op_test.py 66.2 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

B
baojun 已提交
17
import os
18
import unittest
19
import warnings
20
import numpy as np
21
import random
M
minqiyang 已提交
22
import six
23
import struct
24
import time
25
import itertools
Y
Yu Yang 已提交
26
import collections
M
minqiyang 已提交
27
from collections import defaultdict
28

29
import paddle
30 31
import paddle.fluid as fluid
import paddle.fluid.core as core
32 33 34
from paddle.fluid.backward import append_backward
from paddle.fluid.op import Operator
from paddle.fluid.executor import Executor
35
from paddle.fluid.framework import Program, OpProtoHolder, Variable
36
from testsuite import create_op, set_input, append_input_output, append_loss_ops
37
from paddle.fluid import unique_name
38
from white_list import op_accuracy_white_list, check_shape_white_list, compile_vs_runtime_white_list, no_check_set_white_list
39
from white_list import op_threshold_white_list, no_grad_set_white_list
40 41


42 43 44 45 46 47 48 49
def _set_use_system_allocator(value=None):
    USE_SYSTEM_ALLOCATOR_FLAG = "FLAGS_use_system_allocator"
    old_value = core.globals()[USE_SYSTEM_ALLOCATOR_FLAG]
    value = old_value if value is None else value
    core.globals()[USE_SYSTEM_ALLOCATOR_FLAG] = value
    return old_value


50 51 52 53
def randomize_probability(batch_size, class_num, dtype='float32'):
    prob = np.random.uniform(
        0.1, 1.0, size=(batch_size, class_num)).astype(dtype)
    prob_sum = prob.sum(axis=1)
M
minqiyang 已提交
54
    for i in six.moves.xrange(len(prob)):
55 56 57 58
        prob[i] /= prob_sum[i]
    return prob


59 60
def get_numeric_gradient(place,
                         scope,
61 62 63
                         op,
                         inputs,
                         input_to_check,
Y
Yancey 已提交
64
                         output_names,
65
                         delta=0.005,
C
chengduo 已提交
66
                         in_place=False):
Y
Yu Yang 已提交
67
    # FIXME: change this method by compile time concepts
68
    set_input(scope, op, inputs, place)
69 70

    def product(dim):
M
minqiyang 已提交
71
        return six.moves.reduce(lambda a, b: a * b, dim, 1)
72 73

    tensor_to_check = scope.find_var(input_to_check).get_tensor()
Y
yuyang18 已提交
74 75
    tensor_size = product(tensor_to_check.shape())
    tensor_to_check_dtype = tensor_to_check._dtype()
76
    if tensor_to_check_dtype == core.VarDesc.VarType.FP32:
77
        tensor_to_check_dtype = np.float32
78
    elif tensor_to_check_dtype == core.VarDesc.VarType.FP64:
79
        tensor_to_check_dtype = np.float64
D
dzhwinter 已提交
80 81 82 83
    elif tensor_to_check_dtype == core.VarDesc.VarType.FP16:
        tensor_to_check_dtype = np.float16
        # set delta as np.float16, will automatic convert to float32, float64
        delta = np.array(delta).astype(np.float16)
84 85 86 87
    else:
        raise ValueError("Not supported data type " + str(
            tensor_to_check_dtype))

C
chengduo 已提交
88 89 90 91 92 93 94 95 96
    def get_output():
        sum = []
        op.run(scope, place)
        for output_name in output_names:
            sum.append(
                np.array(scope.find_var(output_name).get_tensor()).astype(
                    tensor_to_check_dtype).mean())
        return tensor_to_check_dtype(np.array(sum).sum() / len(output_names))

97 98 99
    gradient_flat = np.zeros(shape=(tensor_size, ), dtype=tensor_to_check_dtype)

    def __get_elem__(tensor, i):
D
dzhwinter 已提交
100 101 102 103 104
        if tensor_to_check_dtype == np.float16:
            numpy_tensor = np.array(tensor).astype(np.float16)
            numpy_tensor = numpy_tensor.flatten()
            return numpy_tensor[i]
        elif tensor_to_check_dtype == np.float32:
Y
yuyang18 已提交
105
            return tensor._get_float_element(i)
106
        else:
Y
yuyang18 已提交
107
            return tensor._get_double_element(i)
108 109

    def __set_elem__(tensor, i, e):
D
dzhwinter 已提交
110 111 112 113 114
        if tensor_to_check_dtype == np.float16:
            numpy_tensor = np.array(tensor).astype(np.float16)
            shape = numpy_tensor.shape
            numpy_tensor = numpy_tensor.flatten()
            numpy_tensor[i] = e
115
            numpy_tensor = numpy_tensor.reshape(shape)
D
dzhwinter 已提交
116 117
            tensor.set(numpy_tensor, place)
        elif tensor_to_check_dtype == np.float32:
Y
yuyang18 已提交
118
            tensor._set_float_element(i, e)
119
        else:
Y
yuyang18 已提交
120
            tensor._set_double_element(i, e)
121

122 123
    # we only compute gradient of one element each time.
    # we use a for loop to compute the gradient of every element.
M
minqiyang 已提交
124
    for i in six.moves.xrange(tensor_size):
125
        if in_place:
126
            set_input(scope, op, inputs, place)
127 128

        # get one input element throw it's index i.
129
        origin = __get_elem__(tensor_to_check, i)
130 131
        # add delta to it, run op and then get the sum of the result tensor.
        x_pos = origin + delta
132
        __set_elem__(tensor_to_check, i, x_pos)
133 134 135
        y_pos = get_output()

        if in_place:
136
            set_input(scope, op, inputs, place)
137 138

        x_neg = origin - delta
139
        __set_elem__(tensor_to_check, i, x_neg)
140 141
        y_neg = get_output()

142
        __set_elem__(tensor_to_check, i, origin)
143 144
        gradient_flat[i] = (y_pos - y_neg) / delta / 2

Y
yuyang18 已提交
145
    return gradient_flat.reshape(tensor_to_check.shape())
146 147


148 149
def skip_check_grad_ci(reason=None):
    """Decorator to skip check_grad CI.
C
cc 已提交
150

151
       Check_grad is required for Op test cases. However, there are some special
C
cc 已提交
152
       cases that do not need to do check_grad. This decorator is used to skip the
153
       check_grad of the above cases.
C
cc 已提交
154 155

       Note: the execution of unit test will not be skipped. It just avoids check_grad
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
       checking in tearDownClass method by setting a `no_need_check_grad` flag.

       Example:
           @skip_check_grad_ci(reason="For inference, check_grad is not required.")
           class TestInference(OpTest):
    """
    if not isinstance(reason, str):
        raise AssertionError("The reason for skipping check_grad is required.")

    def wrapper(cls):
        cls.no_need_check_grad = True
        return cls

    return wrapper


172 173 174 175 176 177 178 179 180 181 182 183
def copy_bits_from_float_to_uint16(f):
    return struct.unpack('<I', struct.pack('<f', f))[0] >> 16


def convert_float_to_uint16(float_list):
    new_output = []
    for x in np.nditer(float_list):
        new_output.append(np.uint16(copy_bits_from_float_to_uint16(x)))

    return np.reshape(new_output, float_list.shape).view(np.uint16)


184
class OpTest(unittest.TestCase):
185 186 187 188 189
    @classmethod
    def setUpClass(cls):
        '''Fix random seeds to remove randomness from tests'''
        cls._np_rand_state = np.random.get_state()
        cls._py_rand_state = random.getstate()
190
        cls.call_once = False
191
        cls.dtype = None
192
        cls.outputs = {}
193
        cls.input_shape_is_large = True
194 195 196 197

        np.random.seed(123)
        random.seed(124)

198 199
        cls._use_system_allocator = _set_use_system_allocator(True)

200 201
    @classmethod
    def tearDownClass(cls):
Y
yuyang18 已提交
202
        """Restore random seeds"""
203 204 205
        np.random.set_state(cls._np_rand_state)
        random.setstate(cls._py_rand_state)

206 207
        _set_use_system_allocator(cls._use_system_allocator)

208 209 210 211
        def is_empty_grad_op(op_type):
            all_op_kernels = core._get_all_register_op_kernels()
            grad_op = op_type + '_grad'
            if grad_op in all_op_kernels.keys():
J
juncaipeng 已提交
212
                if is_mkldnn_op_test():
213 214 215 216 217 218 219 220
                    grad_op_kernels = all_op_kernels[grad_op]
                    for grad_op_kernel in grad_op_kernels:
                        if 'MKLDNN' in grad_op_kernel:
                            return False
                else:
                    return False
            return True

