parallel_executor.py 8.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

import core
import multiprocessing
import framework
import executor
J
JiayiFeng 已提交
19
import warnings
Y
Yu Yang 已提交
20
import sys
21 22 23 24 25

__all__ = ['ParallelExecutor']


class ParallelExecutor(object):
X
Xin Pan 已提交
26 27
    def __init__(self,
                 use_cuda,
28 29
                 loss_name=None,
                 main_program=None,
X
Xin Pan 已提交
30
                 num_threads=None,
31
                 allow_op_delay=False,
Y
Yu Yang 已提交
32 33
                 share_vars_from=None,
                 customize_loss_grad=False):
34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66
        """
        ParallelExecutor can run program in parallel.

        Args:
            use_cuda(bool): Whether to use CUDA or not.
            loss_name(str, default None): The loss name must set in training.
            main_program(Program, default None): The program that need to run,
                if not provided, then default_main_program will be used.
            num_threads(int, default None): How many threads are used for
                training.
            allow_op_delay(bool, default False): Whether to delay and buffer
                some operators together for scheduling or not, which may
                improve performance in some cases, defalut False.
            share_vars_from(ParallelExecutor, default None): If provied,
                it will share variables from the specified ParallelExecutor.

        Returns:
            A ParallelExecutor object.

        Raises:
            TypeError: If share_vars_from is provided, but not ParallelExecutor
                object.

        Examples:
            .. code-block:: python

              train_exe = fluid.ParallelExecutor(
                  use_cuda=True, loss_name=loss.name)
              test_exe = fluid.ParallelExecutor(
                  use_cuda=True,
                  main_program=test_program,
                  share_vars_from=train_exe)

67 68
              train_loss, = train_exe.run([loss.name], feed=feed_dict)
              test_loss, = test_exe.run([loss.name], feed=feed_dict)
69 70
        """

X
Xin Pan 已提交
71 72
        self._places = []
        self._act_places = []
73 74 75
        if use_cuda:
            for i in xrange(core.get_cuda_device_count()):
                p = core.Place()
X
Xin Pan 已提交
76 77 78
                self._act_places.append(core.CUDAPlace(i))
                p.set_place(self._act_places[-1])
                self._places.append(p)
79 80 81
        else:
            for i in xrange(multiprocessing.cpu_count()):
                p = core.Place()
X
Xin Pan 已提交
82 83 84 85
                self._act_places.append(core.CPUPlace(i))
                p.set_place(self._act_places[-1])
                self._places.append(p)
        assert self._places, "no place for execution"
86 87

        if num_threads is None:
X
Xin Pan 已提交
88 89 90
            if use_cuda:
                # Experiments on se-resnext shows that too many threads hurt
                # performance. Worth tunning for other models in the future.
X
Xin Pan 已提交
91
                num_threads = len(self._places)
X
Xin Pan 已提交
92
            else:
93 94
                num_threads = min(
                    len(self._places) * 2, multiprocessing.cpu_count())
95

96 97
        main = main_program
        main = main if main else framework.default_main_program()
98 99
        scope = executor.global_scope()

100 101 102 103 104 105
        if share_vars_from and not isinstance(share_vars_from,
                                              ParallelExecutor):
            raise TypeError("share_vars_from must be ParallelExecutor.")
        local_scopes = share_vars_from.executor.local_scopes(
        ) if share_vars_from else []

T
typhoonzero 已提交
106
        self.persistable_vars = [
107
            v.name
108 109
            for v in filter(
                lambda var: var.persistable and var.type != core.VarDesc.VarType.RAW,
T
typhoonzero 已提交
110
                main.list_vars())
111 112
        ]

113 114 115
        self.executor = core.ParallelExecutor(
            num_threads,
            True if use_cuda else False,  # use_event
X
Xin Pan 已提交
116
            self._places,
117 118 119 120
            set([
                p.name for p in main.global_block().iter_parameters()
                if not p.stop_gradient
            ]),
T
typhoonzero 已提交
121
            set(self.persistable_vars),
122
            main.desc,
123
            loss_name if loss_name else '',
X
Xin Pan 已提交
124
            scope,
125
            local_scopes,
Y
Yu Yang 已提交
126 127
            allow_op_delay,
            customize_loss_grad)
128 129
        self.scope = scope

Y
Yu Yang 已提交
130
    def run(self, fetch_list, feed=None, feed_dict=None):
X
Xin Pan 已提交
131
        """
Y
Yu Yang 已提交
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
        Run a parallel executor with fetch_list.

        The feed parameter can be a dict or a list. If feed is a dict, the
        feed data will be split into multiple devices. If feed is a list, we
        assume the data has been splitted into multiple devices, the each
        element in the list will be copied to each device directly.

        For example, if the feed is a dict:
        >>> exe = ParallelExecutor()
        >>> # the image will be splitted into devices. If there is two devices
        >>> # each device will process an image with shape (24, 1, 28, 28)
        >>> exe.run(feed={'image': numpy.random.random(size=(48, 1, 28, 28))})

        For example, if the feed is a list:
        >>> exe = ParallelExecutor()
        >>> # each device will process each element in the list.
        >>> # the 1st device will process an image with shape (48, 1, 28, 28)
        >>> # the 2nd device will process an image with shape (32, 1, 28, 28)
        >>> #
        >>> # you can use exe.device_count to get the device number.
        >>> exe.run(feed=[{"image": numpy.random.random(size=(48, 1, 28, 28))},
        >>>               {"image": numpy.random.random(size=(32, 1, 28, 28))},
        >>>              ])

X
Xin Pan 已提交
156

Y
Yu Yang 已提交
157 158
        Args:
            fetch_list(list): The fetched variable names
Y
Yu Yang 已提交
159 160 161 162
            feed(list|dict|None): The feed variables. If the feed is a dict,
                tensors in that dict will be splitted into each devices. If
                the feed is a list, each element of the list will be copied
                to each device.
Y
Yu Yang 已提交
163
            feed_dict: Alias for feed parameter, for backward compatibility.
Y
Yu Yang 已提交
164
                This parameter is deprecated.
Y
Yu Yang 已提交
165 166 167

        Returns: fetched result list.

X
Xin Pan 已提交
168
        """
169
        if feed is None and feed_dict is not None:
J
JiayiFeng 已提交
170
            feed = feed_dict
Y
Yu Yang 已提交
171
            print >> sys.stderr, "`feed_dict` is deprecated. Please use `feed=`"
Y
Yu Yang 已提交
172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207

        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

            self.executor.feed_and_split_tensor_into_local_scopes(
                feed_tensor_dict)
        elif isinstance(feed, list) or isinstance(feed, tuple):
            if len(feed) != len(self._act_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._act_places[i])
                        tensor = tmp
                    res_dict[feed_name] = tensor
                res.append(res_dict)
            self.executor.feed_tensors_into_local_scopes(res)
X
Xin Pan 已提交
208

209
        fetch_var_name = '@FETCHED_VAR_NAME@'
Y
Yu Yang 已提交
210
        self.executor.run(fetch_list, fetch_var_name)
211 212
        arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array()
        return [arr[i] for i in range(len(arr))]
T
typhoonzero 已提交
213 214 215

    def bcast_params(self):
        self.executor.bcast_params(set(self.persistable_vars))
Y
Yu Yang 已提交
216 217 218 219

    @property
    def device_count(self):
        return len(self._act_places)