executor.py 39.2 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

Z
Zeng Jinle 已提交
17
import logging
18 19
import os
import multiprocessing
C
chengduo 已提交
20
import sys
21
import warnings
D
dzhwinter 已提交
22
import numpy as np
S
rename  
sneaxiy 已提交
23
from .wrapped_decorator import signature_safe_contextmanager
24
import six
25
from .framework import Program, default_main_program, Variable
26
from . import core
27 28
from . import compiler
from .. import compat as cpt
29
from .trainer_factory import TrainerFactory
30

T
Tink_Y 已提交
31
__all__ = ['Executor', 'global_scope', 'scope_guard']
Y
Yu Yang 已提交
32

Y
Yu Yang 已提交
33
g_scope = core.Scope()
F
flame 已提交
34 35
InferNativeConfig = core.NativeConfig
InferAnalysisConfig = core.AnalysisConfig
Y
Yu Yang 已提交
36

Y
Yu Yang 已提交
37

Y
Yang Yu 已提交
38
def global_scope():
Y
yuyang18 已提交
39 40 41 42
    """
    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`

43 44 45 46 47 48 49 50 51
    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy

          fluid.global_scope().var("data").get_tensor().set(numpy.ones((2, 2)), fluid.CPUPlace())
          numpy.array(fluid.global_scope().find_var("data").get_tensor())

Y
yuyang18 已提交
52 53 54
    Returns:
        Scope: The global/default scope instance.
    """
Y
Yang Yu 已提交
55 56 57
    return g_scope


58
def _switch_scope(scope):
Y
Yang Yu 已提交
59 60 61 62 63 64
    global g_scope
    ex = g_scope
    g_scope = scope
    return ex


S
rename  
sneaxiy 已提交
65
@signature_safe_contextmanager
Y
Yang Yu 已提交
66
def scope_guard(scope):
Y
yuyang18 已提交
67 68 69 70
    """
    Change the global/default scope instance by Python `with` statement. All
    variable in runtime will assigned to the new scope.

L
lujun 已提交
71 72 73
    Args:
        scope: The new global/default scope.

Y
yuyang18 已提交
74
    Examples:
75 76
        .. code-block:: python

77
            import paddle.fluid as fluid
L
lujun 已提交
78
            import numpy
Y
yuyang18 已提交
79

L
lujun 已提交
80 81 82 83
            new_scope = fluid.Scope()
            with fluid.scope_guard(new_scope):
                 fluid.global_scope().var("data").get_tensor().set(numpy.ones((2, 2)), fluid.CPUPlace())
            numpy.array(new_scope.find_var("data").get_tensor())
Y
yuyang18 已提交
84
    """
L
lujun 已提交
85

86
    ex = _switch_scope(scope)
Y
Yang Yu 已提交
87
    yield
88
    _switch_scope(ex)
Y
Yang Yu 已提交
89 90


D
dzhwinter 已提交
91
def as_numpy(tensor):
92 93 94
    """
    Convert a Tensor to a numpy.ndarray, its only support Tensor without LoD information.
    For higher dimensional sequence data, please use LoDTensor directly.
95

96
    Examples:
97 98 99 100 101 102 103 104 105 106
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy

          new_scope = fluid.Scope()
          with fluid.scope_guard(new_scope):
              fluid.global_scope().var("data").get_tensor().set(numpy.ones((2, 2)), fluid.CPUPlace())
          tensor = new_scope.find_var("data").get_tensor()
          fluid.executor.as_numpy(tensor) # or numpy.array(new_scope.find_var("data").get_tensor())
107 108 109 110 111 112 113