221 222 223
        def is_xpu_op_test():
            return hasattr(cls, "use_xpu") and cls.use_xpu == True

J
juncaipeng 已提交
224
        def is_mkldnn_op_test():
225
            return hasattr(cls, "use_mkldnn") and cls.use_mkldnn == True
J
juncaipeng 已提交
226

227 228
        if not hasattr(cls, "op_type"):
            raise AssertionError(
229 230
                "This test do not have op_type in class attrs, "
                "please set self.__class__.op_type=the_real_op_type manually.")
231

J
juncaipeng 已提交
232 233
        # case in NO_FP64_CHECK_GRAD_CASES and op in NO_FP64_CHECK_GRAD_OP_LIST should be fixed
        if not hasattr(cls, "no_need_check_grad") \
234
            and not is_empty_grad_op(cls.op_type):
J
juncaipeng 已提交
235
            if cls.dtype is None or \
236 237
                (cls.dtype == np.float16 \
                    and cls.op_type not in op_accuracy_white_list.NO_FP16_CHECK_GRAD_OP_LIST \
J
juncaipeng 已提交
238 239 240 241
                    and not hasattr(cls, "exist_check_grad")):
                raise AssertionError("This test of %s op needs check_grad." %
                                     cls.op_type)

242
            # check for op test with fp64 precision, but not check mkldnn op test for now
J
juncaipeng 已提交
243 244
            if cls.dtype in [np.float32, np.float64] \
                and cls.op_type not in op_accuracy_white_list.NO_FP64_CHECK_GRAD_OP_LIST \
245
                and not hasattr(cls, 'exist_fp64_check_grad') \
246
                and not is_xpu_op_test() \
J
juncaipeng 已提交
247
                and not is_mkldnn_op_test():
J
juncaipeng 已提交
248 249 250 251
                raise AssertionError(
                    "This test of %s op needs check_grad with fp64 precision." %
                    cls.op_type)

252
            if not cls.input_shape_is_large \
253 254 255 256
                and cls.op_type not in check_shape_white_list.NEED_TO_FIX_OP_LIST:
                raise AssertionError(
                    "Input's shape should be large than or equal to 100 for " +
                    cls.op_type + " Op.")
257

258 259 260 261 262
    def try_call_once(self, data_type):
        if not self.call_once:
            self.call_once = True
            self.dtype = data_type

263 264 265 266 267
    def is_bfloat16_op(self):
        return self.dtype == np.uint16 or (
            hasattr(self, 'mkldnn_data_type') and
            getattr(self, 'mkldnn_data_type') is "bfloat16")

268
    def infer_dtype_from_inputs_outputs(self, inputs, outputs):
J
juncaipeng 已提交
269 270 271 272
        def is_np_data(input):
            return isinstance(input, (np.ndarray, np.generic))

        def infer_dtype(numpy_dict, dtype_set):
273 274 275
            assert isinstance(
                numpy_dict,
                dict), "self.inputs, self.outputs must be numpy_dict"
J
juncaipeng 已提交
276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
            # the inputs are as follows:
            # case 1: inputs = {'X': x}
            # case 2: inputs = {'X': (x, x_lod)}
            # case 3: inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]}
            # case 4: inputs = {'X': [("x1", (x1, [x1_lod1])), ("x2", (x2, [x2_.lod2]))]}
            # TODO(juncaipeng) infer dtype from inputs maybe obtain wrong type.
            for _, var_value in six.iteritems(numpy_dict):
                if is_np_data(var_value):  # case 1
                    dtype_set.add(var_value.dtype)
                elif isinstance(var_value, (list, tuple)):  # case 2, 3, 4
                    for sub_val_value in var_value:
                        if is_np_data(sub_val_value):  # case 2
                            dtype_set.add(sub_val_value.dtype)
                        elif len(sub_val_value) > 1 and is_np_data(
                                sub_val_value[1]):  # case 3
                            dtype_set.add(sub_val_value[1].dtype)
                        elif len(sub_val_value) > 1 and isinstance(sub_val_value[1], (list, tuple)) \
                            and is_np_data(sub_val_value[1][0]): # case 4
                            dtype_set.add(sub_val_value[1][0].dtype)

        # infer dtype from inputs, and dtype means the precision of the test
        # collect dtype of all inputs
        dtype_set = set()
        infer_dtype(inputs, dtype_set)
        dtype_list = [
            np.dtype(np.float64), np.dtype(np.float32), np.dtype(np.float16),
302 303 304
            np.dtype(np.int64), np.dtype(np.int32), np.dtype(np.uint16),
            np.dtype(np.int16), np.dtype(np.int8), np.dtype(np.uint8),
            np.dtype(np.bool)
J
juncaipeng 已提交
305 306 307 308 309 310
        ]
        # check the dtype in dtype_list in order, select the first dtype that in dtype_set
        for dtype in dtype_list:
            if dtype in dtype_set:
                self.dtype = dtype
                break
311 312
        # save dtype in class attr
        self.__class__.dtype = self.dtype
313

Y
Yang Yang(Tony) 已提交
314 315 316 317 318 319
    def feed_var(self, input_vars, place):
        feed_map = {}
        for var_name in input_vars:
            if isinstance(input_vars[var_name], list):
                for name, np_value in self.inputs[var_name]:
                    tensor = core.LoDTensor()
320
                    if isinstance(np_value, tuple):
321
                        tensor.set(np_value[0], place)
322
                        tensor.set_recursive_sequence_lengths(np_value[1])
323
                    else:
324
                        tensor.set(np_value, place)
Y
Yang Yang(Tony) 已提交
325 326 327 328
                    feed_map[name] = tensor
            else:
                tensor = core.LoDTensor()
                if isinstance(self.inputs[var_name], tuple):
329
                    tensor.set(self.inputs[var_name][0], place)
330 331
                    tensor.set_recursive_sequence_lengths(self.inputs[var_name][
                        1])
Y
Yang Yang(Tony) 已提交
332
                else:
333
                    tensor.set(self.inputs[var_name], place)
Y
Yang Yang(Tony) 已提交
334 335 336
                feed_map[var_name] = tensor
        return feed_map

337
    def _append_ops(self, block):
J
juncaipeng 已提交
338
        self.__class__.op_type = self.op_type  # for ci check, please not delete it for now
339 340 341 342
        if (hasattr(self, "use_mkldnn") and self.use_mkldnn == True) or \
            (hasattr(self, "attrs") and "use_mkldnn" in self.attrs and \
                    self.attrs["use_mkldnn"] == True):
            self.__class__.use_mkldnn = True
C
cc 已提交
343

344 345 346 347 348
        if (hasattr(self, "use_xpu") and self.use_xpu == True) or \
            (hasattr(self, "attrs") and "use_xpu" in self.attrs and \
                    self.attrs["use_xpu"] == True):
            self.__class__.use_xpu = True

Y
Yang Yang(Tony) 已提交
349
        op_proto = OpProtoHolder.instance().get_op_proto(self.op_type)
350 351 352 353 354 355
        "infer datatype from inputs and outputs for this test case"
        self.infer_dtype_from_inputs_outputs(self.inputs, self.outputs)
        inputs = append_input_output(block, op_proto, self.inputs, True,
                                     self.dtype)
        outputs = append_input_output(block, op_proto, self.outputs, False,
                                      self.dtype)
P
phlrain 已提交
356 357 358 359 360 361 362 363 364

        if hasattr(self, "cache_name_list"):
            for name in self.cache_name_list:
                inputs[name] = block.create_var(
                    name=name,
                    persistable=True,
                    type=core.VarDesc.VarType.RAW,
                    stop_gradient=True)

Y
Yang Yang(Tony) 已提交
365 366 367 368 369
        op = block.append_op(
            type=self.op_type,
            inputs=inputs,
            outputs=outputs,
            attrs=self.attrs if hasattr(self, "attrs") else dict())
C
cc 已提交
370
        # infer variable type and infer shape in compile-time
Q
QI JUN 已提交
371 372
        op.desc.infer_var_type(block.desc)
        op.desc.infer_shape(block.desc)
Y
Yang Yang(Tony) 已提交
373

374 375
        return op

376 377
    def _get_io_vars(self, block, numpy_inputs):
        inputs = {}
M
minqiyang 已提交
378
        for name, value in six.iteritems(numpy_inputs):
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
            if isinstance(value, list):
                var_list = [
                    block.var(sub_name) for sub_name, sub_value in value
                ]
                inputs[name] = var_list
            else:
                inputs[name] = block.var(name)
        return inputs

    def _get_inputs(self, block):
        return self._get_io_vars(block, self.inputs)

    def _get_outputs(self, block):
        return self._get_io_vars(block, self.outputs)

    def calc_output(self, place):
        outs, _ = self._calc_output(place)
        return outs

