communicator.py 6.2 KB
Newer Older
1
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14
#
# 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.

15 16 17 18 19 20 21 22 23 24 25 26 27 28
# Copyright(c) 2019 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.

29
from .executor import global_scope
30 31 32 33
"""
Communicator is used for async distribute training in distribute_transpiler mode.
It's a wrapper of a cpp class Communicator and should be used inside fleet API.
"""
34
from . import core
35
from paddle.fluid.incubate.fleet.parameter_server.mode import DistributedMode
36

37
__all__ = ['Communicator', 'LargeScaleKV']
38 39 40


class Communicator(object):
41
    def __init__(self, mode, kwargs=None, envs=None):
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63
        """
        Communicator is used for async distribute training in distribute_transpiler mode.
        It's a wrapper of a cpp class Communicator and should be used inside fleet API.

        Args:
            program(Program): the trainers program after transpile of distribute_transpiler.
            It's used by communicator to extract the information to do communication.

        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.Program()
                comm = fluid.communicator.Communicator(prog)
                comm.start()
                comm.stop()
        """
        # set all recv op to not_run mode
64

65 66
        if mode == DistributedMode.SYNC:
            envs["pserver_endpoints"] = ','.join(kwargs["pserver_endpoints"])
67

68 69
        envs["trainers"] = str(kwargs["trainers"])
        envs["trainer_id"] = str(kwargs["trainer_id"])
70
        envs["need_global_step"] = str(kwargs["need_global_step"])
71
        envs["barrier_table_id"] = str(kwargs["barrier_table_id"])
72

73 74 75 76 77 78 79 80 81 82 83
        mode_str = None

        if mode == DistributedMode.SYNC:
            mode_str = "SYNC"
        elif mode == DistributedMode.ASYNC:
            mode_str = "ASYNC"
        elif mode == DistributedMode.HALF_ASYNC:
            mode_str = "HALF_ASYNC"
        elif mode == DistributedMode.GEO:
            mode_str = "GEO"

84 85 86
        self.mode = mode_str
        self.envs = envs
        self.communicator_ = None
87 88 89 90 91 92 93 94 95 96 97 98 99 100
        self.send_ctx_ = None
        self.recv_ctx_ = None

    def init_with_ctx(self,
                      send_ctx,
                      recv_ctx,
                      proto_txt,
                      unit64_hosts,
                      scope=global_scope()):
        self.communicator_ = core.DistCommunicator(self.mode, proto_txt,
                                                   unit64_hosts, send_ctx,
                                                   recv_ctx, scope, self.envs)
        self.send_ctx_ = send_ctx
        self.recv_ctx_ = recv_ctx
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138

    def start(self):
        """
        Start communicator. Should call before training process.

        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.Program()
                comm = fluid.communicator.Communicator(prog)
                comm.start()
                comm.stop()
        """
        self.communicator_.start()

    def stop(self):
        """
        Stop communicator. Should call after training process.

        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.Program()
                comm = fluid.communicator.Communicator(prog)
                comm.start()
                comm.stop()
        """
        self.communicator_.stop()
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156

    def is_running(self):
        """
        Get communicator is running or stop.

        Returns:
            bool

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.Program()
                comm = fluid.communicator.Communicator(prog)
                comm.is_running()
        """
        self.communicator_.is_running()
157 158 159 160

    def recv(self):
        self.communicator_.recv()

161 162 163
    def init_params(self, context):
        self.communicator_.init_params(context)

164 165 166
    def pull_dense(self, context):
        self.communicator_.pull_dense(context)

167 168 169 170 171 172 173 174 175 176 177
    def push_sparse_param(self, var_name, table_id=-1, scope=global_scope()):
        if not self.is_running():
            raise ValueError(
                "Communicator should init first. Using fleet.init_worker() before push_sparse_param()"
            )
        assert isinstance(var_name, str)
        assert isinstance(table_id, int)
        if table_id == -1:
            table_id = self.send_ctx_[var_name].table_id()
        self.communicator_.push_sparse_param(var_name, table_id, scope)

178 179 180 181 182 183 184 185 186 187

class LargeScaleKV(object):
    def __init__(self):
        self.scale_kv = core.LargeScaleKV()

    def save(self, varname, dirname):
        self.scale_kv.save(varname, dirname)

    def load(self, varname, dirname):
        self.scale_kv.load(varname, dirname)
188 189 190

    def size(self, varname):
        return self.scale_kv.size(varname)
191 192 193


class HeterClient(object):
194 195 196
    def __init__(self, endpoint, previous_endpoint, trainer_id):
        self.heter_client_ = core.HeterClient(endpoint, previous_endpoint,
                                              trainer_id)
197 198 199

    def stop(self):
        self.heter_client_.stop()
反馈
建议
客服 返回
顶部