    Args:
       tensor(Variable): a instance of Tensor

    Returns:
        numpy.ndarray
    """
C
chengduo 已提交
114 115
    if isinstance(tensor, core.LoDTensorArray):
        return [as_numpy(t) for t in tensor]
D
dzhwinter 已提交
116 117 118 119
    if isinstance(tensor, list):
        return [as_numpy(t) for t in tensor]
    assert isinstance(tensor, core.LoDTensor)
    lod = tensor.lod()
120
    if len(lod) > 0:
D
dzhwinter 已提交
121
        raise RuntimeError("Some of your fetched tensors hold LoD information. \
122 123 124
            They can not be completely cast to Python ndarray. \
            Please set the parameter 'return_numpy' as 'False' to \
            return LoDTensor itself directly.")
Q
qingqing01 已提交
125 126 127 128
    if tensor._is_initialized():
        return np.array(tensor)
    else:
        return None
D
dzhwinter 已提交
129 130


131 132 133 134 135 136 137 138 139 140 141 142
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
143 144
        feed_holder_name: the name of the variable that holds the data of
            all feed targets. The type of this feed_holder variable is
145 146 147
            FEED_MINIBATCH, which is essentially vector<LoDTensor>.

    Returns:
X
xuwei06 已提交
148
        A boolean value that indicates whether a block has feed operators
149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
        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
171

172 173 174 175 176 177 178 179 180
    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
181 182 183
        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>.
184

X
xuwei06 已提交
185 186 187
    Return:
        A boolean value that indicates whether a block has fetch operators
        that match the info contained in fetch_targets and fetch_holder_name.
188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208
    """

    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
209
def _fetch_var(name, scope=None, return_numpy=True):
X
xuwei06 已提交
210
    """
C
chengduoZH 已提交
211 212 213
    Fetch the value of the variable with the given name from the
    given scope.

X
xuwei06 已提交
214
    Args:
215 216 217 218
        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
219 220 221 222
            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
223 224 225 226 227 228
    Returns:
       LodTensor|numpy.ndarray
    """
    assert isinstance(name, str)
    if scope is None:
        scope = global_scope()
S
sneaxiy 已提交
229
    assert isinstance(scope, core._Scope)
X
xuwei06 已提交
230

Y
Yibing Liu 已提交
231
    var = scope.find_var(name)
232 233 234 235
    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
236 237 238 239 240 241
    tensor = var.get_tensor()
    if return_numpy:
        tensor = as_numpy(tensor)
    return tensor


X
polish  
Xin Pan 已提交
242 243 244 245 246 247 248 249 250
def _to_name_str(var):
    if isinstance(var, Variable):
        return var.desc.name()
    elif isinstance(var, str):
        return var
    elif isinstance(var, six.string_types):
        return str(var)
    else:
        raise TypeError(str(var) + " should be Variable or str")
Q
qiaolongfei 已提交
251 252


253 254 255 256
def _get_strong_program_cache_key(program, feed, fetch_list):
    return str(id(program)) + _get_program_cache_key(feed, fetch_list)


X
polish  
Xin Pan 已提交
257 258 259
def _get_program_cache_key(feed, fetch_list):
    feed_var_names = list(feed.keys())
    fetch_var_names = list(map(_to_name_str, fetch_list))
Q
qiaolongfei 已提交
260 261 262 263

    return str(feed_var_names + fetch_var_names)


W
Wu Yi 已提交
264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
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
295
class Executor(object):
296
    """
297 298 299 300 301 302 303 304 305 306 307
    An Executor in Python, supports single/multiple-GPU running,
    and single/multiple-CPU running. Python executor takes a program,
    adds 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 wants
    to get after program runs. Note: the executor will run all operators
    in the program but not only the operators dependent by the fetch_list.
    It stores the global variables into the global scope, and creates a
    local scope for the temporary variables. The contents in local scope
    may be discarded after every minibatch forward/backward finished.
    But the global scope variables will be persistent through different runs.
S
Fix doc  
sneaxiy 已提交
308

309
    Examples:
S
Fix doc  
sneaxiy 已提交
310 311
        .. code-block:: python

312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
          import paddle.fluid as fluid
          import paddle.fluid.compiler as compiler
          import numpy
          import os

          use_cuda = True
          place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
          exe = fluid.Executor(place)

          train_program = fluid.Program()
          startup_program = fluid.Program()
          with fluid.program_guard(train_program, startup_program):
              data = fluid.layers.data(name='X', shape=[1], dtype='float32')
              hidden = fluid.layers.fc(input=data, size=10)
              loss = fluid.layers.mean(hidden)
              fluid.optimizer.SGD(learning_rate=0.01).minimize(loss)