M
minqiyang 已提交
398 399 400 401
    def _create_var_from_numpy(self, value):
        if isinstance(value, tuple):
            data = value[0]
            lod = value[1]
L
lujun 已提交
402
            v = fluid.dygraph.base.to_variable(value=data)
403
            v.value().get_tensor().set_recursive_sequence_lengths(lod)
M
minqiyang 已提交
404 405
            return v
        else:
L
lujun 已提交
406
            return fluid.dygraph.base.to_variable(value)
M
minqiyang 已提交
407

408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425
    def get_sequence_batch_size_1_input(self, lod=None, shape=None):
        """Get LoD input data whose batch size is 1.
        All sequence related OP unittests should call this function to contain the case of batch size = 1.
        Args:
            lod (list[list of int], optional): Length-based LoD, length of lod[0] should be 1. Default: [[13]].
            shape (list, optional): Shape of input, shape[0] should be equals to lod[0][0]. Default: [13, 23].
        Returns:
            tuple (ndarray, lod) : LoD input data whose batch size is 1.
        """
        if lod is None:
            lod = [[13]]
        if shape is None:
            shape = [13, 23]
        assert len(lod[0]) == 1
        assert lod[0][0] == shape[0]
        x = np.random.uniform(0.1, 1, shape).astype('float32')
        return (x, lod)

426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
    def lod_has_single_zero(self, lod):
        for i in range(len(lod) - 2):
            if lod[i] != 0 and lod[i + 1] == 0 and lod[i + 2] != 0:
                return True
        return False

    def lod_has_continuous_zero(self, lod):
        for i in range(len(lod) - 3):
            if lod[i] != 0 and lod[i + 1] == 0 and lod[i + 2] == 0 and lod[
                    i + 3] != 0:
                return True
        return False

    def get_sequence_instance_size_0_input(self, lod=None, shape=None):
        """Get LoD input data whose instance size is 0.
        All sequence related OP unittests should call this function to contain the case of instance size is 0.
        Args:
            lod (list[list of int], optional): Length-based LoD, lod[0]'s size must at least eight, lod[0] must at least two zeros at the beginning and at least two zeros at the end, the middle position of lod[0] contains a single zero and multiple zero. Default: [[0, 0, 4, 0, 3, 0, 0, 5, 0, 0]].
            shape (list, optional): Shape of input, shape[0] should be equals to lod[0][0]. Default: [13, 23].
        Returns:
            tuple (ndarray, lod): LoD input data whose instance size is 0.
        """
        if lod is None:
            lod = [[0, 0, 4, 0, 3, 0, 0, 5, 0, 0]]
        if shape is None:
            shape = [12, 10]
        assert len(lod[0]) >= 8
        assert lod[0][0] == 0 and lod[0][1] == 0 and lod[0][-1] == 0 and lod[0][
            -2] == 0
        assert self.lod_has_single_zero(lod[0]) is True
        assert self.lod_has_continuous_zero(lod[0]) is True
        assert sum(lod[0]) == shape[0]

        x = np.random.uniform(0.1, 1, shape).astype('float32')
        return (x, lod)

462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477
    def append_input_output_for_dygraph(self, op_proto, np_list, is_input,
                                        if_return_inputs_grad_dict, block):
        def create_var(np_value, name, is_input, if_return_inputs_grad_dict):
            np_value_temp = np_value
            has_lod = False
            lod_temp = None
            if isinstance(np_value, tuple):
                np_value_temp = np_value[0]
                has_lod = True
                lod_temp = np_value[1]

            if is_input:
                v = self._create_var_from_numpy(np_value_temp)
                if if_return_inputs_grad_dict:
                    v.stop_gradient = False
                if has_lod:
478
                    v.value().get_tensor().set_recursive_sequence_lengths(
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
                        lod_temp)
            else:
                v = block.create_var(
                    name=name,
                    dtype=np_value_temp.dtype,
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    persistable=False,
                    stop_gradient=False)

            return v

        # prepare variable for input or output
        var_dict = defaultdict(list)
        if if_return_inputs_grad_dict:
            inputs_grad_dict = defaultdict()
        proto_list = op_proto.inputs if is_input else op_proto.outputs
        for var_proto in proto_list:
            name = var_proto.name
            if (name not in np_list) and var_proto.dispensable:
                continue
            if name not in np_list:
                assert var_proto.intermediate, "{} not found".format(name)
                v = block.create_var(
                    dtype='float32', type=core.VarDesc.VarType.LOD_TENSOR)
                var_dict[name].append(v)
                if if_return_inputs_grad_dict:
                    inputs_grad_dict[name] = v
                continue
            if var_proto.duplicable:
                assert isinstance(
                    np_list[name],
                    list), "Duplicable {} should be set as list".format(name)
                var_list = []
                slot_name = name
                for (name, np_value) in np_list[name]:
                    v = create_var(np_value, name, is_input,
                                   if_return_inputs_grad_dict)
                    var_list.append(v)
                    if if_return_inputs_grad_dict:
                        inputs_grad_dict[name] = v
                var_dict[slot_name] = var_list
            else:
                nplist_value_temp = None
                name_temp = None
                if isinstance(np_list[name], list):
                    nplist_value_temp = np_list[name][0]
                    name_temp = name
                else:
                    nplist_value_temp = np_list[name]
                    name_temp = unique_name.generate("%s_out" % (name))
                v = create_var(nplist_value_temp, name_temp, is_input,
                               if_return_inputs_grad_dict)
                var_dict[name].append(v)
                if if_return_inputs_grad_dict:
                    inputs_grad_dict[name] = v

        if if_return_inputs_grad_dict:
            return var_dict, inputs_grad_dict
        else:
            return var_dict

L
lujun 已提交
540
    def _calc_dygraph_output(self, place, parallel=False, no_check_set=None):
J
juncaipeng 已提交
541
        self.__class__.op_type = self.op_type  # for ci check, please not delete it for now
L
lujun 已提交
542
        with fluid.dygraph.base.guard(place=place):
M
minqiyang 已提交
543 544
            block = fluid.default_main_program().global_block()

545
            op_proto = OpProtoHolder.instance().get_op_proto(self.op_type)
M
minqiyang 已提交
546

547 548 549
            # prepare input variable
            inputs = self.append_input_output_for_dygraph(op_proto, self.inputs,
                                                          True, False, block)
M
minqiyang 已提交
550 551

            # prepare output variable
552 553 554 555 556 557 558 559 560
            outputs = self.append_input_output_for_dygraph(
                op_proto, self.outputs, False, False, block)

            # prepare attrbutes
            attrs_outputs = {}
            if hasattr(self, "attrs"):
                for attrs_name in self.attrs:
                    if self.attrs[attrs_name] is not None:
                        attrs_outputs[attrs_name] = self.attrs[attrs_name]
M
minqiyang 已提交
561 562 563 564
            block.append_op(
                type=self.op_type,
                inputs=inputs,
                outputs=outputs,
565
                attrs=attrs_outputs if hasattr(self, "attrs") else None)
M
minqiyang 已提交
566
            return outputs
567

568 569 570 571 572 573
    def _calc_output(self,
                     place,
                     parallel=False,
                     no_check_set=None,
                     loss=None,
                     enable_inplace=None,
574
                     for_inplace_test=None):
575 576
        program = Program()
        block = program.global_block()
577
        op = self._append_ops(block)
578 579 580 581 582

        inputs = self._get_inputs(block)
        outputs = self._get_outputs(block)
        feed_map = self.feed_var(inputs, place)

