千言数据集：阅读理解之dureader_robust

1 介绍

本项目参考自官方的example，在此基础上增加了注释和实验。 DuReaderrobust数据集是单篇章、抽取式阅读理解数据集，具体的任务定义为：对于一个给定的问题q和一个篇章p，参赛系统需要根据篇章内容，给出该问题的答案a。数据集中的每个样本，是一个三元组<q, p, a>，例如：

问题 q: 乔丹打了多少个赛季
篇章 p: 迈克尔.乔丹在NBA打了15个赛季。他在84年进入nba，期间在1993年10月6日第一次退役改打棒球，95年3月18日重新回归，在99年1月13日第二次退役，后于2001年10月31日复出，在03年最终退役…
参考答案 (a): [‘15个’,‘15个赛季’]

2 准备和修改相关代码

2.1 准备相关代码

# 更新paddlenlp
!pip install --upgrade paddlenlp > /dev/null
# 新建文件夹code
!mkdir ~/code/ > /dev/null

mkdir: cannot create directory ‘/home/aistudio/code/’: File exists

# 下载example文件并将相关示例代码copy到code目录下
!git clone https://gitee.com/paddlepaddle/PaddleNLP.git ~/data 
!cp -r -n ~/data/examples/machine_reading_comprehension/DuReader-robust/. ~/code/
# 进入code目录
%cd ~/code

Cloning into '/home/aistudio/data'...
remote: Enumerating objects: 7736, done.[K
remote: Total 7736 (delta 0), reused 0 (delta 0), pack-reused 7736[K
Receiving objects: 100% (7736/7736), 54.66 MiB | 7.87 MiB/s, done.
Resolving deltas: 100% (4760/4760), done.
Checking connectivity... done.
/home/aistudio/code

2.2 修改相关代码进行训练

修改代码~/code/run_du.py，使其支持保存最佳结果，详细修改内容见源码(修改处有注释)。
新建~/code/train.sh，里面设置参数，然后进行训练

# 进入code目录
%cd ~/code
#训练开始
!bash train.sh

2.3 增加代码进行测试

修改代码~/code/args.py，增加参数test_model_path用来保存训练好的参数路径
新建~/code/test.sh，里面设置参数，然后进行测试

# 进入code目录
%cd ~/code
#测试开始
!bash test.sh

/home/aistudio/code
-----------  Configuration Arguments -----------
gpus: 0
heter_worker_num: None
heter_workers: 
http_port: None
ips: 127.0.0.1
log_dir: log
nproc_per_node: None
run_mode: None
server_num: None
servers: 
training_script: run_du.py
training_script_args: ['--task_name', 'dureader_robust', '--model_type', 'ernie_gram', '--model_name_or_path', 'ernie-gram-zh', '--test_model_path', './ckp/dureader-robust/best_0.889560.pdparams', '--max_seq_length', '384', '--batch_size', '12', '--learning_rate', '3e-5', '--num_train_epochs', '5', '--logging_steps', '10', '--valid_steps', '200', '--warmup_proportion', '0.1', '--weight_decay', '0.01', '--output_dir', './ckp/dureader-robust/', '--do_predict', '--device', 'gpu']
worker_num: None
workers: 
------------------------------------------------
WARNING 2021-07-14 18:28:51,347 launch.py:357] Not found distinct arguments and compiled with cuda or xpu. Default use collective mode
launch train in GPU mode!
INFO 2021-07-14 18:28:51,349 launch_utils.py:510] Local start 1 processes. First process distributed environment info (Only For Debug): 
    +=======================================================================================+
    |                        Distributed Envs                      Value                    |
    +---------------------------------------------------------------------------------------+
    |                       PADDLE_TRAINER_ID                        0                      |
    |                 PADDLE_CURRENT_ENDPOINT                 127.0.0.1:39019               |
    |                     PADDLE_TRAINERS_NUM                        1                      |
    |                PADDLE_TRAINER_ENDPOINTS                 127.0.0.1:39019               |
    |                     PADDLE_RANK_IN_NODE                        0                      |
    |                 PADDLE_LOCAL_DEVICE_IDS                        0                      |
    |                 PADDLE_WORLD_DEVICE_IDS                        0                      |
    |                     FLAGS_selected_gpus                        0                      |
    |             FLAGS_selected_accelerators                        0                      |
    +=======================================================================================+

INFO 2021-07-14 18:28:51,349 launch_utils.py:514] details abouts PADDLE_TRAINER_ENDPOINTS can be found in log/endpoints.log, and detail running logs maybe found in log/workerlog.0
launch proc_id:12491 idx:0
[2021-07-14 18:28:52,788] [    INFO] - Found /home/aistudio/.paddlenlp/models/ernie-gram-zh/vocab.txt
[2021-07-14 18:28:52,800] [    INFO] - Already cached /home/aistudio/.paddlenlp/models/ernie-gram-zh/ernie_gram_zh.pdparams
W0714 18:28:52.801362 12491 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0714 18:28:52.804977 12491 device_context.cc:422] device: 0, cuDNN Version: 7.6.
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py:1297: UserWarning: Skip loading for classifier.weight. classifier.weight is not found in the provided dict.
  warnings.warn(("Skip loading for {}. ".format(key) + str(err)))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py:1297: UserWarning: Skip loading for classifier.bias. classifier.bias is not found in the provided dict.
  warnings.warn(("Skip loading for {}. ".format(key) + str(err)))
Loaded parameters from ./ckp/dureader-robust/best_0.889560.pdparams