          # Run the startup program once and only once.
          # Not need to optimize/compile the startup program.
          startup_program.random_seed=1
          exe.run(startup_program)

          # Run the main program directly without compile.
          x = numpy.random.random(size=(10, 1)).astype('float32')
          loss_data, = exe.run(train_program,
                               feed={"X": x},
                               fetch_list=[loss.name])

          # Or, compiled the program and run. See `CompiledProgram`
          # for more detail.
          # NOTE: If you use CPU to run the program, you need
          # to specify the CPU_NUM, otherwise, fluid will use
          # all the number of the logic core as the CPU_NUM,
          # in that case, the batch size of the input should be
          # greater than CPU_NUM, if not, the process will be
          # failed by an exception.
          if not use_cuda:
              os.environ['CPU_NUM'] = str(2)

          compiled_prog = compiler.CompiledProgram(
              train_program).with_data_parallel(
              loss_name=loss.name)
          loss_data, = exe.run(compiled_prog,
                               feed={"X": x},
                               fetch_list=[loss.name])
X
add doc  
Xin Pan 已提交
357

358
    Args:
359 360
        place(fluid.CPUPlace|fluid.CUDAPlace(n)): indicate the executor run on which device.

361 362
    """

D
dzhwinter 已提交
363 364
    def __init__(self, place):
        self.place = place
Q
qiaolongfei 已提交
365
        self.program_caches = dict()
366
        self.ctx_caches = dict()
367 368
        self.scope_caches = dict()
        self.var_caches = dict()
369 370 371
        p = core.Place()
        p.set_place(self.place)
        self._default_executor = core.Executor(p)
Y
Yancey1989 已提交
372
        self._closed = False
D
dzhwinter 已提交
373

374 375 376 377 378 379
    def _get_var_cache(self, program_cache_key):
        return self.var_caches.get(program_cache_key, None)

    def _get_scope_cache(self, program_cache_key):
        return self.scope_caches.get(program_cache_key, None)

380 381 382
    def _get_ctx_cache(self, program_cache_key):
        return self.ctx_caches.get(program_cache_key, None)

Q
Qiao Longfei 已提交
383 384 385 386 387 388
    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

389 390 391
    def _add_ctx_cache(self, ctx_cache_key, ctx):
        self.ctx_caches[ctx_cache_key] = ctx

392 393 394 395 396 397
    def _add_scope_cache(self, scope_cache_key, scope):
        self.scope_caches[scope_cache_key] = scope

    def _add_var_cache(self, var_cache_key, var):
        self.var_caches[var_cache_key] = var

Q
Qiao Longfei 已提交
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
    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
424
                global_block._prepend_op(
Q
Qiao Longfei 已提交
425 426 427 428 429 430 431 432
                    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
433 434 435
                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
436 437 438 439 440 441 442 443 444 445 446 447 448 449 450
                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
451
                    cur_feed = _as_lodtensor(cur_feed, self.place)
Q
Qiao Longfei 已提交
452 453 454 455 456 457 458 459
                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
460
            for i in six.moves.range(len(fetch_list))
Q
Qiao Longfei 已提交
461 462 463
        ]
        return outs

S
Fix doc  
sneaxiy 已提交
464 465 466 467 468 469
    '''
    TODO(typhoonzero): Define "no longer use" meaning? Can user create
    a new Executor for the same program and run?
    TODO(panyx0718): Why ParallelExecutor doesn't have close?
    '''

Y
Yancey1989 已提交
470 471 472 473
    def close(self):
        """
        Close this executor.

X
fix  
Xin Pan 已提交
474
        You can no longer use this executor after calling this method.
475 476 477 478 479 480 481 482 483 484 485 486
        For the distributed training, this method would free the resource
        on PServers related to the current Trainer.