583
        if for_inplace_test:
C
cc 已提交
584 585
            # Some variables' tensors hold no buffer (tensor's _holder is NULL), like XShape in reshape2 op,
            # and the shapes of those variables contain 0 (eg. Xshape.shape = [0, 2, 5]).
586 587
            # Set persistable for those variables in order to get them from global_scope for inplace grad test directly other than feed them,
            # since feed op calls check_memory_size() which fails when tensor's holder_ is NULL.
588 589
            for out_name in op.output_arg_names:
                var = block.var(out_name)
590 591
                if 0 in var.shape:
                    var.persistable = True
592
        original_program = program
593 594
        if parallel:
            use_cuda = False
595
            if isinstance(place, fluid.CUDAPlace):
596
                use_cuda = True
597 598 599
            compiled_prog = fluid.CompiledProgram(program).with_data_parallel(
                loss_name=loss.name if loss else None, places=place)
            program = compiled_prog
600 601 602 603
        fetch_list = getattr(self, "fetch_list", [])
        # if the fetch_list is customized by user, we use it directly.
        # if not, fill the fetch_list by the user configured outputs in test.
        if len(fetch_list) == 0:
M
minqiyang 已提交
604
            for var_name, var in six.iteritems(outputs):
605 606
                if no_check_set is not None and var_name in no_check_set:
                    continue
Y
Yang Yang(Tony) 已提交
607 608
                if isinstance(var, list):
                    for v in var:
609
                        fetch_list.append(v.name)
Y
Yang Yang(Tony) 已提交
610
                else:
611
                    fetch_list.append(var.name)
612 613 614 615
        # if the fetch_list still empty, fill the fetch_list by the operator output.
        if len(fetch_list) == 0:
            for out_name, out_dup in Operator.get_op_outputs(self.op_type):
                fetch_list.append(str(out_name))
616 617 618 619 620 621 622 623 624

        if enable_inplace is not None:
            build_strategy = fluid.BuildStrategy()
            build_strategy.enable_inplace = enable_inplace

            compiled_prog = fluid.CompiledProgram(program).with_data_parallel(
                build_strategy=build_strategy, places=place)
            program = compiled_prog

625
        executor = Executor(place)
626 627 628 629
        outs = executor.run(program,
                            feed=feed_map,
                            fetch_list=fetch_list,
                            return_numpy=False)
630 631
        self.op = op
        self.program = original_program
632 633 634 635
        if for_inplace_test:
            return outs, fetch_list, feed_map, original_program, op.desc
        else:
            return outs, fetch_list
Y
Yang Yang(Tony) 已提交
636

637 638 639 640 641 642 643 644 645
    def _compare_expect_and_actual_outputs(self,
                                           place,
                                           fetch_list,
                                           expect_outs,
                                           actual_outs,
                                           inplace_atol=None):
        """Compare expect outs and actual outs of an tested op.

        Args:
C
cc 已提交
646
            place (CPUPlace | CUDAPlace): The place where the op runs.
647 648 649 650 651 652 653 654 655 656
            fetch_list (list): The outputs of tested op.
            expect_outs (list): The expect outs of tested op.
            actual_outs (list): The actual outs of tested op.
            inplace_atol (float): The tolerable error, only set when tested op doesn't ensure computational consistency, like group_norm op.

        Returns:
            None.
        """
        # compare expect_outs and actual_outs
        for i, name in enumerate(fetch_list):
C
cc 已提交
657
            # Note(zhiqiu): inplace_atol should be only set when op doesn't ensure
L
Leo Chen 已提交
658 659 660
            # computational consistency.
            # When inplace_atol is not None, the inplace check uses numpy.allclose
            # to check inplace result instead of numpy.array_equal.
661 662
            expect_out = np.array(expect_outs[i])
            actual_out = np.array(actual_outs[i])
663 664 665
            if inplace_atol is not None:
                self.assertTrue(
                    np.allclose(
666
                        expect_out, actual_out, atol=inplace_atol),
667 668
                    "Output (" + name + ") has diff at " + str(place) +
                    " when using and not using inplace" + "\nExpect " +
669 670
                    str(expect_out) + "\n" + "But Got" + str(actual_out) +
                    " in class " + self.__class__.__name__)
671 672
            else:
                self.assertTrue(
673
                    np.array_equal(expect_out, actual_out),
674 675
                    "Output (" + name + ") has diff at " + str(place) +
                    " when using and not using inplace" + "\nExpect " +
676 677
                    str(expect_out) + "\n" + "But Got" + str(actual_out) +
                    " in class " + self.__class__.__name__ + '\n')
678 679 680 681 682 683 684 685

    def _construct_grad_program_from_forward(self, fwd_program, grad_op_desc,
                                             op_grad_to_var):
        """Generate grad_program which contains the grad_op.

        Args:
            fwd_program (tuple): The program that contains grad_op_desc's corresponding forward op.
            grad_op_desc (OpDesc): The OpDesc of grad op.
C
cc 已提交
686
            op_grad_to_var (dict): The relation of variables in grad op and its forward op.
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

        Returns:
            grad_program (program): The program which contains the grad_op.
        """
        grad_program = Program()
        grad_block = grad_program.global_block()
        new_op_desc = grad_block.desc.append_op()
        new_op_desc.copy_from(grad_op_desc)
        grad_program._sync_with_cpp()

        # Create grad vars based on fwd vars (shape and dtype)
        for arg in grad_op_desc.input_arg_names(
        ) + grad_op_desc.output_arg_names():
            fwd_var_name = op_grad_to_var.get(arg, None)
            if fwd_var_name is None:
                fwd_var_name = arg
            fwd_var = fwd_program.global_block().vars.get(fwd_var_name)
            assert fwd_var is not None, "{} cannot be found".format(
                fwd_var_name)
            grad_var = grad_block.create_var(
                name=arg,
                dtype=fwd_var.dtype,
                shape=fwd_var.shape,
                type=fwd_var.type,
                persistable=False)

C
cc 已提交
713 714
            # Some variables' tensors hold no buffer (tensor's _holder is NULL), like XShape in reshape2 op,
            # and the shapes of those variables contain 0 (eg. Xshape.shape = [0, 2, 5]).
715 716 717 718 719 720 721 722 723 724 725 726 727 728 729
            # Set persistable for those variables in order to get them from global_scope for inplace grad test directly other than feed them,
            # since feed op calls check_memory_size() which fails when tensor's holder_ is NULL.
            if 0 in grad_var.shape:
                grad_var.persistable = True
        grad_program._sync_with_cpp()
        return grad_program

    def _construct_grad_feed_map_from_forward(self, place, fwd_res,
                                              grad_op_desc, op_grad_to_var):
        """Generate grad_feed_map for grad_program.

        since we don`t really check gradient accuracy, but check the consistency when using and not using inplace,
        we use fwd outs (also inputs sometimes) to construct grad inputs.

        Args:
C
cc 已提交
730
            place (CPUPlace | CUDAPlace): The place where the op runs.
731 732 733
            fwd_res (tuple): The outputs of its forward op, in the same form as returns of _calc_outputs() when for_inplace_test is True.
                i.e., tuple(fwd_outs, fwd_fetch_list, fwd_feed_map, fwd_program, fwd_op_desc)
            grad_op_desc (OpDesc): The OpDesc of grad op.
C
cc 已提交
734
            op_grad_to_var (dict): The relation of variables in grad op and its fwd_op.
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

        Returns:
            grad_feed_map (dict): The feed_map of grad_op.
        """
        fwd_outs, fwd_fetch_list, fwd_feed_map, fwd_program, fwd_op_desc = fwd_res
        p = core.Place()
        p.set_place(place)
        grad_feed_map = {}
        for arg in grad_op_desc.input_arg_names():
            if arg in fwd_feed_map.keys():
                grad_feed_map[arg] = fwd_feed_map[arg]._copy(p)
            else:
                fwd_var_name = op_grad_to_var.get(arg, None)
                if fwd_var_name is None:
                    fwd_var_name = arg

                for i, out_name in enumerate(fwd_fetch_list):
                    if out_name == fwd_var_name:
                        # don't feed variables whose tensors hold no buffer (shape contains 0 like shape = [0,2,5] and holder_ is NULL), like XShape in reshape2 op.
                        # get them from global_scope directly since we have set them persistable in fwd execution
                        if 0 in fwd_program.global_block().var(out_name).shape:
                            continue
                        else:
                            grad_feed_map[arg] = fwd_outs[i]._copy(p)
        return grad_feed_map

    def _get_need_run_ops(self, op_desc, fwd_op_desc=None):
        """Postorder traversal of the 'grad' tree to get all ops that need to run during inplace test.
        An op needs to run druing inplace check if,
        (1) it has infer_inplace,
        (2) it has infer_inplace in its grad descendants. (since we need its outputs as to construct its grad's inputs)
C
cc 已提交
766

767
        Args:
C
cc 已提交
768 769
            op_desc (OpDesc): The op_desc of current op.
            fwd_op_desc (OpDesc): The op_desc of current op's forward op, None if current op has no forward op.
770
                Eg. relu's fwd_op is None, relu_grad's fwd_op is relu, relu_grad_grad's fwd_op is relu_grad, etc.
C
cc 已提交
771

