Created by: Angus07

使用paddle 1.6.1 在预测压测，发现内存持续增长。

代码如下， 首先是建立一个全局的网络，包括reader

def main_test():
    """
    创建网络
    """
    global test_exe
    global reader
    global test_prog
    global test_pyreader, test_graph_vars
    global place
    ernie_config = ErnieConfig(args.ernie_config_path)
    ernie_config.print_config()

    gpus = os.getenv("FLAGS_selected_gpus").split(",")
    gpu_id = int(gpus[0])
    gpus = fluid.core.get_cuda_device_count()
    if args.use_cuda:
        place = fluid.CUDAPlace(gpu_id)
        print("use gpu")
        dev_count = len(gpus) if args.is_distributed else gpus
    else:
        place = fluid.CPUPlace()
        print("use cpu")
        dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))

    reader = task_reader.Seq2SeqReader(
        vocab_path=args.vocab_path,
        label_map_config=args.label_map_config,
        max_seq_len=args.max_seq_len,
        do_lower_case=args.do_lower_case,
        in_tokens=args.in_tokens,
        random_seed=args.random_seed,
        tgt_type_id=args.tgt_type_id,
        tokenizer=args.tokenizer,
        multi_segment=args.multi_segment)
    startup_prog = fluid.Program()
    if args.random_seed is not None:
        startup_prog.random_seed = args.random_seed
    test_prog = fluid.Program()
    with fluid.program_guard(test_prog, startup_prog):
        with fluid.unique_name.guard():
            test_pyreader, test_graph_vars = create_model(
                args,
                pyreader_name='test_reader',
                ernie_config=ernie_config,
                is_prediction=args.do_dec)

    test_prog = test_prog.clone(for_test=True)

    exe = fluid.Executor(place)
    exe.run(startup_prog)
    if not args.init_checkpoint:
        raise ValueError("args 'init_checkpoint' should be set if"
                         "only doing validation or testing!")
    init_checkpoint(
        exe,
        args.init_checkpoint,
        main_program=startup_prog,
        use_fp16=args.use_fp16)
    test_exe = exe

在回调函数中实现主体预测程序。

      test_pyreader.decorate_tensor_provider(
            reader.data_generator_from_input(
                src_contents,
                batch_size=args.batch_size,
                epoch=1,
                dev_count=1,
                shuffle=False,
                phase='test',
                do_dec=args.do_dec,
                place=place))
        output = infer_test(test_exe, test_prog, test_pyreader, test_graph_vars, "test", do_dec=args.do_dec,
                            vocab_path=args.vocab_path, features=reader.get_features("test"),
                            decode_path=args.decode_path + "/test_final_preds")

其中Infer_test封装的代码如下：

    pyreader.start()
    while True:
        try:
            outputs = exe.run(program=program, fetch_list=fetch_list,
                              return_numpy=return_numpy)
            seq_ids, seq_scores, data_ids = outputs
            seq_ids_list, seq_scores_list = [seq_ids], [
                seq_scores] if isinstance(
                seq_ids, paddle.fluid.core.LoDTensor) else (seq_ids, seq_scores)

            data_ids = np.array(data_ids).reshape(-1).tolist()
            data_idx = 0
            ......
                        score = np.array(seq_scores)[sub_end - 1]
                        data_id = data_ids[data_idx]
                        # not support batch
                        dec_out[data_id].append((tokens, score))
                        if cnt >= 2:
                            break
                        cnt += 1
                    data_idx += 1

        except fluid.core.EOFException:
            pyreader.reset()
            break

补充代码：

pyreader = fluid.layers.py_reader(
capacity=50,
shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, args.max_seq_len, args.max_seq_len],
 [-1, args.max_seq_len, args.max_seq_len],
[-1, 1], [-1, 1]],
dtypes=['int64', 'int64', 'int64', 'float32', 'float32', 'int64', 'int64'],
od_levels=[0, 0, 0, 0, 0, 0, 0],
name=pyreader_name,
use_double_buffer=True)