        Examples:
            .. code-block:: python

              import paddle.fluid as fluid

              cpu = fluid.CPUPlace()
              exe = fluid.Executor(cpu)
              # execute training or testing
              exe.close()
Y
Yancey1989 已提交
487
        """
488 489
        if not self._closed:
            self._default_executor.close()
Y
Yancey1989 已提交
490
            self._closed = True
Y
Yancey1989 已提交
491

X
fix  
Xin Pan 已提交
492
    def _run_parallel(self, program, scope, feed, fetch_list, fetch_var_name,
X
polish  
Xin Pan 已提交
493
                      return_numpy):
494
        exe = program._executor
495 496 497 498 499 500
        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()
501
                    # always set to CPU place, since the tensor need to be split
502
                    # it is fast in CPU
503 504 505
                    assert isinstance( feed[feed_name], np.ndarray ), \
                        "The input({}) should be numpy.array, but not {}.".format(
                        feed_name, type(feed[feed_name]))
506 507 508
                    feed_tensor.set(feed[feed_name], core.CPUPlace())
                feed_tensor_dict[feed_name] = feed_tensor

509
            exe.feed_and_split_tensor_into_local_scopes(feed_tensor_dict)
510
        elif isinstance(feed, list) or isinstance(feed, tuple):
X
fix  
Xin Pan 已提交
511
            if len(feed) != len(program._places):
512 513 514 515 516 517 518 519 520 521 522 523 524 525
                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()
526 527 528
                        assert isinstance(each[feed_name], np.ndarray), \
                            "The input({}) should be numpy.array, but not {}.".format(
                            feed_name, type(each[feed_name]))
X
fix  
Xin Pan 已提交
529
                        tmp.set(tensor, program._places[i])
530 531 532
                        tensor = tmp
                    res_dict[feed_name] = tensor
                res.append(res_dict)
533
            exe.feed_tensors_into_local_scopes(res)
534

X
polish  
Xin Pan 已提交
535
        fetch_var_names = list(map(_to_name_str, fetch_list))
536
        tensors = exe.run(fetch_var_names)._move_to_list()
537
        return as_numpy(tensors) if return_numpy else tensors
538

Y
Yu Yang 已提交
539
    def run(self,
Y
Yu Yang 已提交
540
            program=None,
541 542
            feed=None,
            fetch_list=None,
Y
Yu Yang 已提交
543
            feed_var_name='feed',
Y
Yu Yang 已提交
544
            fetch_var_name='fetch',
D
dzhwinter 已提交
545
            scope=None,
546 547
            return_numpy=True,
            use_program_cache=False):
548
        """
549 550 551 552
        Run program by this Executor. Feed data by feed map, fetch result by
        fetch_list. 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
553 554
        the variables(or names) that user want to get after program run.

555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579
        Note: the executor will run all operators in the program but not
        only the operators dependent by the fetch_list.

        Examples:
            .. code-block:: python

              import paddle.fluid as fluid
              import numpy

              # First create the Executor.
              place = fluid.CPUPlace() # fluid.CUDAPlace(0)
              exe = fluid.Executor(place)

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

              # Run the startup program once and only once.
              exe.run(fluid.default_startup_program())

              x = numpy.random.random(size=(10, 1)).astype('float32')
              outs = exe.run(feed={'X': x},
                             fetch_list=[loss.name])
Q
qiaolongfei 已提交
580

581
        Args:
X
add doc  
Xin Pan 已提交
582
            program(Program|CompiledProgram): the program that need to run,
X
fix  
Xin Pan 已提交
583
                if not provided, then default_main_program (not compiled) will be used.
X
add doc  
Xin Pan 已提交
584
            feed(dict): feed variable map, e.g. {"image": ImageData, "label": LabelData}
Z
Zeng Jinle 已提交
585 586 587 588 589 590 591 592
            fetch_list(list): a list of variable or variable names that user 
                wants to get, this method 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
593
            return_numpy(bool): if convert the fetched tensor to numpy
Z
Zeng Jinle 已提交
594 595 596 597 598 599
            use_program_cache(bool): whether to use the cached program 
                settings across batches. Setting it be true would be faster 
                only when (1) the program is not compiled with data parallel, 
                and (2) program, feed variable names and fetch_list variable 
                names do not changed compared to the last step. 
                
