Created by: minhozhou

• 版本、环境信息：    1）PaddlePaddle版本：1.6.1，静态图模式    2）CPU：厂内CPU    3）系统环境：python3.7
• 训练信息    1）单机
• 复现信息：如为报错，请给出复现环境、复现步骤
• 问题描述：请详细描述您的问题，同步贴出报错信息、日志、可复现的代码片段 优化器部分代码：

def optimization(cloud_train, base_lr, loss, train_steps, optimizer='adam'):
    decayed_lr = fluid.layers.polynomial_decay(base_lr, train_steps, 0.0001)

    if optimizer == 'sgd':
        optimizer = fluid.optimizer.SGD(
            decayed_lr,
            regularization=fluid.regularizer.L2DecayRegularizer(
                regularization_coeff=0.0025))
    elif optimizer == 'adam':
        # dont use gpu's lazy mode
        optimizer = fluid.optimizer.Adam(decayed_lr)
    else:
        raise ValueError
    log.info('learning rate:%f' % (base_lr))
    optimizer.minimize(loss)
    return decayed_lr

训练部分代码

def train_loop(train_exe, exe, decayed_lr, program, loss, data_reader, args, src_embed, input_data_list):
    """ train
    """
    model_save_dir = args.output_path
    if not os.path.exists(model_save_dir):
        os.makedirs(model_save_dir)

    trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
    step = 0
    epoch = 0
    for epoch in range(args.epoch):
        for data in data_reader():
            print(decayed_lr)
            print(train_exe.run(fetch_list=[decayed_lr.name]))
            begin_time = time.time()
            loss_val = train_exe.run(fetch_list=[loss], feed=data)
            step += 1
            if step % 10 == 0:
                log.info("epoch %s: step %s: loss %.5f lr: %.5f speed: %.5f s/step reader qsize: %s" %
                         (epoch, step, np.mean(loss_val), lr, time.time() - begin_time, data_reader.queue.size()))

报错信息

如果把 print(train_exe.run(fetch_list=[decayed_lr.name]))这句话在训练循环中去掉，模型可以正常训练，