772 773 774 775 776 777 778 779 780 781 782 783 784 785
        Returns:
            need_run_ops (list[(op_desc, fwd_op_desc)]): The ops that need to run during inplace test.
        """
        need_run_ops = []
        visited_ops = []

        def _dfs_grad_op(op_desc, fwd_op_desc=None):
            visited_ops.append(op_desc.type())
            has_infer_inplace = fluid.core.has_infer_inplace(op_desc.type())
            has_grad_op_maker = fluid.core.has_grad_op_maker(op_desc.type())
            has_infer_inplace_in_grad_descendants = False
            if not has_grad_op_maker:
                has_infer_inplace_in_descendants = False
            else:
C
cc 已提交
786
                # get grad_op_desc
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
                grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc(
                    op_desc, set(), [])
                if not grad_op_desc_list:
                    has_infer_inplace_in_grad_descendants = False
                else:
                    for i, grad_op_desc in enumerate(grad_op_desc_list):
                        if grad_op_desc.type(
                        ) not in visited_ops and _dfs_grad_op(
                                grad_op_desc, fwd_op_desc=op_desc):
                            has_infer_inplace_in_grad_descendants = True
            if has_infer_inplace or has_infer_inplace_in_grad_descendants:
                need_run_ops.append((op_desc, fwd_op_desc))
                return True
            else:
                return False

        _dfs_grad_op(op_desc, fwd_op_desc=fwd_op_desc)
        return need_run_ops

    def _check_forward_inplace(self,
                               place,
                               no_check_set=None,
                               inplace_atol=None):
        """Chech the inplace correctness of given op (self.op_type).
        Run the op twice with same inputs, one enable inplace and another disable, compare their outputs.
C
cc 已提交
812

813
        Args:
C
cc 已提交
814
            place (CPUPlace | CUDAPlace): The place where the op runs.
815 816 817 818
            no_check_set (list): The names of outputs that needn't check, like XShape of reshape op.
            inplace_atol (float): The tolerable error, only set when op doesn't ensure computational consistency, like group_norm op.

        Returns:
C
cc 已提交
819 820
            expect_res (tuple(outs, fetch_list, feed_map, program, op_desc)): The results of given op.
                We return this to construct grad_program and grad_feed_map for grad inplace check.
821 822
        """
        # _calc_output() returns in the form tuple(outs, fetch_list, feed_map, program, op_desc) when for_inplace_test=True.
823 824 825 826 827 828 829 830 831 832
        expect_res = self._calc_output(
            place,
            no_check_set=no_check_set,
            enable_inplace=False,
            for_inplace_test=True)
        actual_res = self._calc_output(
            place,
            no_check_set=no_check_set,
            enable_inplace=True,
            for_inplace_test=True)
833
        # compare expect_outs and actual_outs
834 835 836 837 838 839
        self._compare_expect_and_actual_outputs(
            place,
            expect_res[1],
            expect_res[0],
            actual_res[0],
            inplace_atol=inplace_atol)
840 841 842 843 844 845 846 847 848 849 850 851 852
        return expect_res

    def _calc_grad_output(self,
                          place,
                          fwd_res,
                          grad_op_desc,
                          enable_inplace=None):
        """Calculate grad_output for given grad_op_desc.

        since we don`t really check gradient accuracy, but check the consistency when using and not using inplace,
        we use fwd outs (also inputs sometimes) to construct grad inputs.

        Args:
C
cc 已提交
853
            place (CPUPlace | CUDAPlace): The place where the op runs.
854 855 856 857 858 859 860 861 862
            fwd_res (tuple): The outputs of its forward op, in the same form as returns of _calc_outputs() when for_inplace_test is True.
                i.e., tuple(fwd_outs, fwd_fetch_list, fwd_feed_map, fwd_program, fwd_op_desc).
            grad_op_desc (OpDesc): The OpDesc of grad op.
            enable_inplace (bool): Enable inplace or not.

        Returns:
            res (tuple(outs, fetch_list, feed_map, program, op_desc)): The results of given grad_op_desc.
        """
        fwd_outs, fwd_fetch_list, fwd_feed_map, fwd_program, fwd_op_desc = fwd_res
863
        grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc(fwd_op_desc,
864
                                                                  set(), [])
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
        grad_program = self._construct_grad_program_from_forward(
            fwd_program, grad_op_desc, op_grad_to_var)
        grad_feed_map = self._construct_grad_feed_map_from_forward(
            place, fwd_res, grad_op_desc, op_grad_to_var)
        grad_fetch_list = grad_op_desc.output_arg_names()
        exe = Executor(place)
        program = grad_program
        if enable_inplace is not None:
            build_strategy = fluid.BuildStrategy()
            build_strategy.enable_inplace = enable_inplace
            compiled_program = fluid.CompiledProgram(
                grad_program).with_data_parallel(
                    loss_name="", build_strategy=build_strategy, places=place)
            program = compiled_program
        outs = exe.run(program,
                       feed=grad_feed_map,
                       fetch_list=grad_fetch_list,
                       return_numpy=False)
        return outs, grad_fetch_list, grad_feed_map, grad_program, grad_op_desc

    def _check_grad_inplace(self,
                            place,
                            fwd_res,
                            grad_op_desc,
                            inplace_atol=None):
        """Chech the inplace correctness of given grad_op_desc.

        Run the grad op twice with same inputs, one enable inplace and another disable, compare their outputs.
        It works like _check_forward_inplace, but the way to construct program and feed_map differs.
        So we define a new function for grad, grad_grad, etc.

        Args:
C
cc 已提交
897
            place (CPUPlace | CUDAPlace): The place where the op runs.
898 899 900 901 902 903
            fwd_res (tuple): The outputs of its forward op, in the same form as returns of _calc_outputs() when for_inplace_test is True.
                i.e., tuple(fwd_outs, fwd_fetch_list, fwd_feed_map, fwd_program, fwd_op_desc).
            grad_op_desc (OpDesc): The OpDesc of grad op.
            inplace_atol (float): The tolerable error, only set when op doesn't ensure computational consistency, like group_norm op.

        Returns:
C
cc 已提交
904 905
            expect_res (tuple(outs, fetch_list, feed_map, program, op_desc)): The results of given op.
                We return this to construct grad_program and grad_feed_map for grad inplace check.
906 907 908 909 910 911 912 913 914 915 916 917
        """
        expect_res = self._calc_grad_output(
            place, fwd_res, grad_op_desc, enable_inplace=False)
        actual_res = self._calc_grad_output(
            place, fwd_res, grad_op_desc, enable_inplace=True)
        self._compare_expect_and_actual_outputs(
            place,
            expect_res[1],
            expect_res[0],
            actual_res[0],
            inplace_atol=inplace_atol)
        return expect_res
918

919 920 921 922 923 924 925 926 927 928
    def check_inplace_output_with_place(self,
                                        place,
                                        no_check_set=None,
                                        inplace_atol=None):
        """Chech the inplace correctness of given op, its grad op, its grad_grad op, etc.

        (1) Get all ops need to run. (see conditions in _get_need_run_ops())
        (2) Run op in need_run_ops, and do inplace check if it has infer_inplace.

        Args:
C
cc 已提交
929
            place (CPUPlace | CUDAPlace): The place where the op runs.
930 931 932 933 934 935 936 937 938 939 940 941 942 943 944
            no_check_set (list): The names of outputs that needn't check, like XShape of reshape op.
            inplace_atol (float): The tolerable error, only set when op doesn't ensure computational consistency, like group_norm op.

