注意
本项目代码包含多个文件, Fork并使用GPU环境来运行后, 才能看到项目完整代码, 并正确运行:
并请检查相关参数设置, 例如use_gpu, fluid.CUDAPlace(0)等处是否设置正确.
下载安装命令
## CPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/cpu paddlepaddle
## GPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/gpu paddlepaddle-gpu Auto Dialogue Evaluation
简介
任务说明
对话自动评估（Auto Dialogue Evaluation）评估开放领域对话系统的回复质量，能够帮助企业或个人快速评估对话系统的回复质量，减少人工评估成本。
- 在无标注数据的情况下，利用负采样训练匹配模型作为评估工具，实现对多个对话系统回复质量排序；
- 利用少量标注数据（特定对话系统或场景的人工打分），在匹配模型基础上进行微调，可以显著提高该对话系统或场景的评估效果。
效果说明
我们以四个不同的对话系统（seq2seq_naive／seq2seq_att／keywords／human）为例，使用对话自动评估工具进行自动评估。
-
无标注数据情况下，直接使用预训练好的评估工具进行评估； 在四个对话系统上，自动评估打分和人工评估打分spearman相关系数，如下：
/ seq2seq_naive seq2seq_att keywords human cor 0.361 0.343 0.324 0.288 对四个系统平均得分排序：
人工评估 k(0.591)<n(0.847)<a(1.116)<h(1.240) 自动评估 k(0.625)<n(0.909)<a(1.399)<h(1.683) -
利用少量标注数据微调后，自动评估打分和人工打分spearman相关系数，如下：
/ seq2seq_naive seq2seq_att keywords human cor 0.474 0.477 0.443 0.378
任务定义与建模
对话自动评估任务输入是文本对（上文，回复），输出是回复质量得分。
模型原理介绍
匹配任务（预测上下文是否匹配）和自动评估任务有天然的联系，该项目利用匹配任务作为自动评估的预训练；
利用少量标注数据，在匹配模型基础上微调。
数据格式说明
训练、预测、评估使用的数据示例如下，数据由三列组成，以制表符（'\t'）分隔，第一列是以空格分开的上文id，第二列是以空格分开的回复id，第三列是标签
723 236 7823 12 8 887 13 77 4 2
8474 13 44 34 2 87 91 23 0
注：本项目额外提供了分词预处理脚本（在preprocess目录下），可供用户使用，具体使用方法如下：
python tokenizer.py --test_data_dir ./test.txt.utf8 --batch_size 1 > test.txt.utf8.seg
代码结构说明
main.py：该项目的主函数，封装包括训练、预测、评估的部分
config.py：定义了该项目模型的相关配置，包括具体模型类别、以及模型的超参数
reader.py：定义了读入数据，加载词典的功能
evaluation.py：定义评估函数
init.py:定义模型load函数
run.sh：训练、预测、评估运行脚本
文件介绍
auto_dialogue_evaluation/: 存放对话自动评估模型的主要执行文件
auto_dialogue_evaluation/data目录下存放unlabel_data(train.ids/val.ids/test.ids)，lable_data(四个任务数据train.ids/val.ids/test.ids)，以及word2ids，
auto_dialogue_evaluation/model_files存放预训练模型
auto_dialogue_evaluation/model_files_tmp存放训练好的模型
models:共享的模型集合
preprocess:共享的数据预处理流程
本例算法运行基于GPU，若采用CPU，请调用run_CPU.sh
In[1]
#可以通过以下方式查看数据格式，此处仅展示部分
f = open('auto_dialogue_evaluation/data/label_data/seq2seq_att/train.ids', "r")
i=0
for line in f:
i=i+1
print(line)
if i > 10:
break8 644 10 98 4494 3 19 218 18 182 464 3 20 32 15 837 880 10 27 32 1128 105 32 10 98 63 25 32 7 63 0 0 13863 1348 21 693 20 515 82518 181 1590 3 911 16015 5 2 5681 3 12274 8077 0 186 9 7 322 9761 3 20 622 5 1135 3 7726 17 9 22 51 3 1610 2 5374 5714 1 1224 134 3 1109 458 528 92 42 51 99 19 4 4621 94 73 2 149 151 94 73 4 53 720 25 453 123 104 4621 81 1 943 2 3303 723 4 4 130 3228 4 7 10 12 11750 3 0 1850 2612 534 12199 62 55 187 118 29 1 32138 820 9651 375 149 37 64 481 25 10 12 1048 3 1683 0 1494 5535 1479 1395 213 26815 944 1574 110886 183 38 5535 733 19 11 18 5535 2 5535 1 2543 663 50 64 10114 94 73 11 138 8 5223 11 0 753 6 21 5 2 4028 213 257 249 753 1955 913 2 18 16566 2 37 184 11 8 554 19 797 1 7 26 59 116 586 15663 11634 2 375 4831 3 2 25471 1534 13 26 275 30 564 1
In[1]
#基于示例的数据集，可以运行下面的命令，进行第一阶段训练,训练参数详见run.sh文件，可根据自己的需要，修改训练参数。
!cd auto_dialogue_evaluation && sh run.sh train----------- Configuration Arguments ----------- batch_size: 256 do_infer: False do_train: True do_val: False emb_size: 256 hidden_size: 256 init_model: None learning_rate: 0.001 loss_type: CLS max_len: 50 num_scan_data: 50 out_path: None print_step: 50 sample_pro: 1 save_path: model_files_tmp/matching_pretrained save_step: 10 test_path: None train_path: data/unlabel_data/train.ids use_cuda: True val_path: data/unlabel_data/val.ids vocab_size: 484016 word_emb_init: None ------------------------------------------------ context_rep shape: (-1L, 256L) response_rep shape: (-1L, 256L) logits shape: (-1L, 1L) before reduce mean loss shape: (-1L, 1L) after reduce mean loss shape: (1L,) begin memory optimization ... 