|
| 1 | +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | +import argparse |
| 15 | +import time |
| 16 | +import os |
| 17 | +import math |
| 18 | +import numpy as np |
| 19 | + |
| 20 | +import paddle.fluid as F |
| 21 | +import paddle.fluid.layers as L |
| 22 | +from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import fleet |
| 23 | +from paddle.fluid.transpiler.distribute_transpiler import DistributeTranspilerConfig |
| 24 | +import paddle.fluid.incubate.fleet.base.role_maker as role_maker |
| 25 | +from pgl.utils.logger import log |
| 26 | + |
| 27 | +from model import Metapath2vecModel |
| 28 | +from graph import m2vGraph |
| 29 | +from utils import load_config |
| 30 | +from walker import multiprocess_data_generator |
| 31 | + |
| 32 | + |
| 33 | +def init_role(): |
| 34 | + # reset the place according to role of parameter server |
| 35 | + training_role = os.getenv("TRAINING_ROLE", "TRAINER") |
| 36 | + paddle_role = role_maker.Role.WORKER |
| 37 | + place = F.CPUPlace() |
| 38 | + if training_role == "PSERVER": |
| 39 | + paddle_role = role_maker.Role.SERVER |
| 40 | + |
| 41 | + # set the fleet runtime environment according to configure |
| 42 | + ports = os.getenv("PADDLE_PORT", "6174").split(",") |
| 43 | + pserver_ips = os.getenv("PADDLE_PSERVERS").split(",") # ip,ip... |
| 44 | + eplist = [] |
| 45 | + if len(ports) > 1: |
| 46 | + # local debug mode, multi port |
| 47 | + for port in ports: |
| 48 | + eplist.append(':'.join([pserver_ips[0], port])) |
| 49 | + else: |
| 50 | + # distributed mode, multi ip |
| 51 | + for ip in pserver_ips: |
| 52 | + eplist.append(':'.join([ip, ports[0]])) |
| 53 | + |
| 54 | + pserver_endpoints = eplist # ip:port,ip:port... |
| 55 | + worker_num = int(os.getenv("PADDLE_TRAINERS_NUM", "0")) |
| 56 | + trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) |
| 57 | + role = role_maker.UserDefinedRoleMaker( |
| 58 | + current_id=trainer_id, |
| 59 | + role=paddle_role, |
| 60 | + worker_num=worker_num, |
| 61 | + server_endpoints=pserver_endpoints) |
| 62 | + fleet.init(role) |
| 63 | + |
| 64 | + |
| 65 | +def optimization(base_lr, loss, train_steps, optimizer='sgd'): |
| 66 | + decayed_lr = L.learning_rate_scheduler.polynomial_decay( |
| 67 | + learning_rate=base_lr, |
| 68 | + decay_steps=train_steps, |
| 69 | + end_learning_rate=0.0001 * base_lr, |
| 70 | + power=1.0, |
| 71 | + cycle=False) |
| 72 | + if optimizer == 'sgd': |
| 73 | + optimizer = F.optimizer.SGD(decayed_lr) |
| 74 | + elif optimizer == 'adam': |
| 75 | + optimizer = F.optimizer.Adam(decayed_lr, lazy_mode=True) |
| 76 | + else: |
| 77 | + raise ValueError |
| 78 | + |
| 79 | + log.info('learning rate:%f' % (base_lr)) |
| 80 | + #create the DistributeTranspiler configure |
| 81 | + config = DistributeTranspilerConfig() |
| 82 | + config.sync_mode = False |
| 83 | + #config.runtime_split_send_recv = False |
| 84 | + |
| 85 | + config.slice_var_up = False |
| 86 | + #create the distributed optimizer |
| 87 | + optimizer = fleet.distributed_optimizer(optimizer, config) |
| 88 | + optimizer.minimize(loss) |
| 89 | + |
| 90 | + |
| 91 | +def build_complied_prog(train_program, model_loss): |
| 92 | + num_threads = int(os.getenv("CPU_NUM", 10)) |
| 93 | + trainer_id = int(os.getenv("PADDLE_TRAINER_ID", 0)) |
| 94 | + exec_strategy = F.ExecutionStrategy() |
| 95 | + exec_strategy.num_threads = num_threads |
| 96 | + #exec_strategy.use_experimental_executor = True |
| 97 | + build_strategy = F.BuildStrategy() |
| 98 | + build_strategy.enable_inplace = True |
| 99 | + #build_strategy.memory_optimize = True |
| 100 | + build_strategy.memory_optimize = False |
| 101 | + build_strategy.remove_unnecessary_lock = False |