600 601 602
        Returns:

            list(numpy.array): fetch result according to fetch_list.
603
        """
C
chengduo 已提交
604 605 606 607 608 609 610 611 612 613 614 615
        try:
            return self._run_impl(
                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)
        except Exception as e:
            if not isinstance(e, core.EOFException):
616
                print("!!!A non-EOF exception is thrown.")
617
            six.reraise(*sys.exc_info())
C
chengduo 已提交
618 619 620

    def _run_impl(self, program, feed, fetch_list, feed_var_name,
                  fetch_var_name, scope, return_numpy, use_program_cache):
Y
Yancey1989 已提交
621 622 623
        if self._closed:
            raise RuntimeError("Attempted to use a closed Executor")

C
chengduo 已提交
624
        use_default_main_program = program is None
625 626
        if program is None:
            program = default_main_program()
C
chengduo 已提交
627
        if isinstance(program, Program) and \
628
                        len(program.global_block().ops) == 0:
C
chengduo 已提交
629 630 631 632
            error_info = "The current program is empty."
            if use_default_main_program:
                error_info += " Maybe you should pass the Program or the CompiledProgram manually."
            warnings.warn(error_info)
633

634 635
        if scope is None:
            scope = global_scope()
636 637 638 639 640 641 642 643 644

        if fetch_list is not None:
            if isinstance(fetch_list, Variable) or isinstance(fetch_list, str):
                fetch_list = [fetch_list]
            assert isinstance(fetch_list, tuple) or isinstance(fetch_list, list), \
                "Currently , The fetch_list type only should be list or tuple, \n"\
                "but the input type is {}. For more information please refer to \n"\
                "the executor.run(...).".format(type(fetch_list))
        else:
X
polish  
Xin Pan 已提交
645
            fetch_list = []
646

X
polish  
Xin Pan 已提交
647 648
        compiled = isinstance(program, compiler.CompiledProgram)
        # For backward compatibility, run directly.
649
        if not compiled:
C
chengduo 已提交
650
            return self._run_program(
651 652 653 654 655 656 657 658 659 660
                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)
C
chengduo 已提交
661 662 663
        if program._is_inference:
            return self._run_inference(program._executor, feed)
        else:
664
            return self._run_parallel(
X
fix  
Xin Pan 已提交
665
                program,
666 667 668
                scope=scope,
                feed=feed,
                fetch_list=fetch_list,
X
polish  
Xin Pan 已提交
669
                fetch_var_name=fetch_var_name,
670 671
                return_numpy=return_numpy)

C
chengduo 已提交
672
    def _run_program(self, program, feed, fetch_list, feed_var_name,
C
chengduo 已提交
673
                     fetch_var_name, scope, return_numpy, use_program_cache):
674

675 676
        if feed is None:
            feed = {}
S
sneaxiy 已提交
677 678 679 680
        elif isinstance(feed, (list, tuple)):
            assert len(feed) == 1, "Not compiled with data parallel"
            feed = feed[0]

Q
qiaolongfei 已提交
681
        if not isinstance(feed, dict):
D
dzhwinter 已提交
682 683 684
            raise TypeError(
                "feed requires dict as its Parameter. But you passed in %s" %
                (type(feed)))
Y
Yu Yang 已提交
685

686
        assert program is not None, "The program should not be Empty"
Y
Yu Yang 已提交
687
        if not isinstance(program, Program):
D
dzhwinter 已提交
688 689 690
            raise TypeError(
                "Executor requires Program as its Parameter. But you passed in %s"
                % (type(program)))
Y
Yu Yang 已提交
691