2020-03-03 20:15:27,444-WARNING: Caution! paddle.fluid.memory_optimize() is deprecated and not maintained any more, since it is not stable! This API would not take any memory optimizations on your Program now, since we have provided default strategies for you. The newest and stable memory optimization strategies (they are all enabled by default) are as follows: 1. Garbage collection strategy, which is enabled by exporting environment variable FLAGS_eager_delete_tensor_gb=0 (0 is the default value). 2. Inplace strategy, which is enabled by setting build_strategy.enable_inplace=True (True is the default value) when using CompiledProgram or ParallelExecutor. end memory optimization ... context_rep shape: (-1L, 256L) response_rep shape: (-1L, 256L) logits shape: (-1L, 1L) before reduce mean loss shape: (-1L, 1L) after reduce mean loss shape: (1L,) device count 1 theoretical memory usage: (1519.714105796814, 1592.081444168091, 'MB') W0303 20:15:28.347313 92 device_context.cc:236] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 9.2, Runtime API Version: 9.0 W0303 20:15:28.350337 92 device_context.cc:244] device: 0, cuDNN Version: 7.3. start loading data ... I0303 20:15:29.676050 92 parallel_executor.cc:421] The number of CUDAPlace, which is used in ParallelExecutor, is 1. And the Program will be copied 1 copies I0303 20:15:29.677879 92 build_strategy.cc:363] SeqOnlyAllReduceOps:0, num_trainers:1 I0303 20:15:29.679220 92 parallel_executor.cc:285] Inplace strategy is enabled, when build_strategy.enable_inplace = True I0303 20:15:29.680267 92 parallel_executor.cc:368] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0 Pass 0, pass_time_cost 0.05 sec Pass 1, pass_time_cost 0.03 sec Pass 2, pass_time_cost 0.03 sec Pass 3, pass_time_cost 0.03 sec Pass 4, pass_time_cost 0.03 sec Pass 5, pass_time_cost 0.03 sec Pass 6, pass_time_cost 0.03 sec Pass 7, pass_time_cost 0.03 sec Pass 8, pass_time_cost 0.03 sec share_vars_from is set, scope is ignored. I0303 20:15:29.968084 92 parallel_executor.cc:421] The number of CUDAPlace, which is used in ParallelExecutor, is 1. And the Program will be copied 1 copies I0303 20:15:29.968631 92 build_strategy.cc:363] SeqOnlyAllReduceOps:0, num_trainers:1 I0303 20:15:29.968953 92 parallel_executor.cc:285] Inplace strategy is enabled, when build_strategy.enable_inplace = True I0303 20:15:29.969275 92 parallel_executor.cc:368] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0 length=25 evaluation recall result: 1_in_2: 0.6 1_in_10: 0.12 2_in_10: 0.32 5_in_10: 0.68 Save model at step 10 ... 2020-03-03 20:15:33 Pass 9, pass_time_cost 3.89 sec Pass 10, pass_time_cost 0.03 sec Pass 11, pass_time_cost 0.03 sec Pass 12, pass_time_cost 0.03 sec Pass 13, pass_time_cost 0.03 sec Pass 14, pass_time_cost 0.03 sec Pass 15, pass_time_cost 0.03 sec Pass 16, pass_time_cost 0.03 sec Pass 17, pass_time_cost 0.03 sec Pass 18, pass_time_cost 0.03 sec length=25 evaluation recall result: 1_in_2: 0.6 1_in_10: 0.16 2_in_10: 0.28 5_in_10: 0.68 Save model at step 20 ... 