| 102 | + if num_threads > 1: |
| 103 | + build_strategy.reduce_strategy = F.BuildStrategy.ReduceStrategy.Reduce |
| 104 | + |
| 105 | + compiled_prog = F.compiler.CompiledProgram( |
| 106 | + train_program).with_data_parallel(loss_name=model_loss.name) |
| 107 | + return compiled_prog |
| 108 | + |
| 109 | + |
| 110 | +def train_prog(exe, program, loss, node2vec_pyreader, args, train_steps): |
| 111 | + trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) |
| 112 | + step = 0 |
| 113 | + if not os.path.exists(args.save_path): |
| 114 | + os.makedirs(args.save_path) |
| 115 | + while True: |
| 116 | + try: |
| 117 | + begin_time = time.time() |
| 118 | + loss_val, = exe.run(program, fetch_list=[loss]) |
| 119 | + log.info("step %s: loss %.5f speed: %.5f s/step" % |
| 120 | + (step, np.mean(loss_val), time.time() - begin_time)) |
| 121 | + step += 1 |
| 122 | + except F.core.EOFException: |
| 123 | + node2vec_pyreader.reset() |
| 124 | + |
| 125 | + if step % args.steps_per_save == 0 or step == train_steps: |
| 126 | + save_path = args.save_path |
| 127 | + if trainer_id == 0: |
| 128 | + model_path = os.path.join(save_path, "%s" % step) |
| 129 | + fleet.save_persistables(exe, model_path) |
| 130 | + |
| 131 | + if step == train_steps: |
| 132 | + break |
| 133 | + |
| 134 | + |
| 135 | +def main(args): |
| 136 | + log.info("start") |
| 137 | + |
| 138 | + worker_num = int(os.getenv("PADDLE_TRAINERS_NUM", "0")) |
| 139 | + num_devices = int(os.getenv("CPU_NUM", 10)) |
| 140 | + |
| 141 | + model = Metapath2vecModel(config=args) |
| 142 | + pyreader = model.pyreader |
| 143 | + loss = model.forward() |
| 144 | + |
| 145 | + # init fleet |
| 146 | + init_role() |
| 147 | + |
| 148 | + train_steps = math.ceil(args.num_nodes * args.epochs / args.batch_size / |
| 149 | + num_devices / worker_num) |
| 150 | + log.info("Train step: %s" % train_steps) |
| 151 | + |
| 152 | + real_batch_size = args.batch_size * args.walk_len * args.win_size |
| 153 | + if args.optimizer == "sgd": |
| 154 | + args.lr *= real_batch_size |
| 155 | + optimization(args.lr, loss, train_steps, args.optimizer) |
| 156 | + |
| 157 | + # init and run server or worker |
| 158 | + if fleet.is_server(): |
| 159 | + fleet.init_server(args.warm_start_from_dir) |
| 160 | + fleet.run_server() |
| 161 | + |
| 162 | + if fleet.is_worker(): |
| 163 | + log.info("start init worker done") |
| 164 | + fleet.init_worker() |
| 165 | + #just the worker, load the sample |
| 166 | + log.info("init worker done") |
| 167 | + |
| 168 | + exe = F.Executor(F.CPUPlace()) |
| 169 | + exe.run(fleet.startup_program) |
| 170 | + log.info("Startup done") |
| 171 | + |
| 172 | + dataset = m2vGraph(args) |
| 173 | + log.info("Build graph done.") |
| 174 | + |
| 175 | + data_generator = multiprocess_data_generator(args, dataset) |
| 176 | + |
| 177 | + cur_time = time.time() |
| 178 | + for idx, _ in enumerate(data_generator()): |
| 179 | + log.info("iter %s: %s s" % (idx, time.time() - cur_time)) |
| 180 | + cur_time = time.time() |
| 181 | + if idx == 100: |
| 182 | + break |
| 183 | + |
| 184 | + pyreader.decorate_tensor_provider(data_generator) |
| 185 | + pyreader.start() |
| 186 | + |
| 187 | + compiled_prog = build_complied_prog(fleet.main_program, loss) |
| 188 | + train_prog(exe, compiled_prog, loss, pyreader, args, train_steps) |
| 189 | + |
| 190 | + |
| 191 | +if __name__ == '__main__': |
| 192 | + parser = argparse.ArgumentParser(description='metapath2vec') |
| 193 | + parser.add_argument("-c", "--config", type=str, default="./config.yaml") |
| 194 | + args = parser.parse_args() |
| 195 | + config = load_config(args.config) |
| 196 | + log.info(config) |
| 197 | + main(config) |
0 commit comments