executor.py 20.9 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 os
import multiprocessing
D
dzhwinter 已提交
19
import numpy as np
Y
Yang Yu 已提交
20
import contextlib
21
import six
22
from .framework import Program, default_main_program, Variable
23
from . import core
24 25
from . import compiler
from .. import compat as cpt
26

T
Tink_Y 已提交
27
__all__ = ['Executor', 'global_scope', 'scope_guard']
Y
Yu Yang 已提交
28

Y
Yu Yang 已提交
29 30
g_scope = core.Scope()

Y
Yu Yang 已提交
31

Y
Yang Yu 已提交
32
def global_scope():
Y
yuyang18 已提交
33 34 35 36 37 38 39
    """
    Get the global/default scope instance. There are a lot of APIs use
    :code:`global_scope` as its default value, e.g., :code:`Executor.run`

    Returns:
        Scope: The global/default scope instance.
    """
Y
Yang Yu 已提交
40 41 42
    return g_scope


43
def _switch_scope(scope):
Y
Yang Yu 已提交
44 45 46 47 48 49 50 51
    global g_scope
    ex = g_scope
    g_scope = scope
    return ex


@contextlib.contextmanager
def scope_guard(scope):
Y
yuyang18 已提交
52 53 54 55 56 57 58 59 60 61 62 63 64
    """
    Change the global/default scope instance by Python `with` statement. All
    variable in runtime will assigned to the new scope.

    Examples:
        >>> import paddle.fluid as fluid
        >>> new_scope = fluid.Scope()
        >>> with fluid.scope_guard(new_scope):
        >>>     ...

    Args:
        scope: The new global/default scope.
    """
65
    ex = _switch_scope(scope)
Y
Yang Yu 已提交
66
    yield
67
    _switch_scope(ex)
Y
Yang Yu 已提交
68 69


D
dzhwinter 已提交
70
def as_numpy(tensor):
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
    """
    Convert a Tensor to a numpy.ndarray, its only support Tensor without LoD information.
    For higher dimensional sequence data, please use LoDTensor directly.
    Examples:
        >>> import paddle.fluid as fluid
        >>> outs = executor.run(...)
        >>> np_outs = map(lambda x: as_numpy(x), outs)
        >>>     ...

    Args:
       tensor(Variable): a instance of Tensor

    Returns:
        numpy.ndarray
    """
C
chengduo 已提交
86 87
    if isinstance(tensor, core.LoDTensorArray):
        return [as_numpy(t) for t in tensor]
D
dzhwinter 已提交
88 89 90 91
    if isinstance(tensor, list):
        return [as_numpy(t) for t in tensor]
    assert isinstance(tensor, core.LoDTensor)
    lod = tensor.lod()
92
    if len(lod) > 0:
D
dzhwinter 已提交
93
        raise RuntimeError("Some of your fetched tensors hold LoD information. \
94 95 96 97
            They can not be completely cast to Python ndarray. \
            Please set the parameter 'return_numpy' as 'False' to \
            return LoDTensor itself directly.")
    return np.array(tensor)
D
dzhwinter 已提交
98 99


100 101 102 103 104 105 106 107 108 109 110 111
def has_feed_operators(block, feed_targets, feed_holder_name):
    """ Check whether the block already has feed operators.

    Return false if the block does not have any feed operators.
    If some feed operators have been prepended to the block, check that
    the info contained in these feed operators matches the feed_targets
    and feed_holder_name. Raise exception when any mismatch is found.
    Return true when the block has feed operators with matching info.

    Args:
        block: a block instance (typically global block of a program)
        feed_targets: a dictionary of {feed_target_name: feed_target_data}
X
xuwei06 已提交
112 113
        feed_holder_name: the name of the variable that holds the data of
            all feed targets. The type of this feed_holder variable is
114 115 116
            FEED_MINIBATCH, which is essentially vector<LoDTensor>.