        Returns:
            None
        """
        has_infer_inplace = fluid.core.has_infer_inplace(self.op_type)
        has_grad_op_maker = fluid.core.has_grad_op_maker(self.op_type)

        fwd_res = self._calc_output(
            place, no_check_set=no_check_set, for_inplace_test=True)
        op_desc = fwd_res[4]
        need_run_ops = self._get_need_run_ops(op_desc)

        res = {}
945 946
        if hasattr(self, 'attrs') and bool(self.attrs.get('use_xpu', False)):
            return
947 948 949 950 951 952 953 954 955 956 957 958 959
        for op_desc, father_op_desc in reversed(need_run_ops):
            # The first one is the forward op
            has_infer_inplace = fluid.core.has_infer_inplace(op_desc.type())
            if op_desc.type() == self.op_type:
                if has_infer_inplace:
                    res[op_desc] = self._check_forward_inplace(
                        place,
                        no_check_set=no_check_set,
                        inplace_atol=inplace_atol)
                else:
                    res[op_desc] = self._calc_output(
                        place, no_check_set=no_check_set, for_inplace_test=True)
            else:
960 961
                # TODO(zhiqiu): enhance inplace_grad test for ops (sum and activation) using mkldnn
                # skip op that use_mkldnn currently
962
                flags_use_mkldnn = fluid.core.globals()["FLAGS_use_mkldnn"]
963 964 965 966 967 968 969 970 971 972 973 974
                attrs_use_mkldnn = hasattr(
                    self,
                    'attrs') and bool(self.attrs.get('use_mkldnn', False))
                if flags_use_mkldnn or attrs_use_mkldnn:
                    warnings.warn(
                        "check inplace_grad for ops using mkldnn is not supported"
                    )
                    continue
                if has_infer_inplace:
                    fwd_res = res[father_op_desc]
                    res[op_desc] = self._check_grad_inplace(
                        place, fwd_res, op_desc, inplace_atol=inplace_atol)
975
                else:
976 977
                    res[op_desc] = self._calc_grad_output(place, fwd_res,
                                                          op_desc)
978

979 980
    def check_output_with_place(self,
                                place,
981
                                atol=0,
982
                                no_check_set=None,
M
minqiyang 已提交
983
                                equal_nan=False,
984
                                check_dygraph=True,
985
                                inplace_atol=None):
986 987 988 989 990
        self.infer_dtype_from_inputs_outputs(self.inputs, self.outputs)
        if self.dtype == np.float64 and \
            self.op_type not in op_threshold_white_list.NEED_FIX_FP64_CHECK_OUTPUT_THRESHOLD_OP_LIST:
            atol = 0

991 992 993 994 995 996 997 998
        if self.is_bfloat16_op():
            check_dygraph = False
            if hasattr(self, 'force_fp32_output') and getattr(
                    self, 'force_fp32_output'):
                atol = 1e-2
            else:
                atol = 2

999 1000 1001 1002
        if no_check_set is not None:
            if self.op_type not in no_check_set_white_list.no_check_set_white_list:
                raise AssertionError(
                    "no_check_set of op %s must be set to None." % self.op_type)
1003

L
lujun 已提交
1004 1005
        if check_dygraph:
            dygraph_outs = self._calc_dygraph_output(
M
minqiyang 已提交
1006
                place, no_check_set=no_check_set)
1007
        outs, fetch_list = self._calc_output(place, no_check_set=no_check_set)
Y
Yang Yang(Tony) 已提交
1008
        for out_name, out_dup in Operator.get_op_outputs(self.op_type):
1009 1010
            if out_name not in self.outputs:
                continue
1011 1012
            if no_check_set is not None and out_name in no_check_set:
                continue
1013

1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
            def find_imperative_actual(target_name, dygraph_outs, place):
                with fluid.dygraph.base.guard(place=place):
                    for name in dygraph_outs:
                        if name == target_name:
                            return dygraph_outs[name][0]
                        var_list = dygraph_outs[name]
                        for i, var in enumerate(var_list):
                            if var.name == target_name:
                                return dygraph_outs[name][i]
                    self.assertTrue(False, "Found failed {} {}".format(
                        dygraph_outs.keys(), target_name))

Y
Yang Yang(Tony) 已提交
1026 1027
            def find_actual(target_name, fetch_list):
                found = [
1028 1029
                    i for i, var_name in enumerate(fetch_list)
                    if var_name == target_name
Y
Yang Yang(Tony) 已提交
1030 1031 1032 1033 1034 1035
                ]
                self.assertTrue(
                    len(found) == 1, "Found {} {}".format(
                        len(found), target_name))
                return found[0]

1036 1037
            if out_dup:
                sub_out = self.outputs[out_name]
Y
Yancey 已提交
1038 1039 1040
                if not isinstance(sub_out, list):
                    raise AssertionError("sub_out type %s is not list",
                                         type(sub_out))
1041 1042
                for item in sub_out:
                    sub_out_name, expect = item[0], item[1]
L
lujun 已提交
1043
                    if check_dygraph:
1044 1045
                        imperative_actual = find_imperative_actual(
                            sub_out_name, dygraph_outs, place)
1046 1047
                        imperative_actual_t = np.array(imperative_actual.value()
                                                       .get_tensor())
Y
Yang Yang(Tony) 已提交
1048
                    idx = find_actual(sub_out_name, fetch_list)
Q
QI JUN 已提交
1049 1050
                    actual = outs[idx]
                    actual_t = np.array(actual)
1051 1052
                    expect_t = expect[0] \
                        if isinstance(expect, tuple) else expect
1053 1054
                    self.assertTrue(
                        np.allclose(
1055
                            actual_t, expect_t, atol=atol, equal_nan=equal_nan),
Y
Yang Yang(Tony) 已提交
1056 1057
                        "Output (" + sub_out_name + ") has diff at " +
                        str(place))
L
lujun 已提交
1058
                    if check_dygraph:
M
minqiyang 已提交
1059 1060 1061 1062 1063 1064 1065
                        self.assertTrue(
                            np.allclose(
                                imperative_actual_t,
                                expect_t,
                                atol=atol,
                                equal_nan=equal_nan),
                            "Output (" + sub_out_name + ") has diff at " +
L
lujun 已提交
1066
                            str(place) + " in dygraph mode")
1067 1068
                    if isinstance(expect, tuple):
                        self.assertListEqual(
1069 1070
                            actual.recursive_sequence_lengths(), expect[1],
                            "Output (" + sub_out_name +
Q
QI JUN 已提交
1071
                            ") has different lod at " + str(place))
1072 1073
                        if check_dygraph:
                            self.assertListEqual(
1074
                                imperative_actual.value().get_tensor()
1075 1076 1077 1078
                                .recursive_sequence_lengths(), expect[1],
                                "Output (" + out_name +
                                ") has different lod at " + str(place) +
                                " in dygraph mode")
1079
            else:
L
lujun 已提交
1080
                if check_dygraph:
1081 1082
                    imperative_actual = find_imperative_actual(
                        out_name, dygraph_outs, place)
1083 1084
                    imperative_actual_t = np.array(imperative_actual.value()
                                                   .get_tensor())
Y
Yang Yang(Tony) 已提交
1085
                idx = find_actual(out_name, fetch_list)
Q
QI JUN 已提交
1086 1087
                actual = outs[idx]
                actual_t = np.array(actual)
1088
                expect = self.outputs[out_name]
1089
                expect_t = expect[0] if isinstance(expect, tuple) else expect
1090 1091
                self.assertTrue(
                    np.allclose(
1092
                        actual_t, expect_t, atol=atol, equal_nan=equal_nan),
E
emailweixu 已提交
1093
                    "Output (" + out_name + ") has diff at " + str(place) +
D
dzhwinter 已提交
1094
                    "\nExpect " + str(expect_t) + "\n" + "But Got" +
1095
                    str(actual_t) + " in class " + self.__class__.__name__)
L
lujun 已提交
1096
                if check_dygraph:
1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112
                    if six.moves.reduce(
                            lambda x, y: x * y, imperative_actual_t.shape,
                            1) == 0 and six.moves.reduce(
                                lambda x, y: x * y, expect_t.shape, 1) == 0:
                        pass
                    else:
                        self.assertTrue(
                            np.allclose(
                                imperative_actual_t,
                                expect_t,
                                atol=atol,
                                equal_nan=equal_nan),
                            "Output (" + out_name + ") has diff at " +
                            str(place) + "\nExpect " + str(expect_t) + "\n" +
                            "But Got" + str(imperative_actual_t) + " in class "
                            + self.__class__.__name__)
1113
                if isinstance(expect, tuple):
1114 1115
                    self.assertListEqual(actual.recursive_sequence_lengths(),
                                         expect[1], "Output (" + out_name +
1116
                                         ") has different lod at " + str(place))
L
lujun 已提交
1117
                    if check_dygraph:
M
minqiyang 已提交
1118
                        self.assertListEqual(
1119
                            imperative_actual.value().get_tensor()
M
minqiyang 已提交
1120 1121
                            .recursive_sequence_lengths(), expect[1],
                            "Output (" + out_name + ") has different lod at " +
L
lujun 已提交
1122
                            str(place) + " in dygraph mode")
1123