2020-03-03 20:15:37 Pass 19, pass_time_cost 3.87 sec Pass 20, pass_time_cost 0.03 sec Pass 21, pass_time_cost 0.03 sec Pass 22, pass_time_cost 0.03 sec Pass 23, pass_time_cost 0.03 sec Pass 24, pass_time_cost 0.03 sec Pass 25, pass_time_cost 0.03 sec Pass 26, pass_time_cost 0.03 sec Pass 27, pass_time_cost 0.03 sec Pass 28, pass_time_cost 0.03 sec length=25 evaluation recall result: 1_in_2: 0.52 1_in_10: 0.16 2_in_10: 0.32 5_in_10: 0.64 Pass 29, pass_time_cost 0.04 sec Pass 30, pass_time_cost 0.03 sec Pass 31, pass_time_cost 0.03 sec Pass 32, pass_time_cost 0.03 sec Pass 33, pass_time_cost 0.03 sec Pass 34, pass_time_cost 0.03 sec Pass 35, pass_time_cost 0.03 sec Pass 36, pass_time_cost 0.03 sec Pass 37, pass_time_cost 0.03 sec Pass 38, pass_time_cost 0.03 sec length=25 evaluation recall result: 1_in_2: 0.52 1_in_10: 0.12 2_in_10: 0.2 5_in_10: 0.6 Pass 39, pass_time_cost 0.04 sec Pass 40, pass_time_cost 0.03 sec Pass 41, pass_time_cost 0.03 sec Pass 42, pass_time_cost 0.03 sec Pass 43, pass_time_cost 0.03 sec Pass 44, pass_time_cost 0.03 sec Pass 45, pass_time_cost 0.03 sec Pass 46, pass_time_cost 0.03 sec Pass 47, pass_time_cost 0.03 sec Pass 48, pass_time_cost 0.03 sec length=25 evaluation recall result: 1_in_2: 0.44 1_in_10: 0.08 2_in_10: 0.08 5_in_10: 0.56 training step 50 avg loss 0.0003279643505811691 Pass 49, pass_time_cost 0.04 sec
In[2]
#以召回率的形式对模型结果进行评估，也可以通过其他形式评估，具体参见run.sh文件
!cd auto_dialogue_evaluation && sh run.sh eval_recall----------- Configuration Arguments ----------- batch_size: 256 do_infer: False do_train: False do_val: True emb_size: 256 hidden_size: 256 init_model: model_files_tmp/matching_pretrained learning_rate: 0.001 loss_type: CLS max_len: 50 num_scan_data: 50 out_path: None print_step: 50 sample_pro: 1 save_path: tmp save_step: 10 test_path: data/unlabel_data/test.ids train_path: None use_cuda: True val_path: None vocab_size: 484016 word_emb_init: None ------------------------------------------------ W0303 20:16:00.303758 142 device_context.cc:236] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 9.2, Runtime API Version: 9.0 W0303 20:16:00.307708 142 device_context.cc:244] device: 0, cuDNN Version: 7.3. init model model_files_tmp/matching_pretrained len scores: 256 len labels: 256 length=25 mean score: -0.12624257184143062 evaluation recall result: 1_in_2: 0.48 1_in_10: 0.12 2_in_10: 0.24 5_in_10: 0.4 finish evaluate model:model_files_tmp/matching_pretrained on data:data/unlabel_data/test.ids time_cost(s):2.76
In[3]
#基于预训练模型和数据，可以运行下面的命令直接对对话数据进行打分
!cd auto_dialogue_evaluation && sh run.sh infer----------- Configuration Arguments ----------- batch_size: 256 do_infer: True do_train: False do_val: False emb_size: 256 hidden_size: 256 init_model: model_files_tmp/human_finetuned learning_rate: 0.001 loss_type: CLS max_len: 50 num_scan_data: 50 out_path: None print_step: 50 sample_pro: 1 save_path: tmp save_step: 10 test_path: data/label_data/human/test.ids train_path: None use_cuda: True val_path: None vocab_size: 484016 word_emb_init: None ------------------------------------------------ W0303 20:17:02.407306 180 device_context.cc:236] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 9.2, Runtime API Version: 9.0 W0303 20:17:02.411096 180 device_context.cc:244] device: 0, cuDNN Version: 7.3. finish infer model:model_files_tmp/human_finetuned out file: data/label_data/human/test.ids.infer time_cost(s):2.92
In[4]
#在第一阶段训练基础上，可利用少量标注数据进行第二阶段训练，进行finetune