    Returns:
X
xuwei06 已提交
117
        A boolean value that indicates whether a block has feed operators
118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139
        that match the info contained in feed_targets and feed_holder_name.
    """

    feed_count = 0
    for op in block.ops:
        if op.desc.type() == 'feed':
            feed_count += 1
            assert op.desc.input('X')[0] == feed_holder_name
            feed_target_name = op.desc.output('Out')[0]
            if feed_target_name not in feed_targets:
                raise Exception("'feed_targets' does not have {} variable".
                                format(feed_target_name))
        else:
            break
    if feed_count > 0 and feed_count != len(feed_targets):
        raise Exception(
            "Feed operators in program desc do not match 'feed_targets'")
    return feed_count > 0


def has_fetch_operators(block, fetch_targets, fetch_holder_name):
    """ Check whether the block already has fetch operators.
X
xuwei06 已提交
140

141 142 143 144 145 146 147 148 149
    Return false if the block does not have any fetch operators.
    If some fetch operators have been appended to the block, check that
    the info contained in these fetch operators matches the fetch_targets
    and fetch_holder_name. Raise exception when any mismatch is found.
    Return true when the block has fetch operators with matching info.

    Args:
        block: a block instance (typically global block of a program)
        fetch_targets: a dictionary of {fetch_target_name: fetch_target_data}
X
xuwei06 已提交
150 151 152
        fetch_holder_name: the name of the variable that holds the data of
            all fetch targets. The type of this fetch_holder variable is
            FETCH_LIST, which is essentially vector<LoDTensor>.
153

X
xuwei06 已提交
154 155 156
    Return:
        A boolean value that indicates whether a block has fetch operators
        that match the info contained in fetch_targets and fetch_holder_name.
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177
    """

    fetch_count = 0
    for op in block.ops:
        if op.desc.type() == 'fetch':
            fetch_count += 1
            assert op.desc.output('Out')[0] == fetch_holder_name
            fetch_target_name = op.desc.input('X')[0]
            if fetch_target_name not in [
                    var.desc.name() for var in fetch_targets
            ]:
                raise Exception("'fetch_targets' does not have {} variable".
                                format(fetch_target_name))
            idx = op.desc.attr('col')
            assert fetch_target_name == fetch_targets[idx].desc.name()
    if fetch_count > 0 and fetch_count != len(fetch_targets):
        raise Exception(
            "Fetch operators in program desc do not match 'fetch_targets'")
    return fetch_count > 0


W
Wu Yi 已提交
178
def _fetch_var(name, scope=None, return_numpy=True):
X
xuwei06 已提交
179
    """
C
chengduoZH 已提交
180 181 182
    Fetch the value of the variable with the given name from the
    given scope.

X
xuwei06 已提交
183
    Args:
184 185 186 187
        name(str): name of the variable. Typically, only persistable variables
            can be found in the scope used for running the program.
        scope(core.Scope|None): scope object. It should be the scope where
            you pass to Executor.run() when running your program.
C
chengduoZH 已提交
188 189 190 191
            If None, global_scope() will be used. Default None.
        return_numpy(bool): whether convert the tensor to numpy.ndarray.
            Default True.

X
xuwei06 已提交
192 193 194 195 196 197
    Returns:
       LodTensor|numpy.ndarray
    """
    assert isinstance(name, str)
    if scope is None:
        scope = global_scope()
S
sneaxiy 已提交
198
    assert isinstance(scope, core._Scope)
X
xuwei06 已提交
199

Y
Yibing Liu 已提交
200
    var = scope.find_var(name)
201 202 203 204
    assert var is not None, (
        "Cannot find " + name + " in scope. Perhaps you need to make the"
        " variable persistable by using var.persistable = True in your"
        " program.")
X
xuwei06 已提交
205 206 207 208 209 210
    tensor = var.get_tensor()
    if return_numpy:
        tensor = as_numpy(tensor)
    return tensor


211
def _get_program_cache_key(feed, fetch_list):
212
    feed_var_names = list(feed.keys())
Q
qiaolongfei 已提交
213 214 215 216 217 218

    def to_name_str(var):
        if isinstance(var, Variable):
            return var.desc.name()
        elif isinstance(var, str):
            return var
219
        elif isinstance(var, six.string_types):
220
            return str(var)
Q
qiaolongfei 已提交
221 222 223
        else:
            raise TypeError(str(var) + " should be Variable or str")

224
    fetch_var_names = list(map(to_name_str, fetch_list))
Q
qiaolongfei 已提交
225 226 227 228

    return str(feed_var_names + fetch_var_names)


W
Wu Yi 已提交
229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
def _as_lodtensor(data, place):
    """
        Convert numpy.ndarray to Tensor, its only support Tensor without LoD information.
        For higher dimensional sequence data, please use LoDTensor directly.

        Examples:
            >>> import paddle.fluid as fluid
            >>> place = fluid.CPUPlace()
            >>> exe = fluid.executor(place)
            >>> data = np.array(size=(100, 200, 300))
            >>> np_outs = map(lambda x: fluid.executor._as_lodtensor(x, place), data)
            >>>     ...