  0%|          | 0/20038 [00:00<?, ?it/s]
 32%|███▏      | 6426/20038 [00:00<00:00, 64256.54it/s]
 58%|█████▊    | 11568/20038 [00:00<00:00, 53066.41it/s]
 88%|████████▊ | 17568/20038 [00:00<00:00, 54971.04it/s]
100%|██████████| 20038/20038 [00:00<00:00, 53731.53it/s]
Processing example: 1000
time per 1000: 19.861122369766235
Processing example: 2000
time per 1000: 18.7887282371521
Processing example: 3000
time per 1000: 18.533133268356323
Processing example: 4000
time per 1000: 18.655021905899048
Processing example: 5000
time per 1000: 19.41499638557434
Processing example: 6000
time per 1000: 18.56358242034912
Processing example: 7000
time per 1000: 18.46160054206848
Processing example: 8000
time per 1000: 18.02328634262085
Processing example: 9000
time per 1000: 18.121554136276245
Processing example: 10000
time per 1000: 18.77463173866272
Processing example: 11000
time per 1000: 18.854174852371216
Processing example: 12000
time per 1000: 18.3427631855011
Processing example: 13000
time per 1000: 18.139728546142578
Processing example: 14000
time per 1000: 18.13940954208374
Processing example: 15000
time per 1000: 18.333805322647095
Processing example: 16000
time per 1000: 18.27450466156006
Processing example: 17000
time per 1000: 18.2143452167511
Processing example: 18000
time per 1000: 18.3362033367157
Processing example: 19000
time per 1000: 17.52031445503235
Processing example: 20000
time per 1000: 18.308919668197632
Processing example: 21000
time per 1000: 17.93433904647827
Processing example: 22000
time per 1000: 18.61131501197815
Processing example: 23000
time per 1000: 17.81761932373047
Processing example: 24000
time per 1000: 18.169885396957397
Processing example: 25000
time per 1000: 18.06153106689453
Processing example: 26000
time per 1000: 18.65974760055542
Processing example: 27000
time per 1000: 18.549249172210693
Processing example: 28000
time per 1000: 18.370241165161133
Processing example: 29000
time per 1000: 17.514158964157104
Processing example: 30000
time per 1000: 18.186832666397095
Processing example: 31000
time per 1000: 18.02071237564087
Processing example: 32000
time per 1000: 18.077165365219116
Processing example: 33000
time per 1000: 17.92798161506653
Processing example: 34000
time per 1000: 18.24454164505005
Processing example: 35000
time per 1000: 17.971704721450806
Processing example: 36000
time per 1000: 18.037253618240356
Processing example: 37000
time per 1000: 17.7172589302063
Processing example: 38000
time per 1000: 17.67001986503601
Processing example: 39000
time per 1000: 18.9000563621521
Processing example: 40000
time per 1000: 18.156458139419556
Processing example: 41000
time per 1000: 17.52651047706604
Processing example: 42000
time per 1000: 18.486640453338623
Processing example: 43000
time per 1000: 18.064385890960693
Processing example: 44000
time per 1000: 17.500937700271606
Processing example: 45000
time per 1000: 18.119248390197754
Processing example: 46000
time per 1000: 17.99043583869934
Processing example: 47000
time per 1000: 17.735183000564575
Processing example: 48000
time per 1000: 18.903116703033447
Processing example: 49000
time per 1000: 18.059256076812744
Processing example: 50000
time per 1000: 18.150585889816284
Processing example: 51000
time per 1000: 18.13495898246765
Processing example: 52000
time per 1000: 18.12674856185913
Processing example: 53000
time per 1000: 17.845890283584595
Processing example: 54000
time per 1000: 18.334473133087158
Processing example: 55000
time per 1000: 18.027803659439087
Processing example: 56000
time per 1000: 17.905311107635498
Processing example: 57000
time per 1000: 17.747634649276733
Processing example: 58000
time per 1000: 18.281975507736206
Processing example: 59000
time per 1000: 17.63224148750305
Processing example: 60000
time per 1000: 18.4889976978302
Processing example: 61000
time per 1000: 17.537192583084106
Processing example: 62000
time per 1000: 18.301725387573242
Processing example: 63000
time per 1000: 17.706255435943604
Processing example: 64000
time per 1000: 18.26703381538391
Processing example: 65000
time per 1000: 18.21987748146057
Processing example: 66000
time per 1000: 18.159775733947754
Processing example: 67000
time per 1000: 18.190679788589478
Processing example: 68000
time per 1000: 17.921314477920532
Processing example: 69000
time per 1000: 18.660756587982178
Processing example: 70000
time per 1000: 18.100632667541504
Processing example: 71000
time per 1000: 17.699832677841187
Processing example: 72000
time per 1000: 18.08768057823181
Processing example: 73000
time per 1000: 18.130647897720337
Processing example: 74000
time per 1000: 17.599804401397705
Processing example: 75000
time per 1000: 17.33307909965515
Processing example: 76000
time per 1000: 18.606958389282227
Processing example: 77000
time per 1000: 17.928603887557983
Processing example: 78000
time per 1000: 18.561939239501953
Processing example: 79000
time per 1000: 17.997174739837646
Processing example: 80000
time per 1000: 17.839795112609863
Processing example: 81000
time per 1000: 18.1271710395813
Processing example: 82000
time per 1000: 18.352195501327515
INFO 2021-07-14 18:58:41,044 launch.py:266] Local processes completed.