已提交
1060
        self.filter_size = [24, f_c, 3, 3]
L
liym27 已提交
1061 1062 1063 1064 1065 1066 1067
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1068
class TestDepthwiseConvWithDilation_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1069 1070 1071 1072 1073 1074 1075 1076 1077
    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
Z
zhupengyang 已提交
1078
        self.filter_size = [24, f_c, 3, 3]
L
liym27 已提交
1079 1080 1081 1082 1083 1084 1085
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1086
class TestDepthwiseConvWithDilation2_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1087 1088 1089 1090 1091 1092 1093 1094 1095
    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
Z
zhupengyang 已提交
1096
        self.filter_size = [24, f_c, 3, 3]
L
liym27 已提交
1097 1098 1099 1100 1101 1102 1103
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1104
class TestDepthwiseConvandFuse_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1105 1106 1107 1108 1109 1110 1111 1112 1113
    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
Z
zhupengyang 已提交
1114
        self.filter_size = [12, f_c, 3, 3]
L
liym27 已提交
1115 1116 1117 1118 1119 1120 1121
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1122
class TestDepthwiseConv2andFuse_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1123 1124 1125 1126 1127 1128 1129 1130 1131
    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
Z
zhupengyang 已提交
1132
        self.filter_size = [12, f_c, 3, 3]
L
liym27 已提交
1133 1134 1135 1136 1137 1138 1139
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1140
class TestDepthwiseConv3andFuse_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1141 1142 1143 1144 1145 1146 1147 1148 1149
    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
Z
zhupengyang 已提交
1150
        self.filter_size = [24, f_c, 3, 3]
L
liym27 已提交
1151 1152 1153 1154 1155 1156 1157
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1158
class TestDepthwiseConvWithDilationandFuse_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1159 1160 1161 1162 1163 1164 1165 1166 1167 1168
    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
Z
zhupengyang 已提交
1169
        self.filter_size = [24, f_c, 3, 3]
L
liym27 已提交
1170 1171 1172 1173 1174 1175 1176
        self.op_type = "depthwise_conv2d"

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


C
cnn 已提交
1177
class TestDepthwiseConvWithDilation2andFuse_AsyPadding(TestConv2DOp_v2):
L
liym27 已提交
1178 1179 1180 1181 1182 1183 1184 1185 1186 1187
    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
Z
zhupengyang 已提交
1188
        self.filter_size = [24, f_c, 3, 3]
L
liym27 已提交
1189 1190 1191 1192 1193 1194 1195 1196
        self.op_type = "depthwise_conv2d"

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


#---------- test SAME VALID -----------
C
cnn 已提交
1197
create_test_padding_SAME_class(TestConv2DOp_AsyPadding)
L
liym27 已提交
1198 1199 1200 1201 1202
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)

C
cnn 已提交
1203
create_test_padding_VALID_class(TestConv2DOp_AsyPadding)
L
liym27 已提交
1204 1205 1206 1207 1208
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)

C
cnn 已提交
1209
create_test_cudnn_padding_SAME_class(TestConv2DOp_AsyPadding)
L
liym27 已提交
1210 1211 1212 1213 1214
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)

C
cnn 已提交
1215
create_test_cudnn_padding_VALID_class(TestConv2DOp_AsyPadding)
L
liym27 已提交
1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233
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 ---------
C
cnn 已提交
1234
create_test_channel_last_class(TestConv2DOp_AsyPadding)
L
liym27 已提交
1235 1236 1237 1238 1239 1240 1241 1242 1243 1244
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)

C
cnn 已提交
1245
create_test_cudnn_channel_last_class(TestConv2DOp_AsyPadding)
L
liym27 已提交
1246 1247 1248 1249 1250
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)

1251
create_test_cudnn_channel_last_fp16_class(
C
cnn 已提交
1252
    TestConv2DOp_AsyPadding, grad_check=False)
1253 1254 1255 1256 1257 1258 1259 1260 1261
create_test_cudnn_channel_last_fp16_class(
    TestWithPad_AsyPadding, grad_check=False)
create_test_cudnn_channel_last_fp16_class(
    TestWithStride_AsyPadding, grad_check=False)
create_test_cudnn_channel_last_fp16_class(
    TestWithGroup_AsyPadding, grad_check=False)
create_test_cudnn_channel_last_fp16_class(
    TestWithDilation_AsyPadding, grad_check=False)

L
liym27 已提交
1262 1263

# --------- test python API ---------------
C
cnn 已提交
1264
class TestConv2DAPI(unittest.TestCase):
L
liym27 已提交
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
    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")

1339 1340 1341 1342 1343 1344 1345 1346 1347 1348
    def test_depthwise_conv2d(self):
        x_var = paddle.uniform((2, 8, 8, 4), dtype='float32', min=-1., max=1.)
        conv = paddle.nn.Conv2D(
            in_channels=4,
            out_channels=4,
            kernel_size=(3, 3),
            groups=4,
            data_format='NHWC')
        y_var = conv(x_var)

L
liym27 已提交
1349

C
cnn 已提交
1350
class TestConv2DAPI_Error(unittest.TestCase):
L
liym27 已提交
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 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467
    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)


1468 1469 1470 1471 1472
# --------- test environment variable ------
@unittest.skipIf(
    not (core.is_compiled_with_cuda() or core.is_compiled_with_rocm()),
    "core is not compiled with CUDA or ROCM")
class TestConv2DEnviron(unittest.TestCase):
1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517
    def run1(self, place):
        with fluid.program_guard(fluid.Program(), fluid.Program()):
            inputs = fluid.layers.data(
                shape=[2, 3, 5, 5],
                append_batch_size=False,
                name="inputs",
                dtype="float32")
            result = fluid.layers.conv2d(
                input=inputs,
                num_filters=4,
                filter_size=[3, 3],
                stride=[1, 1],
                padding=0,
                dilation=[1, 1],
                groups=1,
                data_format="NCHW")
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            fetches = exe.run(fluid.default_main_program(),
                              feed={"inputs": self.input_np},
                              fetch_list=[result])

    def run2(self, place):
        with fluid.dygraph.guard(place):
            inputs = fluid.dygraph.to_variable(self.input_np)
            conv = paddle.nn.Conv2D(
                in_channels=3,
                out_channels=4,
                kernel_size=(3, 3),
                data_format="NCHW")
            result = conv(inputs)

    def run3(self, place):
        with fluid.dygraph.guard(place):
            inputs = fluid.dygraph.to_variable(self.input_np)
            conv = paddle.fluid.dygraph.nn.Conv2D(
                num_channels=3,
                num_filters=4,
                filter_size=(3, 3), )
            result = conv(inputs)

    def run_all(self, place):
        self.run1(place)
        self.run2(place)
        self.run3(place)
1518 1519

    def test_environ(self):
1520 1521 1522 1523 1524 1525
        self.input_np = np.random.random([2, 3, 5, 5]).astype("float32")
        for place in [paddle.CPUPlace(), paddle.CUDAPlace(0)]:
            fluid.set_flags({'FLAGS_conv2d_disable_cudnn': False})
            self.run_all(place)
            fluid.set_flags({'FLAGS_conv2d_disable_cudnn': True})
            self.run_all(place)
1526 1527


1528 1529
if __name__ == '__main__':
    unittest.main()