        Args:
            data(numpy.ndarray): a instance of array

        Returns:
            LoDTensor
        """
    if isinstance(data, list):
        raise RuntimeError("Some of your feed data hold LoD information. \
                They can not be completely cast from a list of Python \
                ndarray to LoDTensor. Please convert data to LoDTensor \
                directly before feeding the data.\
                ")
    # single tensor case
    tensor = core.LoDTensor()
    tensor.set(data, place)
    return tensor


Y
Yu Yang 已提交
260
class Executor(object):
261 262 263 264 265 266 267
    """
    An Executor in Python, only support the single-GPU running. For multi-cards, please refer to
    ParallelExecutor.
    Python executor takes a program, add feed operators and fetch operators to this program according
    to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides
    the variables(or names) that user want to get after program run. Note: the executor will run all
    operators in the program but not only the operators dependent by the fetch_list.
268 269
    It store the global variables into the global scope, and create a local scope for the temporary
    variables. The local scope contents will be discarded after every minibatch forward/backward finished.
270 271 272 273 274 275 276 277 278 279
    But the global scope variables will be persistent through different runs.
    All of ops in program will be running in sequence.

    Args:
        place(core.CPUPlace|core.CUDAPlace(n)): indicate the executor run on which device

    Note: For debugging complicated network in parallel-GPUs, you can test it on the executor.
    They has the exactly same arguments, and expected the same results.
    """

D
dzhwinter 已提交
280 281
    def __init__(self, place):
        self.place = place
Q
qiaolongfei 已提交
282
        self.program_caches = dict()
283
        self.executor = None
Y
Yancey1989 已提交
284
        self._closed = False
D
dzhwinter 已提交
285

Q
Qiao Longfei 已提交
286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317
    def _get_program_cache(self, program_cache_key):
        return self.program_caches.get(program_cache_key, None)

    def _add_program_cache(self, program_cache_key, program):
        self.program_caches[program_cache_key] = program

    def _add_feed_fetch_ops(self, program, feed, fetch_list, feed_var_name,
                            fetch_var_name):
        tmp_program = program.clone()

        global_block = tmp_program.global_block()

        if feed_var_name in global_block.vars:
            feed_var = global_block.var(feed_var_name)
        else:
            feed_var = global_block.create_var(
                name=feed_var_name,
                type=core.VarDesc.VarType.FEED_MINIBATCH,
                persistable=True)

        if fetch_var_name in global_block.vars:
            fetch_var = global_block.var(fetch_var_name)
        else:
            fetch_var = global_block.create_var(
                name=fetch_var_name,
                type=core.VarDesc.VarType.FETCH_LIST,
                persistable=True)

        # prepend feed operators
        if not has_feed_operators(global_block, feed, feed_var_name):
            for i, name in enumerate(feed):
                out = global_block.var(name)
W
Wu Yi 已提交
318
                global_block._prepend_op(
Q
Qiao Longfei 已提交
319 320 321 322 323 324 325 326
                    type='feed',
                    inputs={'X': [feed_var]},
                    outputs={'Out': [out]},
                    attrs={'col': i})

        # append fetch_operators
        if not has_fetch_operators(global_block, fetch_list, fetch_var_name):
            for i, var in enumerate(fetch_list):
M
minqiyang 已提交
327 328 329
                assert isinstance(var, Variable) or isinstance(
                    var, six.string_types), (
                        "Wrong type for fetch_list[%s]: %s" % (i, type(var)))
Q
Qiao Longfei 已提交
330 331 332 333 334 335 336 337 338 339 340 341 342 343 344
                global_block.append_op(
                    type='fetch',
                    inputs={'X': [var]},
                    outputs={'Out': [fetch_var]},
                    attrs={'col': i})

        return tmp_program

    def _feed_data(self, program, feed, feed_var_name, scope):
        # feed var to framework
        for op in program.global_block().ops:
            if op.desc.type() == 'feed':
                feed_target_name = op.desc.output('Out')[0]
                cur_feed = feed[feed_target_name]
                if not isinstance(cur_feed, core.LoDTensor):
W
Wu Yi 已提交
345
                    cur_feed = _as_lodtensor(cur_feed, self.place)
Q
Qiao Longfei 已提交
346 347 348 349 350 351 352 353
                idx = op.desc.attr('col')
                core.set_feed_variable(scope, cur_feed, feed_var_name, idx)
            else:
                break

    def _fetch_data(self, fetch_list, fetch_var_name, scope):
        outs = [
            core.get_fetch_variable(scope, fetch_var_name, i)
M
minqiyang 已提交
354
            for i in six.moves.range(len(fetch_list))
Q
Qiao Longfei 已提交
355 356 357
        ]
        return outs

Y
Yancey1989 已提交
358 359 360 361 362 363 364
    def close(self):
        """
        Close this executor.

