executor.py 17.3 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.

D
dzhwinter 已提交
15
import numpy as np
Y
Yang Yu 已提交
16
import contextlib
17
import six
18
from .framework import Program, default_main_program, Variable
19 20
from . import core

X
xuwei06 已提交
21
__all__ = [
22
    'Executor', 'global_scope', 'scope_guard', '_switch_scope', 'fetch_var'
X
xuwei06 已提交
23
]
Y
Yu Yang 已提交
24

Y
Yu Yang 已提交
25 26
g_scope = core.Scope()

Y
Yu Yang 已提交
27

Y
Yang Yu 已提交
28
def global_scope():
Y
yuyang18 已提交
29 30 31 32 33 34 35
    """
    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
36 37 38
    return g_scope


39
def _switch_scope(scope):
Y
Yang Yu 已提交
40 41 42 43 44 45 46 47
    global g_scope
    ex = g_scope
    g_scope = scope
    return ex


@contextlib.contextmanager
def scope_guard(scope):
Y
yuyang18 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60
    """
    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.
    """
61
    ex = _switch_scope(scope)
Y
Yang Yu 已提交
62
    yield
63
    _switch_scope(ex)
Y
Yang Yu 已提交
64 65


D
dzhwinter 已提交
66
def as_numpy(tensor):
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
    """
    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
82 83
    if isinstance(tensor, core.LoDTensorArray):
        return [as_numpy(t) for t in tensor]
D
dzhwinter 已提交
84 85 86 87
    if isinstance(tensor, list):
        return [as_numpy(t) for t in tensor]
    assert isinstance(tensor, core.LoDTensor)
    lod = tensor.lod()
88
    if len(lod) > 0:
D
dzhwinter 已提交
89
        raise RuntimeError("Some of your fetched tensors hold LoD information. \
90 91 92 93
            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
94 95


96 97 98 99 100 101 102 103 104 105 106 107
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
108 109
        feed_holder_name: the name of the variable that holds the data of
            all feed targets. The type of this feed_holder variable is
110 111 112
            FEED_MINIBATCH, which is essentially vector<LoDTensor>.

    Returns:
X
xuwei06 已提交
113
        A boolean value that indicates whether a block has feed operators
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
        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
136

137 138 139 140 141 142 143 144 145
    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
146 147 148
        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>.
149

X
xuwei06 已提交
150 151 152
    Return:
        A boolean value that indicates whether a block has fetch operators
        that match the info contained in fetch_targets and fetch_holder_name.
153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173
    """

    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


X
xuwei06 已提交
174 175
def fetch_var(name, scope=None, return_numpy=True):
    """
C
chengduoZH 已提交
176 177 178
    Fetch the value of the variable with the given name from the
    given scope.

X
xuwei06 已提交
179
    Args:
180 181 182 183
        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
184 185 186 187
            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
188 189 190 191 192 193 194 195
    Returns:
       LodTensor|numpy.ndarray
    """
    assert isinstance(name, str)
    if scope is None:
        scope = global_scope()
    assert isinstance(scope, core.Scope)

Y
Yibing Liu 已提交
196
    var = scope.find_var(name)
197 198 199 200
    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
201 202 203 204 205 206
    tensor = var.get_tensor()
    if return_numpy:
        tensor = as_numpy(tensor)
    return tensor


207
def _get_program_cache_key(feed, fetch_list):
208
    feed_var_names = list(feed.keys())
Q
qiaolongfei 已提交
209 210 211 212 213 214

    def to_name_str(var):
        if isinstance(var, Variable):
            return var.desc.name()
        elif isinstance(var, str):
            return var
215
        elif isinstance(var, six.string_types):
216
            return str(var)
Q
qiaolongfei 已提交
217 218 219
        else:
            raise TypeError(str(var) + " should be Variable or str")

220
    fetch_var_names = list(map(to_name_str, fetch_list))
Q
qiaolongfei 已提交
221 222 223 224

    return str(feed_var_names + fetch_var_names)


Y
Yu Yang 已提交
225
class Executor(object):
226 227 228 229 230 231 232
    """
    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.
233 234
    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.
235 236 237 238 239 240 241 242 243 244
    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
245 246 247 248 249
    def __init__(self, place):
        self.place = place
        p = core.Place()
        p.set_place(place)
        self.executor = core.Executor(p)
Q
qiaolongfei 已提交
250
        self.program_caches = dict()
Y
Yancey1989 已提交
251
        self._closed = False
D
dzhwinter 已提交
252

D
dzhwinter 已提交
253
    def as_lodtensor(self, data):
254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270
        """
        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
            >>> exe = fluid.executor(fluid.CPUPlace())
            >>> data = np.array(size=(100, 200, 300))
            >>> np_outs = map(lambda x: exe.as_lodtensor(x), data)
            >>>     ...