692
        if use_program_cache:
693
            cache_key = _get_strong_program_cache_key(program, feed, fetch_list)
Q
Qiao Longfei 已提交
694
            cached_program = self._get_program_cache(cache_key)
695
            cached_ctx = self._get_ctx_cache(cache_key)
696 697
            cached_scope = self._get_scope_cache(cache_key)
            cached_var = self._get_var_cache(cache_key)
Q
Qiao Longfei 已提交
698 699 700 701 702 703 704 705
            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)
706
                fetch_list_str = list(map(_to_name_str, fetch_list))
707
                cached_ctx = self._default_executor.prepare_ctx_cache(
708 709 710 711 712 713 714 715 716
                    cached_program.desc, 0, fetch_list_str, False)
                cached_var = self._default_executor.create_variables(
                    cached_program.desc, scope, 0)
                # currently, we cache program, vars, sub_scope here
                # we suppose that in a life cycle of training, a user
                # will not create many programs. So, here the basic
                # rule of caching is to cache all unseen (program, var, scope)
                # when a user use use_program_cache.
                cached_scope = scope.new_scope()
717
                self._add_ctx_cache(cache_key, cached_ctx)
718 719
                self._add_var_cache(cache_key, cached_var)
                self._add_scope_cache(cache_key, cached_scope)
Q
Qiao Longfei 已提交
720
            program = cached_program
721
            ctx = cached_ctx
722 723
            scope = cached_scope
            var = cached_var
724
        else:
Q
Qiao Longfei 已提交
725 726 727 728 729 730 731 732
            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)
733
        if not use_program_cache:
C
chengduo 已提交
734 735
            self._default_executor.run(program.desc, scope, 0, True, True,
                                       fetch_var_name)
736
        else:
C
chengduo 已提交
737 738
            self._default_executor.run_cached_prepared_ctx(ctx, scope, False,
                                                           False, False)
739 740
        arr = scope.find_var(fetch_var_name).get_lod_tensor_array()
        tensors = arr._move_to_list()
D
dzhwinter 已提交
741
        if return_numpy:
742 743 744
            return as_numpy(tensors)
        else:
            return tensors
F
flame 已提交
745

X
Xin Pan 已提交
746 747
    def _run_inference(self, exe, feed):
        return exe.run(feed)
D
dongdaxiang 已提交
748

749 750
    def _dump_debug_info(self, program=None, trainer=None):
        with open(str(id(program)) + "_train_desc.prototxt", "w") as fout:
H
hutuxian 已提交
751
            fout.write(str(trainer))
752 753 754 755
        if program._fleet_opt:
            with open("fleet_desc.prototxt", "w") as fout:
                fout.write(str(program._fleet_opt["fleet_desc"]))

756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771
    def _adjust_pipeline_resource(self, pipeline_opt, dataset, pipeline_num):
        filelist_length = len(dataset.dataset.get_filelist())
        if filelist_length < pipeline_num:
            pipeline_num = filelist_length
            print(
                "Pipeline training: setting the pipeline num to %d is enough because there are only %d files"
                % (filelist_length, filelist_length))
        if filelist_length < pipeline_num * pipeline_opt["concurrency_list"][0]:
            print(
                "Pipeline training: setting the 1st element in concurrency_list to %d is enough because there are only %d files"
                % (filelist_length // pipeline_num, filelist_length))
            pipeline_opt["concurrency_list"][
                0] = filelist_length // pipeline_num
        dataset.set_thread(pipeline_opt["concurrency_list"][0] * pipeline_num)
        return pipeline_num

772 773 774 775 776 777 778 779 780
    def _prepare_trainer(self,
                         program=None,
                         dataset=None,
                         scope=None,
                         thread=0,
                         debug=False,
                         fetch_list=None,
                         fetch_info=None,
                         print_period=100):
D
dongdaxiang 已提交
781 782 783 784
        if scope is None:
            scope = global_scope()
        if fetch_list is None:
            fetch_list = []
D
dongdaxiang 已提交
785 786 787
        if fetch_info is None:
            fetch_info = []
        assert len(fetch_list) == len(fetch_info)
D
dongdaxiang 已提交
788 789
        compiled = isinstance(program, compiler.CompiledProgram)
        if not compiled:
H
hutuxian 已提交
790 791 792 793 794 795
            # TODO: Need a better way to distinguish and specify different execution mode
            if program._pipeline_opt:
                trainer = TrainerFactory()._create_trainer(
                    program._pipeline_opt)
            else:
                trainer = TrainerFactory()._create_trainer(program._fleet_opt)
796
            trainer._set_program(program)
797
        else:
H
hutuxian 已提交
798 799 800 801 802 803
            if program._pipeline_opt:
                trainer = TrainerFactory()._create_trainer(
                    program.program._pipeline_opt)
            else:
                trainer = TrainerFactory()._create_trainer(
                    program.program._fleet_opt)
804
            trainer._set_program(program.program)
H
hutuxian 已提交
805