        You can no long use this executor after calling this method.
        For the distributed training, this method would free the resource on PServers related to
        the current Trainer.
365
        TODO(panyx0718): Why ParallelExecutor doesn't have close?
Y
Yancey1989 已提交
366 367 368 369 370 371 372

        Example:
            >>> cpu = core.CPUPlace()
            >>> exe = Executor(cpu)
            >>> ...
            >>> exe.close()
        """
373
        if not self._closed and self.executor:
Y
Yancey1989 已提交
374 375
            self.executor.close()
            self._closed = True
Y
Yancey1989 已提交
376

377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
    def _run_parallel(self,
                      exe,
                      scope,
                      feed=None,
                      fetch_list=None,
                      return_numpy=True):
        if isinstance(feed, dict):
            feed_tensor_dict = dict()
            for feed_name in feed:
                feed_tensor = feed[feed_name]
                if not isinstance(feed_tensor, core.LoDTensor):
                    feed_tensor = core.LoDTensor()
                    # always set to CPU place, since the tensor need to be splitted
                    # it is fast in CPU
                    feed_tensor.set(feed[feed_name], core.CPUPlace())
                feed_tensor_dict[feed_name] = feed_tensor

            exe.feed_and_split_tensor_into_local_scopes(feed_tensor_dict)
        elif isinstance(feed, list) or isinstance(feed, tuple):
            if len(feed) != len(self._places):
                raise ValueError(
                    "Feed a list of tensor, the list should be the same size as places"
                )

            res = list()
            for i, each in enumerate(feed):
                if not isinstance(each, dict):
                    raise TypeError(
                        "Each element of feed list should be a dict")
                res_dict = dict()
                for feed_name in each:
                    tensor = each[feed_name]
                    if not isinstance(tensor, core.LoDTensor):
                        tmp = core.LoDTensor()
                        tmp.set(tensor, self._places[i])
                        tensor = tmp
                    res_dict[feed_name] = tensor
                res.append(res_dict)
            exe.feed_tensors_into_local_scopes(res)

        fetch_var_name = '@FETCHED_VAR_NAME@'
        exe.run(fetch_list, fetch_var_name)
        arr = scope.find_var(fetch_var_name).get_lod_tensor_array()

        if return_numpy:
            return as_numpy(arr)
        return [arr[i] for i in range(len(arr))]

Y
Yu Yang 已提交
425
    def run(self,
Y
Yu Yang 已提交
426
            program=None,
427 428
            feed=None,
            fetch_list=None,
Y
Yu Yang 已提交
429
            feed_var_name='feed',
Y
Yu Yang 已提交
430
            fetch_var_name='fetch',
D
dzhwinter 已提交
431
            scope=None,
432 433
            return_numpy=True,
            use_program_cache=False):
434 435
        """
        Run program by this Executor. Feed data by feed map, fetch result by fetch_list.
Q
qiaolongfei 已提交
436 437
        Python executor takes a program, add feed operators and fetch operators to this program according
        to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides
438 439 440
        the variables(or names) that user want to get after program run.

        Note: the executor will run all
Q
qiaolongfei 已提交
441 442
        operators in the program but not only the operators dependent by the fetch_list

443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459
        Args:
            program(Program): the program that need to run, if not provied, then default_main_program will be used.
            feed(dict): feed variable map, e.g. {"image": ImageData, "label": LableData}
            fetch_list(list): a list of variable or variable names that user want to get, run will return them according to this list.
            feed_var_name(str): the name for the input variable of feed Operator.
            fetch_var_name(str): the name for the output variable of fetch Operator.
            scope(Scope): the scope used to run this program, you can switch it to different scope. default is global_scope
            return_numpy(bool): if convert the fetched tensor to numpy
            use_program_cache(bool): set use_program_cache to true if program not changed compare to the last step.

        Returns:

            list(numpy.array): fetch result according to fetch_list.


