---
title: Hook的index顺序改变导致训练异常-昇腾社区
description: 训练报错：AttributeError:'xxxxHook' object has no attribute 'xxxx'。
keywords: Hook
url: https://www.hiascend.com/document/caselibrary/detail/topic_0000001344190745
section: (其他)
---

# Hook的index顺序改变导致训练异常-昇腾社区

URL: https://www.hiascend.com/document/caselibrary/detail/topic_0000001344190745
描述: 训练报错：AttributeError:'xxxxHook' object has no attribute 'xxxx'。
关键词: Hook

昇腾文档 [了解详情](https://www.hiascend.com/document)

故障案例 [了解详情](https://www.hiascend.com/document/caselibrary)

Hook的index顺序改变导致训练异常

Hook的index顺序改变导致训练异常

2022/07/26

890

问题信息

| 问题来源 | 产品大类 | 产品子类 | 关键字 |
| --- | --- | --- | ---|
| 官方 | 模型训练 | TensorFlow | Hook |

#### 问题现象描述

训练报错：AttributeError:'xxxxHook' object has no attribute 'xxxx'。

#### 原因分析

迁移工具对于Estimator脚本默认添加了NPUBroadcastHook，而出错的场景中对于hooks这个list，在添加完NPU的hook之后，改变了hooks这个list中的index顺序，所以出现了取hooks[-1]进行其他操作报错的问题。

```
training_hooks.append(LogTrainRunHook(global_batch_size, hvd_rank, FLAGS.save_checkpoints_steps, num_steps_ignore_xla=25))
...
    estimator.train(input_fn=train_input_fn, max_steps=num_train_steps, hooks=npu_hooks_append(hooks_list=training_hooks))
    train_time_elapsed = time.time() - train_start_time
    train_time_wo_overhead = training_hooks[-1].total_time
    avg_sentences_per_second = num_train_steps * global_batch_size * 1.0 / train_time_elapsed
    ss_sentences_per_second = (training_hooks[-1].count - training_hooks[-1].skipped) * global_batch_size * 1.0 / train_time_wo_overhead
```

#### 解决措施

需要修改训练脚本中hook list中的index。针对以上脚本，进行如下修改即可执行成功。

```
training_hooks.append(LogTrainRunHook(global_batch_size, hvd_rank, FLAGS.save_checkpoints_steps, num_steps_ignore_xla=25))
...
    estimator.train(input_fn=train_input_fn, max_steps=num_train_steps, hooks=npu_hooks_append(hooks_list=training_hooks))
    train_time_elapsed = time.time() - train_start_time
    train_time_wo_overhead = training_hooks[-2].total_time
    avg_sentences_per_second = num_train_steps * global_batch_size * 1.0 / train_time_elapsed
    ss_sentences_per_second = (training_hooks[-2].count - training_hooks[-2].skipped) * global_batch_size * 1.0 / train_time_wo_overhead
```

本页内容

- 问题信息
- 问题现象描述
- 原因分析
- 解决措施