cluster_trainer.py 6.5 KB
Newer Older
T
tangwei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
# Copyright (c) 2020 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.

"""
Training use fluid with one node only.
"""

from __future__ import print_function

C
chengmo 已提交
21
import os
T
tangwei 已提交
22
import paddle.fluid as fluid
T
tangwei 已提交
23 24 25
from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import fleet
from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler.distributed_strategy import StrategyFactory
from paddle.fluid.incubate.fleet.base.role_maker import PaddleCloudRoleMaker
T
bug fix  
tangwei12 已提交
26
from paddle.fluid.incubate.fleet.base.role_maker import MPISymetricRoleMaker
T
tangwei 已提交
27

28 29
from paddlerec.core.utils import envs
from paddlerec.core.trainers.transpiler_trainer import TranspileTrainer
T
tangwei 已提交
30 31


T
tangwei 已提交
32
class ClusterTrainer(TranspileTrainer):
T
tangwei 已提交
33
    def processor_register(self):
T
tangwei 已提交
34
        role = PaddleCloudRoleMaker()
T
tangwei 已提交
35 36
        fleet.init(role)

T
tangwei12 已提交
37
        if fleet.is_server():
T
tangwei 已提交
38 39 40 41 42 43
            self.regist_context_processor('uninit', self.instance)
            self.regist_context_processor('init_pass', self.init)
            self.regist_context_processor('server_pass', self.server)
        else:
            self.regist_context_processor('uninit', self.instance)
            self.regist_context_processor('init_pass', self.init)
C
chengmo 已提交
44
            self.regist_context_processor('startup_pass', self.startup)
Y
add din  
yaoxuefeng 已提交
45
            if envs.get_platform() == "LINUX" and envs.get_global_env("dataset_class", None, "train.reader") != "DataLoader":
T
tangwei 已提交
46 47
                self.regist_context_processor('train_pass', self.dataset_train)
            else:
C
chengmo 已提交
48 49
                self.regist_context_processor(
                    'train_pass', self.dataloader_train)
50
            self.regist_context_processor('infer_pass', self.infer)
T
tangwei 已提交
51 52 53
            self.regist_context_processor('terminal_pass', self.terminal)

    def build_strategy(self):
T
tangwei 已提交
54
        mode = envs.get_runtime_environ("train.trainer.strategy")
T
fix bug  
tangwei 已提交
55
        assert mode in ["async", "geo", "sync", "half_async"]
T
tangwei 已提交
56

T
tangwei 已提交
57 58 59 60 61 62 63 64 65 66 67 68 69 70
        strategy = None

        if mode == "async":
            strategy = StrategyFactory.create_async_strategy()
        elif mode == "geo":
            push_num = envs.get_global_env("train.strategy.mode.push_num", 100)
            strategy = StrategyFactory.create_geo_strategy(push_num)
        elif mode == "sync":
            strategy = StrategyFactory.create_sync_strategy()
        elif mode == "half_async":
            strategy = StrategyFactory.create_half_async_strategy()

        assert strategy is not None

T
tangwei 已提交
71
        self.strategy = strategy
T
tangwei 已提交
72 73 74
        return strategy

    def init(self, context):
T
tangwei 已提交
75
        self.model.train_net()
T
tangwei 已提交
76
        optimizer = self.model.optimizer()
C
chengmo 已提交
77 78 79 80 81
        optimizer_name = envs.get_global_env(
            "hyper_parameters.optimizer", None, "train.model")
        if optimizer_name not in ["", "sgd", "SGD", "Sgd"]:
            os.environ["FLAGS_communicator_is_sgd_optimizer"] = '0'

T
tangwei 已提交
82 83
        strategy = self.build_strategy()
        optimizer = fleet.distributed_optimizer(optimizer, strategy)
T
tangwei 已提交
84
        optimizer.minimize(self.model.get_cost_op())
T
tangwei 已提交
85 86 87 88

        if fleet.is_server():
            context['status'] = 'server_pass'
        else:
T
tangwei 已提交
89 90 91 92 93 94 95 96
            self.fetch_vars = []
            self.fetch_alias = []
            self.fetch_period = self.model.get_fetch_period()

            metrics = self.model.get_metrics()
            if metrics:
                self.fetch_vars = metrics.values()
                self.fetch_alias = metrics.keys()
C
chengmo 已提交
97
            context['status'] = 'startup_pass'
T
tangwei 已提交
98 99 100 101 102 103

    def server(self, context):
        fleet.init_server()
        fleet.run_server()
        context['is_exit'] = True

C
chengmo 已提交
104
    def startup(self, context):
T
tangwei 已提交
105
        self._exe.run(fleet.startup_program)
C
chengmo 已提交
106
        context['status'] = 'train_pass'
T
tangwei 已提交
107

C
chengmo 已提交
108
    def dataloader_train(self, context):
T
tangwei 已提交
109 110
        fleet.init_worker()

T
tangwei 已提交
111
        reader = self._get_dataloader()
T
tangwei 已提交
112 113
        epochs = envs.get_global_env("train.epochs")

T
tangwei 已提交
114 115 116
        program = fluid.compiler.CompiledProgram(
            fleet.main_program).with_data_parallel(
            loss_name=self.model.get_cost_op().name,
C
chengmo 已提交
117 118
            build_strategy=self.strategy.get_build_strategy(),
            exec_strategy=self.strategy.get_execute_strategy())
T
tangwei 已提交
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143

        metrics_varnames = []
        metrics_format = []

        metrics_format.append("{}: {{}}".format("epoch"))
        metrics_format.append("{}: {{}}".format("batch"))

        for name, var in self.model.get_metrics().items():
            metrics_varnames.append(var.name)
            metrics_format.append("{}: {{}}".format(name))

        metrics_format = ", ".join(metrics_format)

        for epoch in range(epochs):
            reader.start()
            batch_id = 0
            try:
                while True:
                    metrics_rets = self._exe.run(
                        program=program,
                        fetch_list=metrics_varnames)

                    metrics = [epoch, batch_id]
                    metrics.extend(metrics_rets)

144
                    if batch_id % self.fetch_period == 0 and batch_id != 0:
T
tangwei 已提交
145 146 147 148
                        print(metrics_format.format(*metrics))
                    batch_id += 1
            except fluid.core.EOFException:
                reader.reset()
149
            self.save(epoch, "train", is_fleet=True)
T
tangwei 已提交
150

T
tangwei 已提交
151
        fleet.stop_worker()
152
        context['status'] = 'infer_pass'
T
tangwei 已提交
153

T
tangwei 已提交
154 155 156 157 158 159 160 161 162 163 164 165 166 167
    def dataset_train(self, context):
        fleet.init_worker()

        dataset = self._get_dataset()
        epochs = envs.get_global_env("train.epochs")

        for i in range(epochs):
            self._exe.train_from_dataset(program=fluid.default_main_program(),
                                         dataset=dataset,
                                         fetch_list=self.fetch_vars,
                                         fetch_info=self.fetch_alias,
                                         print_period=self.fetch_period)
            self.save(i, "train", is_fleet=True)
        fleet.stop_worker()
168
        context['status'] = 'infer_pass'
T
tangwei 已提交
169 170 171 172 173

    def terminal(self, context):
        for model in self.increment_models:
            print("epoch :{}, dir: {}".format(model[0], model[1]))
        context['is_exit'] = True