        Examples:

T
Tink_Y 已提交
460 461 462 463 464
            >>> data = fluid.layers.data(name='X', shape=[1], dtype='float32')
            >>> out = fluid.layers.create_tensor(dtype='float32')
            >>> hidden = fluid.layers.fc(input=data, size=10)
            >>> fluid.layers.assign(hidden,out)
            >>> loss = fluid.layers.mean(out)
465
            >>> adam = fluid.optimizer.Adam()
T
Tink_Y 已提交
466
						>>> adam.minimize(loss)
467 468

            >>> cpu = core.CPUPlace()
T
Tink_Y 已提交
469 470
            >>> exe = fluid.Executor(cpu)
            >>> exe.run(fluid.default_startup_program())
471 472 473 474 475

            >>> x = numpy.random.random(size=(10, 1)).astype('float32')
            >>> outs = exe.run(
            >>>     feed={'X': x},
            >>>     fetch_list=[loss.name])
476
        """
Y
Yancey1989 已提交
477 478 479 480

        if self._closed:
            raise RuntimeError("Attempted to use a closed Executor")

481 482 483
        if scope is None:
            scope = global_scope()

X
polish  
Xin Pan 已提交
484 485
        compiled = isinstance(program, compiler.CompiledProgram)
        # For backward compatibility, run directly.
486
        if not compiled:
X
polish  
Xin Pan 已提交
487 488 489 490
            if not self.executor:
                p = core.Place()
                p.set_place(self.place)
                self.executor = core.Executor(p)
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
            return self._run(
                program,
                feed=feed,
                fetch_list=fetch_list,
                feed_var_name=feed_var_name,
                fetch_var_name=fetch_var_name,
                scope=scope,
                return_numpy=return_numpy,
                use_program_cache=use_program_cache)

        program._compile(scope, self.place)
        self.executor = program._executor
        if program._is_data_parallel:
            return self._run_parallel(
                exe=program._executor,
                scope=scope,
                feed=feed,
                fetch_list=fetch_list,
                return_numpy=return_numpy)
        else:
            return self._run(
                program._program,
                feed=feed,
                fetch_list=fetch_list,
                feed_var_name=feed_var_name,
                fetch_var_name=fetch_var_name,
                scope=scope,
                return_numpy=return_numpy,
                use_program_cache=use_program_cache)

    def _run(self, program, feed, fetch_list, feed_var_name, fetch_var_name,
             scope, return_numpy, use_program_cache):

524 525
        if feed is None:
            feed = {}
Q
qiaolongfei 已提交
526
        if not isinstance(feed, dict):
D
dzhwinter 已提交
527 528 529
            raise TypeError(
                "feed requires dict as its Parameter. But you passed in %s" %
                (type(feed)))
530 531
        if fetch_list is None:
            fetch_list = []
Y
Yu Yang 已提交
532
        if program is None:
Y
Yu Yang 已提交
533
            program = default_main_program()
Y
Yu Yang 已提交
534

Y
Yu Yang 已提交
535
        if not isinstance(program, Program):
D
dzhwinter 已提交
536 537 538
            raise TypeError(
                "Executor requires Program as its Parameter. But you passed in %s"
                % (type(program)))
Y
Yu Yang 已提交
539

540
        cache_key = _get_program_cache_key(feed, fetch_list)
541
        if use_program_cache:
Q
Qiao Longfei 已提交
542 543 544 545 546 547 548 549 550 551
            cached_program = self._get_program_cache(cache_key)
            if cached_program is None:
                cached_program = self._add_feed_fetch_ops(
                    program=program,
                    feed=feed,
                    fetch_list=fetch_list,
                    feed_var_name=feed_var_name,
                    fetch_var_name=fetch_var_name)
                self._add_program_cache(cache_key, cached_program)
            program = cached_program
552
        else:
Q
Qiao Longfei 已提交
553 554 555 556 557 558 559 560 561 562 563
            self.program_caches.pop(cache_key, None)
            program = self._add_feed_fetch_ops(
                program=program,
                feed=feed,
                fetch_list=fetch_list,
                feed_var_name=feed_var_name,
                fetch_var_name=fetch_var_name)

        self._feed_data(program, feed, feed_var_name, scope)
        self.executor.run(program.desc, scope, 0, True, True)
        outs = self._fetch_data(fetch_list, fetch_var_name, scope)
D
dzhwinter 已提交
564 565 566
        if return_numpy:
            outs = as_numpy(outs)
        return outs