Created by: shuoyin

hi, 我跑了PGL提供的graphsage的demo，可以正常跑，然后把本地的程序改成了fleet的分布式。网络结构和超参数都没有变，启动一个pserver和一个worker，发现fleet的分布式程序loss不降，请问事什么问题。 下面是我跑本地版graphsage的log 这是我fleet分布式的log 下面是main函数的分布式部分，我只修改了main函数

def main(args):
    data = load_data(args.normalize, args.symmetry)
    log.info("preprocess finish")
    log.info("Train Examples: %s" % len(data["train_index"]))
    log.info("Val Examples: %s" % len(data["val_index"]))
    log.info("Test Examples: %s" % len(data["test_index"]))
    log.info("Num nodes %s" % data["graph"].num_nodes)
    log.info("Num edges %s" % data["graph"].num_edges)
    log.info("Average Degree %s" % np.mean(data["graph"].indegree()))

    place = fluid.CUDAPlace(0) if args.use_cuda else fluid.CPUPlace()
    train_program = fluid.default_main_program()
    startup_program = fluid.default_startup_program()
    samples = []
    if args.samples_1 > 0:
        samples.append(args.samples_1)
    if args.samples_2 > 0:
        samples.append(args.samples_2)

    with fluid.program_guard(train_program, startup_program):
        feature, feature_init = paddle_helper.constant(
            "feat",
            dtype=data['feature'].dtype,
            value=data['feature'],
            hide_batch_size=False)

        graph_wrapper = pgl.graph_wrapper.GraphWrapper(
            "sub_graph", place, node_feat=data['graph'].node_feat_info())
        model_loss, model_acc = build_graph_model(
            graph_wrapper,
            num_class=data["num_class"],
            feature=feature,
            hidden_size=args.hidden_size,
            graphsage_type=args.graphsage_type,
            k_hop=len(samples))

    test_program = train_program.clone(for_test=True)
    
    trainer_id = int(os.environ["PADDLE_TRAINER_ID"])
    trainers = int(os.environ["PADDLE_TRAINERS"])
    training_role = os.environ["PADDLE_TRAINING_ROLE"]
    log.info(training_role )
    training_role = role_maker.Role.WORKER if training_role == "TRAINER" else role_maker.Role.SERVER
    ports = os.getenv("PADDLE_PSERVER_PORTS")
    pserver_ip = os.getenv("PADDLE_PSERVER_IP", "")
    pserver_endpoints = []
    for port in ports.split(","):
        pserver_endpoints.append(':'.join([pserver_ip, port]))
    role = role_maker.UserDefinedRoleMaker(current_id=trainer_id, role=training_role, worker_num=trainers, server_endpoints=pserver_endpoints)
    config = DistributeTranspilerConfig()
    config.sync_mode = True

    fleet.init(role)
    optimizer = fluid.optimizer.SGD(learning_rate=args.lr)
    optimizer = fleet.distributed_optimizer(optimizer, config)
    optimizer.minimize(model_loss)

    exe = fluid.Executor(place)

    if fleet.is_server():
        log.info('running server')
        fleet.init_server()
        fleet.run_server()

    if fleet.is_worker():
        log.info('running worker')
        fleet.init_worker()
        exe.run(fleet.startup_program)
        feature_init(place)