!cd auto_dialogue_evaluation && sh run.sh finetune----------- Configuration Arguments ----------- batch_size: 256 do_infer: False do_train: True do_val: False emb_size: 256 hidden_size: 256 init_model: model_files_tmp/matching_pretrained learning_rate: 0.001 loss_type: L2 max_len: 50 num_scan_data: 50 out_path: None print_step: 1 sample_pro: 1 save_path: model_files_tmp/human_finetuned save_step: 1 test_path: None train_path: data/label_data/human/train.ids use_cuda: True val_path: data/label_data/human/val.ids vocab_size: 484016 word_emb_init: None ------------------------------------------------ context_rep shape: (-1L, 256L) response_rep shape: (-1L, 256L) logits shape: (-1L, 1L) begin memory optimization ... 2020-03-03 20:17:10,301-WARNING: Caution! paddle.fluid.memory_optimize() is deprecated and not maintained any more, since it is not stable! This API would not take any memory optimizations on your Program now, since we have provided default strategies for you. The newest and stable memory optimization strategies (they are all enabled by default) are as follows: 1. Garbage collection strategy, which is enabled by exporting environment variable FLAGS_eager_delete_tensor_gb=0 (0 is the default value). 2. Inplace strategy, which is enabled by setting build_strategy.enable_inplace=True (True is the default value) when using CompiledProgram or ParallelExecutor. end memory optimization ... context_rep shape: (-1L, 256L) response_rep shape: (-1L, 256L) logits shape: (-1L, 1L) device count 1 theoretical memory usage: (1519.720290184021, 1592.087923049927, 'MB') W0303 20:17:11.208235 218 device_context.cc:236] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 9.2, Runtime API Version: 9.0 W0303 20:17:11.211208 218 device_context.cc:244] device: 0, cuDNN Version: 7.3. Load pretraining parameters from model_files_tmp/matching_pretrained. sccuess init model_files_tmp/matching_pretrained start loading data ... I0303 20:17:14.004169 218 parallel_executor.cc:421] The number of CUDAPlace, which is used in ParallelExecutor, is 1. And the Program will be copied 1 copies I0303 20:17:14.006306 218 build_strategy.cc:363] SeqOnlyAllReduceOps:0, num_trainers:1 I0303 20:17:14.007922 218 parallel_executor.cc:285] Inplace strategy is enabled, when build_strategy.enable_inplace = True I0303 20:17:14.009212 218 parallel_executor.cc:368] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0 share_vars_from is set, scope is ignored. I0303 20:17:14.040747 218 parallel_executor.cc:421] The number of CUDAPlace, which is used in ParallelExecutor, is 1. And the Program will be copied 1 copies I0303 20:17:14.041477 218 build_strategy.cc:363] SeqOnlyAllReduceOps:0, num_trainers:1 I0303 20:17:14.041981 218 parallel_executor.cc:285] Inplace strategy is enabled, when build_strategy.enable_inplace = True I0303 20:17:14.042467 218 parallel_executor.cc:368] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0 evaluation cor relevance -0.045655154934289226 training step 1 avg loss 0.20214074850082397 Pass 0, pass_time_cost 0.06 sec evaluation cor relevance -0.03812301431872137 training step 2 avg loss 0.16594290733337402 Pass 1, pass_time_cost 0.04 sec evaluation cor relevance -0.03420949590753319 training step 3 avg loss 0.1503480225801468 Pass 2, pass_time_cost 0.04 sec evaluation cor relevance -0.026657388361296008 training step 4 avg loss 0.13701853156089783 Pass 3, pass_time_cost 0.04 sec evaluation cor relevance -0.026197380997029576 training step 5 avg loss 0.12259398400783539 Pass 4, pass_time_cost 0.04 sec evaluation cor relevance -0.01672844810653696 training step 6 avg loss 0.10547607392072678 Pass 5, pass_time_cost 0.04 sec evaluation cor relevance 0.02767570182543171 Save model at step 7 ... 