806
        if thread <= 0:
D
dongdaxiang 已提交
807 808
            if dataset.thread_num <= 0:
                raise RuntimeError(
809 810
                    "You should set thread num first, either in Dataset"
                    "or in Executor.train_from_dataset")
D
dongdaxiang 已提交
811
            else:
812
                trainer._set_thread(dataset.thread_num)
813
        else:
814
            trainer._set_thread(thread)
H
hutuxian 已提交
815

816 817
        trainer._set_debug(debug)
        trainer._set_fetch_var_and_info(fetch_list, fetch_info, print_period)
818
        return scope, trainer
819 820 821 822 823 824

    def infer_from_dataset(self,
                           program=None,
                           dataset=None,
                           scope=None,
                           thread=0,
825 826 827 828
                           debug=False,
                           fetch_list=None,
                           fetch_info=None,
                           print_period=100):
829 830 831 832 833 834
        """
        The document of infer_from_dataset is almost the same as
        train_from_dataset, except that in distributed training,
        push gradients will be disabled in infer_from_dataset.
        infer_from_dataset() can be used for evaluation in multi-thread
        very easily.
835

836 837 838 839 840
        Args:
            program(Program|CompiledProgram): the program that needs to be run,
               if not provided, then default_main_program (not compiled) will be used.
            dataset(paddle.fluid.Dataset): dataset created outside this function,
               a user should provide a well-defined dataset before calling this function.
841
               Please check the document of Dataset if needed. default is None
842 843 844
            scope(Scope): the scope used to run this program, you can switch it to different scope
               for each run. default is global_scope
            thread(int): number of thread a user wants to run in this function. The actual number
845 846
               of thread will be min(Dataset.thread_num, thread) if thread > 0, default is 0
            debug(bool): whether a user wants to run infer_from_dataset, default is False
847
            fetch_list(Variable List): fetch variable list, each variable
848 849 850
                                       will be printed during training, default is None
            fetch_info(String List): print information for each variable, default is None
            print_period(int): the number of mini-batches for each print, default is 100
851

852 853 854 855
        Returns:
            None

        Examples:
856 857

            .. code-block:: python
858

859
                import paddle.fluid as fluid
860 861

                place = fluid.CPUPlace() # you can set place = fluid.CUDAPlace(0) to use gpu
862
                exe = fluid.Executor(place)
863 864
                x = fluid.layers.data(name="x", shape=[10, 10], dtype="int64")
                y = fluid.layers.data(name="y", shape=[1], dtype="int64", lod_level=1)
865 866
                dataset = fluid.DatasetFactory().create_dataset()
                dataset.set_use_var([x, y])
867 868
                dataset.set_thread(1)
                filelist = [] # you should set your own filelist, e.g. filelist = ["dataA.txt"]
869 870 871 872
                dataset.set_filelist(filelist)
                exe.run(fluid.default_startup_program())
                exe.infer_from_dataset(program=fluid.default_main_program(),
                                       dataset=dataset)        
873

874
        """
875 876 877
        if dataset == None:
            raise RuntimeError("dataset is needed and should be initialized")