C
cc 已提交
1124
        # Note(zhiqiu): inplace_atol should be only set when op doesn't ensure
L
Leo Chen 已提交
1125 1126
        # computational consistency.
        # For example, group_norm uses AtomicAdd on CUDAPlace, which do not ensure
C
cc 已提交
1127
        # computation order when multiple threads write the same address. So the
L
Leo Chen 已提交
1128 1129 1130
        # result of group_norm is non-deterministic when datatype is float.
        # When inplace_atol is not None, the inplace check uses numpy.allclose
        # to check inplace result instead of numpy.array_equal.
1131 1132
        if inplace_atol is not None:
            warnings.warn(
L
Leo Chen 已提交
1133 1134
                "inplace_atol should only be set when op doesn't ensure computational consistency, please check it!"
            )
1135
        # Check inplace for given op, its grad op, its grad_grad op, etc.
C
cc 已提交
1136
        # No effect on original OpTest
1137 1138 1139 1140
        # Currently not support ParallelExecutor on XPUPlace.
        if not paddle.is_compiled_with_xpu():
            self.check_inplace_output_with_place(
                place, no_check_set=no_check_set, inplace_atol=inplace_atol)
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
        if check_dygraph:
            return outs, dygraph_outs, fetch_list
        else:
            return outs, fetch_list

    def check_compile_vs_runtime(self, fetch_list, fetch_outs):
        def find_fetch_index(target_name, fetch_list):
            found = [
                i for i, var_name in enumerate(fetch_list)
                if var_name == target_name
            ]
            if len(found) == 0:
                return -1
            else:
                self.assertTrue(
                    len(found) == 1,
                    "Found {} {}".format(len(found), target_name))
                return found[0]

        for name in self.op.desc.output_names():
            var_names = self.op.desc.output(name)
            for var_name in var_names:
                i = find_fetch_index(var_name, fetch_list)
                if i == -1:
                    # The output is dispensiable or intermediate.
                    break
                out = fetch_outs[i]
                if isinstance(out, core.LoDTensor):
                    lod_level_runtime = len(out.lod())
                else:
                    if isinstance(out, core.LoDTensorArray):
                        warnings.warn(
                            "The check of LoDTensorArray's lod_level is not implemented now!"
                        )
                    lod_level_runtime = 0

                var = self.program.global_block().var(var_name)
                if var.type == core.VarDesc.VarType.LOD_TENSOR:
                    lod_level_compile = var.lod_level
                else:
                    lod_level_compile = 0
                self.assertEqual(
                    lod_level_compile, lod_level_runtime,
                    "The lod_level of Output (" + name +
                    ") is different between compile-time and runtime (" +
                    str(lod_level_compile) + " vs " + str(lod_level_runtime) +
                    ")")

1190
    def _get_places(self):
D
dzhwinter 已提交
1191 1192 1193 1194 1195 1196
        if self.dtype == np.float16:
            if core.is_compiled_with_cuda() and core.op_support_gpu(
                    self.op_type):
                place = core.CUDAPlace(0)
                if core.is_float16_supported(place):
                    return [place]
W
Wu Yi 已提交
1197 1198
                else:
                    return []
D
dzhwinter 已提交
1199 1200
            else:
                return []
1201
        places = [fluid.CPUPlace()]
1202 1203 1204
        cpu_only = self._cpu_only if hasattr(self, '_cpu_only') else False
        if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type)\
           and not cpu_only:
D
dzhwinter 已提交
1205
            places.append(core.CUDAPlace(0))
1206 1207
        return places

M
minqiyang 已提交
1208 1209 1210 1211
    def check_output(self,
                     atol=1e-5,
                     no_check_set=None,
                     equal_nan=False,
1212
                     check_dygraph=True,
1213
                     inplace_atol=None):
1214
        self.__class__.op_type = self.op_type
1215 1216 1217 1218
        if (hasattr(self, "use_mkldnn") and self.use_mkldnn == True) or \
            (hasattr(self, "attrs") and "use_mkldnn" in self.attrs and \
                    self.attrs["use_mkldnn"] == True):
            self.__class__.use_mkldnn = True
C
cc 已提交
1219

1220 1221 1222 1223 1224
        if (hasattr(self, "use_xpu") and self.use_xpu == True) or \
            (hasattr(self, "attrs") and "use_xpu" in self.attrs and \
                    self.attrs["use_xpu"] == True):
            self.__class__.use_xpu = True

1225
        places = self._get_places()
Q
qijun 已提交
1226
        for place in places:
1227 1228 1229 1230 1231 1232
            res = self.check_output_with_place(place, atol, no_check_set,
                                               equal_nan, check_dygraph)
            if check_dygraph:
                outs, dygraph_outs, fetch_list = res
            else:
                outs, fetch_list = res
1233
            if self.op_type not in compile_vs_runtime_white_list.COMPILE_RUN_OP_WHITE_LIST:
1234
                self.check_compile_vs_runtime(fetch_list, outs)
Q
qijun 已提交
1235

1236
    def check_output_customized(self, checker):
1237
        places = self._get_places()
1238 1239 1240
        for place in places:
            outs = self.calc_output(place)
            outs = [np.array(out) for out in outs]
1241
            outs.sort(key=len)
1242 1243
            checker(outs)

D
Dun 已提交
1244 1245
    def _assert_is_close(self, numeric_grads, analytic_grads, names,
                         max_relative_error, msg_prefix):
1246

M
minqiyang 已提交
1247
        for a, b, name in six.moves.zip(numeric_grads, analytic_grads, names):
1248 1249 1250 1251 1252 1253
            # It asserts np.abs(a - b) / np.abs(a) < max_relative_error, in which
            # max_relative_error is 1e-7. According to the value of np.abs(a), we
            # change np.abs(a) to achieve dynamic threshold. For example, if
            # the value of np.abs(a) is between 1e-10 and 1e-8, we set np.abs(a)*=1e4.
            # Therefore, it asserts np.abs(a - b) / (np.abs(a)*1e4) < max_relative_error,
            # which is the same as np.abs(a - b) / np.abs(a) < max_relative_error*1e4.
1254
            abs_a = np.abs(a)
1255 1256 1257 1258 1259 1260 1261
            if self.dtype == np.float64 and \
                self.op_type not in op_threshold_white_list.NEED_FIX_FP64_CHECK_GRAD_THRESHOLD_OP_LIST:
                abs_a[abs_a < 1e-10] = 1e-3
                abs_a[np.logical_and(abs_a > 1e-10, abs_a <= 1e-8)] *= 1e4
                abs_a[np.logical_and(abs_a > 1e-8, abs_a <= 1e-6)] *= 1e2
            else:
                abs_a[abs_a < 1e-3] = 1
1262 1263 1264 1265 1266 1267

            diff_mat = np.abs(a - b) / abs_a
            max_diff = np.max(diff_mat)

            def err_msg():
                offset = np.argmax(diff_mat > max_relative_error)
1268 1269 1270
                return ("Operator %s error, %s variable %s (shape: %s, dtype: %s) max gradient diff %e over limit %e, "
                    "the first error element is %d, expected %e, but got %e.") \
                    % (self.op_type, msg_prefix, name, str(a.shape), self.dtype, max_diff, max_relative_error,
1271
                    offset, a.flatten()[offset], b.flatten()[offset])
1272 1273 1274

            self.assertLessEqual(max_diff, max_relative_error, err_msg())

1275 1276 1277 1278 1279 1280 1281
    def _check_grad_helper(self):
        self.infer_dtype_from_inputs_outputs(self.inputs, self.outputs)
        self.__class__.op_type = self.op_type
        self.__class__.exist_check_grad = True
        if self.dtype == np.float64:
            self.__class__.exist_fp64_check_grad = True

1282 1283
    def check_grad(self,
                   inputs_to_check,
Y
Yancey 已提交
1284
                   output_names,
1285
                   no_grad_set=None,
1286
                   numeric_grad_delta=0.005,
1287
                   in_place=False,
Q
Qiao Longfei 已提交
1288
                   max_relative_error=0.005,
1289 1290
                   user_defined_grads=None,
                   check_dygraph=True):
1291
        self._check_grad_helper()
1292
        places = self._get_places()
1293 1294 1295 1296
        for place in places:
            self.check_grad_with_place(place, inputs_to_check, output_names,
                                       no_grad_set, numeric_grad_delta,
                                       in_place, max_relative_error,
1297
                                       user_defined_grads, check_dygraph)
1298 1299 1300 1301 1302 1303 1304 1305 1306

    def check_grad_with_place(self,
                              place,
                              inputs_to_check,
                              output_names,
                              no_grad_set=None,
                              numeric_grad_delta=0.005,
                              in_place=False,
                              max_relative_error=0.005,
1307 1308
                              user_defined_grads=None,
                              check_dygraph=True):
1309
        self.scope = core.Scope()
Q
qijun 已提交
1310
        op_inputs = self.inputs if hasattr(self, "inputs") else dict()
1311
        op_outputs = self.outputs if hasattr(self, "outputs") else dict()
Q
qijun 已提交
1312
        op_attrs = self.attrs if hasattr(self, "attrs") else dict()
P
phlrain 已提交
1313