2020-03-03 20:17:18 training step 7 avg loss 0.08713535219430923 Pass 6, pass_time_cost 3.88 sec evaluation cor relevance 0.05709543824894198 Save model at step 8 ... 2020-03-03 20:17:21 training step 8 avg loss 0.07621623575687408 Pass 7, pass_time_cost 3.85 sec evaluation cor relevance 0.08186671962085844 Save model at step 9 ... 2020-03-03 20:17:25 training step 9 avg loss 0.0698356181383133 Pass 8, pass_time_cost 3.86 sec evaluation cor relevance 0.06082311061313105 training step 10 avg loss 0.0569886639714241 Pass 9, pass_time_cost 0.04 sec evaluation cor relevance 0.0347839291437123 training step 11 avg loss 0.04764831066131592 Pass 10, pass_time_cost 0.04 sec evaluation cor relevance 0.0381713957276509 training step 12 avg loss 0.03954772651195526 Pass 11, pass_time_cost 0.04 sec evaluation cor relevance 0.02633714951171487 training step 13 avg loss 0.03392593562602997 Pass 12, pass_time_cost 0.04 sec evaluation cor relevance 0.02831387564797829 training step 14 avg loss 0.030515454709529877 Pass 13, pass_time_cost 0.04 sec evaluation cor relevance 0.03238175379241782 training step 15 avg loss 0.027258876711130142 Pass 14, pass_time_cost 0.04 sec evaluation cor relevance 0.0218399823578848 training step 16 avg loss 0.023921750485897064 Pass 15, pass_time_cost 0.04 sec evaluation cor relevance 0.02820405752929699 training step 17 avg loss 0.021650437265634537 Pass 16, pass_time_cost 0.04 sec evaluation cor relevance 0.03996381173465899 training step 18 avg loss 0.019501041620969772 Pass 17, pass_time_cost 0.04 sec evaluation cor relevance 0.04563058025038851 training step 19 avg loss 0.017170080915093422 Pass 18, pass_time_cost 0.04 sec evaluation cor relevance 0.05073597083076118 training step 20 avg loss 0.015909677371382713 Pass 19, pass_time_cost 0.12 sec evaluation cor relevance 0.046565186197487424 training step 21 avg loss 0.013923028483986855 Pass 20, pass_time_cost 0.04 sec evaluation cor relevance 0.05044875421267162 training step 22 avg loss 0.012706486508250237 Pass 21, pass_time_cost 0.04 sec evaluation cor relevance 0.053344727118596 training step 23 avg loss 0.011962947435677052 Pass 22, pass_time_cost 0.04 sec evaluation cor relevance 0.055481956659085944 training step 24 avg loss 0.011046133935451508 Pass 23, pass_time_cost 0.04 sec evaluation cor relevance 0.05437763180129775 training step 25 avg loss 0.010488426312804222 Pass 24, pass_time_cost 0.04 sec evaluation cor relevance 0.0553283648847065 training step 26 avg loss 0.010120008140802383 Pass 25, pass_time_cost 0.04 sec evaluation cor relevance 0.05274495123964428 training step 27 avg loss 0.009528218768537045 Pass 26, pass_time_cost 0.04 sec evaluation cor relevance 0.046217300828517986 training step 28 avg loss 0.009135609492659569 Pass 27, pass_time_cost 0.04 sec evaluation cor relevance 0.04565438697541733 training step 29 avg loss 0.008912760764360428 Pass 28, pass_time_cost 0.04 sec evaluation cor relevance 0.04975528735134844 training step 30 avg loss 0.00856935977935791 Pass 29, pass_time_cost 0.04 sec evaluation cor relevance 0.05003098458635954 training step 31 avg loss 0.008245025761425495 Pass 30, pass_time_cost 0.04 