J
jiaqi 已提交
878
        dataset._prepare_to_run()
879
        scope, trainer = self._prepare_trainer(
880 881 882 883 884 885 886 887
            program=program,
            dataset=dataset,
            scope=scope,
            thread=thread,
            debug=debug,
            fetch_list=fetch_list,
            fetch_info=fetch_info,
            print_period=print_period)
888
        trainer._set_infer(True)
889
        trainer._gen_trainer_desc()
890
        self._dump_debug_info(program=program, trainer=trainer)
891
        dataset._dynamic_adjust_before_train(trainer.proto_desc.thread_num)
892 893 894
        self._default_executor.run_from_dataset(program.desc, scope,
                                                dataset.dataset,
                                                trainer._desc())
895
        dataset._dynamic_adjust_after_train()
J
jiaqi 已提交
896
        dataset._finish_to_run()
897
        return None
898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933

    def train_from_dataset(self,
                           program=None,
                           dataset=None,
                           scope=None,
                           thread=0,
                           debug=False,
                           fetch_list=None,
                           fetch_info=None,
                           print_period=100):
        """
        Train from a pre-defined Dataset. Dataset is defined in paddle.fluid.dataset.
        Given a program, either a program or compiled program, train_from_dataset will
        consume all data samples in dataset. Input scope can be given by users. By default,
        scope is global_scope(). The total number of thread run in training is `thread`.
        Thread number used in training will be minimum value of threadnum in Dataset and
        the value of thread in this interface. Debug can be set so that executor will display
        Run-Time for all operators and the throughputs of current training task.
        
        Note: train_from_dataset will destroy all resources created within executor for each run.

        Args:
            program(Program|CompiledProgram): the program that needs to be run,
               if not provided, then default_main_program (not compiled) will be used.
            dataset(paddle.fluid.Dataset): dataset created outside this function,
               a user should provide a well-defined dataset before calling this function.
               Please check the document of Dataset if needed.
            scope(Scope): the scope used to run this program, you can switch it to different scope
               for each run. default is global_scope
            thread(int): number of thread a user wants to run in this function. The actual number
               of thread will be min(Dataset.thread_num, thread)
            debug(bool): whether a user wants to run train_from_dataset 
            fetch_list(Variable List): fetch variable list, each variable
                                       will be printed during training
            fetch_info(String List): print information for each variable
            print_period(int): the number of mini-batches for each print
934 935 936

        Returns:
            None
937
        
938
        Examples:
939
        
940 941 942
            .. code-block:: python

              import paddle.fluid as fluid
943 944

              place = fluid.CPUPlace() # you can set place = fluid.CUDAPlace(0) to use gpu
945
              exe = fluid.Executor(place)
946 947
              x = fluid.layers.data(name="x", shape=[10, 10], dtype="int64")
              y = fluid.layers.data(name="y", shape=[1], dtype="int64", lod_level=1)
948 949
              dataset = fluid.DatasetFactory().create_dataset()
              dataset.set_use_var([x, y])
950 951
              dataset.set_thread(1)
              filelist = [] # you should set your own filelist, e.g. filelist = ["dataA.txt"]
952 953 954 955
              dataset.set_filelist(filelist)
              exe.run(fluid.default_startup_program())
              exe.train_from_dataset(program=fluid.default_main_program(),
                                     dataset=dataset)
956 957

        """
958 959 960
        if dataset == None:
            raise RuntimeError("dataset is need and should be initialized")

H
hutuxian 已提交
961
        if program._pipeline_opt:
962 963 964
            thread = self._adjust_pipeline_resource(program._pipeline_opt,
                                                    dataset, thread)

J
jiaqi 已提交
965
        dataset._prepare_to_run()
966
        scope, trainer = self._prepare_trainer(
967 968 969 970 971 972 973 974
            program=program,
            dataset=dataset,
            scope=scope,
            thread=thread,
            debug=debug,
            fetch_list=fetch_list,
            fetch_info=fetch_info,
            print_period=print_period)
975
        trainer._gen_trainer_desc()
976
        self._dump_debug_info(program=program, trainer=trainer)
977
        dataset._dynamic_adjust_before_train(trainer.proto_desc.thread_num)
D
dongdaxiang 已提交
978 979 980
        self._default_executor.run_from_dataset(program.desc, scope,
                                                dataset.dataset,
                                                trainer._desc())
981
        dataset._dynamic_adjust_after_train()
J
jiaqi 已提交
982
        dataset._finish_to_run()
983
        return None