1314
        self._check_grad_helper()
1315 1316 1317 1318
        if self.dtype == np.float64 and \
            self.op_type not in op_threshold_white_list.NEED_FIX_FP64_CHECK_GRAD_THRESHOLD_OP_LIST:
            numeric_grad_delta = 1e-5
            max_relative_error = 1e-7
1319

P
phlrain 已提交
1320 1321 1322
        cache_list = None
        if hasattr(self, "cache_name_list"):
            cache_list = self.cache_name_list
1323 1324 1325 1326 1327 1328 1329

        # oneDNN numeric gradient should use CPU kernel
        use_onednn = False
        if "use_mkldnn" in op_attrs and op_attrs["use_mkldnn"] == True:
            op_attrs["use_mkldnn"] = False
            use_onednn = True

P
phlrain 已提交
1330 1331 1332 1333 1334 1335 1336
        self.op = create_op(
            self.scope,
            self.op_type,
            op_inputs,
            op_outputs,
            op_attrs,
            cache_list=cache_list)
Y
Yu Yang 已提交
1337

1338 1339 1340
        if use_onednn:
            op_attrs["use_mkldnn"] = True

1341 1342
        if no_grad_set is None:
            no_grad_set = set()
1343 1344
        else:
            if (self.op_type not in no_grad_set_white_list.NEED_TO_FIX_OP_LIST
1345 1346 1347
                ) and (
                    self.op_type not in no_grad_set_white_list.NOT_CHECK_OP_LIST
                ) and (not self.is_bfloat16_op()):
1348 1349
                raise AssertionError("no_grad_set must be None, op_type is " +
                                     self.op_type + " Op.")
1350

1351 1352 1353 1354 1355 1356 1357 1358
        for input_to_check in inputs_to_check:
            set_input(self.scope, self.op, self.inputs, place)
            tensor_to_check = self.scope.find_var(input_to_check).get_tensor()
            tensor_size = six.moves.reduce(lambda a, b: a * b,
                                           tensor_to_check.shape(), 1)
            if tensor_size < 100:
                self.__class__.input_shape_is_large = False

Y
Yancey 已提交
1359 1360 1361
        if not type(output_names) is list:
            output_names = [output_names]

Q
Qiao Longfei 已提交
1362
        numeric_grads = user_defined_grads or [
1363
            get_numeric_gradient(
1364
                place,
1365 1366 1367 1368
                self.scope,
                self.op,
                self.inputs,
                input_to_check,
Y
Yancey 已提交
1369
                output_names,
1370
                delta=numeric_grad_delta,
C
chengduo 已提交
1371
                in_place=in_place) for input_to_check in inputs_to_check
1372
        ]
1373 1374
        analytic_grads = self._get_gradient(inputs_to_check, place,
                                            output_names, no_grad_set)
D
Dun 已提交
1375 1376 1377
        self._assert_is_close(numeric_grads, analytic_grads, inputs_to_check,
                              max_relative_error,
                              "Gradient Check On %s" % str(place))
Q
qijun 已提交
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 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486
        if check_dygraph:
            dygraph_grad = self._get_dygraph_grad(inputs_to_check, place,
                                                  output_names, no_grad_set)
            self._assert_is_close(numeric_grads, dygraph_grad, inputs_to_check,
                                  max_relative_error,
                                  "Gradient Check On %s" % str(place))

    def _find_var_in_dygraph(self, output_vars, name):
        if name in output_vars:
            return output_vars[name]
        else:
            for output_vars_index in output_vars:
                for output_vars_selected in output_vars[output_vars_index]:
                    if output_vars_selected.name == name:
                        return output_vars_selected

    def _get_dygraph_grad(self,
                          inputs_to_check,
                          place,
                          output_names,
                          no_grad_set=None):
        with fluid.dygraph.base.guard(place=place):
            block = fluid.default_main_program().global_block()

            op_proto = OpProtoHolder.instance().get_op_proto(self.op_type)

            # prepare input variable
            inputs, inputs_grad_dict = self.append_input_output_for_dygraph(
                op_proto, self.inputs, True, True, block)

            # prepare output variable
            outputs = self.append_input_output_for_dygraph(
                op_proto, self.outputs, False, False, block)

            # prepare attrbutes
            attrs_outputs = {}
            if hasattr(self, "attrs"):
                for attrs_name in self.attrs:
                    if self.attrs[attrs_name] is not None:
                        attrs_outputs[attrs_name] = self.attrs[attrs_name]
            block.append_op(
                type=self.op_type,
                inputs=inputs,
                outputs=outputs,
                attrs=attrs_outputs if hasattr(self, "attrs") else None)

            outputs_valid = {}
            for output_name in output_names:
                outputs_valid[output_name] = self._find_var_in_dygraph(
                    outputs, output_name)

            if len(outputs_valid) == 1:
                loss = block.create_var(
                    dtype=self.dtype,
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    persistable=False,
                    stop_gradient=False,
                    shape=[1])
                for outputs_valid_key in outputs_valid:
                    block.append_op(
                        type="mean",
                        inputs={"X": outputs_valid[outputs_valid_key]},
                        outputs={"Out": [loss]},
                        attrs=None)
            else:
                avg_sum = []
                for cur_loss in outputs_valid:
                    cur_avg_loss = block.create_var(
                        dtype=self.dtype,
                        type=core.VarDesc.VarType.LOD_TENSOR,
                        persistable=False,
                        stop_gradient=False)
                    block.append_op(
                        type="mean",
                        inputs={"X": outputs_valid[cur_loss]},
                        outputs={"Out": [cur_avg_loss]},
                        attrs=None)
                    avg_sum.append(cur_avg_loss)
                loss_sum = block.create_var(
                    dtype=self.dtype,
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    persistable=False,
                    stop_gradient=False,
                    shape=[1])
                block.append_op(
                    type='sum',
                    inputs={"X": avg_sum},
                    outputs={"Out": loss_sum},
                    attrs=None)
                loss = block.create_var(
                    dtype=self.dtype,
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    persistable=False,
                    stop_gradient=False,
                    shape=[1])
                block.append_op(
                    type='scale',
                    inputs={"X": loss_sum},
                    outputs={"Out": loss},
                    attrs={'scale': 1.0 / float(len(avg_sum))})
            loss.backward()

            fetch_list_grad = []
            for inputs_to_check_name in inputs_to_check:
                a = inputs_grad_dict[inputs_to_check_name].gradient()
                fetch_list_grad.append(a)
            return fetch_list_grad

Y
Yu Yang 已提交
1487 1488 1489 1490 1491
    @staticmethod
    def _numpy_to_lod_tensor(np_value, lod, place):
        tensor = core.LoDTensor()
        tensor.set(np_value, place)
        if lod is not None:
1492
            tensor.set_recursive_sequence_lengths(lod)
Y
Yu Yang 已提交
1493 1494
        return tensor

K
Kexin Zhao 已提交
1495
    @staticmethod
K
Kexin Zhao 已提交
1496 1497
    def np_dtype_to_fluid_dtype(input):
        return input
K
Kexin Zhao 已提交
1498

D
dzhwinter 已提交
1499 1500 1501 1502 1503 1504 1505 1506
    @staticmethod
    def fluid_dtype_to_np_dtype(self, dtype):
        return dtype

    @staticmethod
    def np_value_to_fluid_value(input):
        return input

1507 1508 1509 1510 1511 1512
    def _get_gradient(self,
                      input_to_check,
                      place,
                      output_names,
                      no_grad_set,
                      parallel=False):
Y
Yu Yang 已提交
1513 1514
        prog = Program()
        block = prog.global_block()
1515 1516
        self._append_ops(block)
        loss = append_loss_ops(block, output_names)
F
fengjiayi 已提交
1517
        param_grad_list = append_backward(
Y
Yu Yang 已提交
1518 1519
            loss=loss, parameter_list=input_to_check, no_grad_set=no_grad_set)

1520 1521
        inputs = self._get_inputs(block)
        feed_dict = self.feed_var(inputs, place)
Y
Yu Yang 已提交
1522 1523

        fetch_list = [g for p, g in param_grad_list]
1524 1525
        if parallel:
            use_cuda = False
1526
            if isinstance(place, fluid.CUDAPlace):
1527
                use_cuda = True
1528 1529 1530 1531
            compiled_prog = fluid.CompiledProgram(prog).with_data_parallel(
                loss_name=loss.name, places=place)
            prog = compiled_prog
        executor = fluid.Executor(place)
1532 1533 1534
        return list(
            map(np.array,
                executor.run(prog, feed_dict, fetch_list, return_numpy=False)))