sec evaluation cor relevance 0.04752513478735894 training step 32 avg loss 0.008075003512203693 Pass 31, pass_time_cost 0.04 sec evaluation cor relevance 0.04078091997435762 training step 33 avg loss 0.007863759063184261 Pass 32, pass_time_cost 0.04 sec evaluation cor relevance 0.03466259164195254 training step 34 avg loss 0.007645013276487589 Pass 33, pass_time_cost 0.04 sec evaluation cor relevance 0.03227270363260842 training step 35 avg loss 0.007561428472399712 Pass 34, pass_time_cost 0.04 sec evaluation cor relevance 0.03320577366196353 training step 36 avg loss 0.007472529541701078 Pass 35, pass_time_cost 0.04 sec evaluation cor relevance 0.035791491183641444 training step 37 avg loss 0.007373311556875706 Pass 36, pass_time_cost 0.04 sec evaluation cor relevance 0.03418108142927299 training step 38 avg loss 0.007349514868110418 Pass 37, pass_time_cost 0.04 sec evaluation cor relevance 0.031033985972238214 training step 39 avg loss 0.007251190487295389 Pass 38, pass_time_cost 0.04 sec evaluation cor relevance 0.029633228989897703 training step 40 avg loss 0.00710717961192131 Pass 39, pass_time_cost 0.04 sec evaluation cor relevance 0.031149179803022797 training step 41 avg loss 0.00701714726164937 Pass 40, pass_time_cost 0.04 sec evaluation cor relevance 0.03398755579355489 training step 42 avg loss 0.006880040280520916 Pass 41, pass_time_cost 0.04 sec evaluation cor relevance 0.038649834104842865 training step 43 avg loss 0.006781894713640213 Pass 42, pass_time_cost 0.04 sec evaluation cor relevance 0.039020758239969214 training step 44 avg loss 0.006729709915816784 Pass 43, pass_time_cost 0.04 sec evaluation cor relevance 0.03814682104375019 training step 45 avg loss 0.006635408382862806 Pass 44, pass_time_cost 0.04 sec evaluation cor relevance 0.03727365180640306 training step 46 avg loss 0.0065619852393865585 Pass 45, pass_time_cost 0.04 sec evaluation cor relevance 0.03844171725055872 training step 47 avg loss 0.006483430974185467 Pass 46, pass_time_cost 0.04 sec evaluation cor relevance 0.03844402112717441 training step 48 avg loss 0.006373999174684286 Pass 47, pass_time_cost 0.04 sec evaluation cor relevance 0.039338693212934664 training step 49 avg loss 0.0063008773140609264 Pass 48, pass_time_cost 0.04 sec evaluation cor relevance 0.03943545603079371 training step 50 avg loss 0.0062194340862333775 Pass 49, pass_time_cost 0.04 sec
In[5]
#使用固化的模型进行预测，结果保存在out_path
#此处仅打印一例结果，分别为上文，回复，回复质量得分
!cd auto_dialogue_evaluation/ && python freeze_infer.py --test_path data/label_data/human/test.ids \
--model_path model_files_tmp/human_finetuned \
--use_cuda False \
--out_path infer_result.txt ----------- Configuration Arguments -----------
batch_size: 256
max_len: 50
model_path: model_files_tmp/human_finetuned
out_path: infer_result.txt
test_path: data/label_data/human/test.ids
use_cuda: 0
------------------------------------------------
('context_wordseq:', '3354 1604 1100 6564 97 694 326 11')
('response_wordseq:', '1604 48 6564 97 2 9 763 97')
('scores:', 1.5645232)
finish infer model:model_files_tmp/human_finetuned out file: infer_result.txt time_cost(s):3.80 点击链接，使用AI Studio一键上手实践项目吧：https://aistudio.baidu.com/aistudio/projectdetail/122301
下载安装命令
## CPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/cpu paddlepaddle
## GPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/gpu paddlepaddle-gpu>> 访问 PaddlePaddle 官网，了解更多相关内容。