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

        Returns:
            LoDTensor
        """
D
dzhwinter 已提交
271 272 273 274 275 276 277 278 279 280
        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, self.place)
        return tensor
Y
Yu Yang 已提交
281

Q
Qiao Longfei 已提交
282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313
    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
314
                global_block._prepend_op(
Q
Qiao Longfei 已提交
315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339
                    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):
                assert isinstance(var, Variable) or isinstance(var, str), (
                    "Wrong type for fetch_list[%s]: %s" % (i, type(var)))
                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):
D
dzhwinter 已提交
340
                    cur_feed = self.as_lodtensor(cur_feed)
Q
Qiao Longfei 已提交
341 342 343 344 345 346 347 348
                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)
349
            for i in range(len(fetch_list))
Q
Qiao Longfei 已提交
350 351 352
        ]
        return outs

Y
Yancey1989 已提交
353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369
    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.

        Example:
            >>> cpu = core.CPUPlace()
            >>> exe = Executor(cpu)
            >>> ...
            >>> exe.close()
        """
        if not self._closed:
            self.executor.close()
            self._closed = True
Y
Yancey1989 已提交
370

Y
Yu Yang 已提交
371
    def run(self,
Y
Yu Yang 已提交
372
            program=None,
373 374
            feed=None,
            fetch_list=None,
Y
Yu Yang 已提交
375
            feed_var_name='feed',
Y
Yu Yang 已提交
376
            fetch_var_name='fetch',
D
dzhwinter 已提交
377
            scope=None,
378 379
            return_numpy=True,
            use_program_cache=False):
380 381
        """
        Run program by this Executor. Feed data by feed map, fetch result by fetch_list.
Q
qiaolongfei 已提交
382 383
        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
384 385 386
        the variables(or names) that user want to get after program run.

        Note: the executor will run all
Q
qiaolongfei 已提交
387 388
        operators in the program but not only the operators dependent by the fetch_list

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
        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:

            >>> data = layers.data(name='X', shape=[1], dtype='float32')
            >>> hidden = layers.fc(input=data, size=10)
            >>> layers.assign(hidden, out)
            >>> loss = layers.mean(out)
            >>> adam = fluid.optimizer.Adam()
            >>> adam.minimize(loss)

            >>> cpu = core.CPUPlace()
            >>> exe = Executor(cpu)
            >>> exe.run(default_startup_program())

            >>> x = numpy.random.random(size=(10, 1)).astype('float32')
            >>> outs = exe.run(
            >>>     feed={'X': x},
            >>>     fetch_list=[loss.name])
421
        """
Y
Yancey1989 已提交
422 423 424 425

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

426 427
        if feed is None:
            feed = {}
Q
qiaolongfei 已提交
428
        if not isinstance(feed, dict):
D
dzhwinter 已提交
429 430 431
            raise TypeError(
                "feed requires dict as its Parameter. But you passed in %s" %
                (type(feed)))
432 433
        if fetch_list is None:
            fetch_list = []
Y
Yu Yang 已提交
434
        if program is None:
Y
Yu Yang 已提交
435
            program = default_main_program()
Y
Yu Yang 已提交
436

Y
Yu Yang 已提交
437
        if not isinstance(program, Program):
D
dzhwinter 已提交
438 439 440
            raise TypeError(
                "Executor requires Program as its Parameter. But you passed in %s"
                % (type(program)))
Y
Yu Yang 已提交
441

Y
Yu Yang 已提交
442
        if scope is None:
Y
Yang Yu 已提交
443
            scope = global_scope()
Y
Yu Yang 已提交
444

445
        cache_key = _get_program_cache_key(feed, fetch_list)
446
        if use_program_cache:
Q
Qiao Longfei 已提交
447 448 449 450 451 452 453 454 455 456
            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
457
        else:
Q
Qiao Longfei 已提交
458 459 460 461 462 463 464 465 466 467 468
            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
469 470 471
        if return_numpy:
            outs = as_numpy(outs)
        return outs