├── GPT-2
  ├── README.md                              // MASS模型介绍
  ├── scripts
  │   ├──run_cbt.sh                          // CBT任务的微调&评估脚本
  │   ├──run_lambada.sh                      // LAMBADA任务的微调&评估脚本
  │   ├──run_language_model.sh               // 语言建模任务的微调&评估脚本
  │   ├──run_read_comprehension.sh           // 阅读理解任务的微调&评估脚本
  │   ├──run_summarization.sh                // 摘要生成任务的微调&评估脚本
  │   ├──run_translation.sh                  // 翻译任务的微调&评估脚本
  ├──src
  │   ├──clip_grad_utils.py                  // 用于梯度裁剪
  |   ├──dataset.py                          // 数据集加载用于微调或推理
  │   ├──finetune_eval_config.py             // 微调和推理配置文件
  │   ├──gpt2_for_finetune.py                // 用于梯度裁剪
  |   ├──GPT2_generation.py                  // 生成模块
  │   ├──GPT2_model.py                       // GPT2模型脚本
  │   ├──GPT2ForCBT.py                       // CBT任务的模型脚本
  │   ├──GPT2ForLanguageModel.py             // 语言建模任务的模型脚本
  │   ├──GPT2ForReadComprehension.py         // 阅读理解任务的模型脚本
  │   ├──GPT2ForSummarization.py             // 摘要生成任务的模型脚本
  │   ├──GPT2ForTranslation.py               // 翻译任务的模型脚本
  │   ├──weight_init.py                      // 初始化权重
  │   ├──utils
  │      ├──bleu_score.py                    // 用于计算BLEU分数
  │      ├──rouge_score.py                   // 用于计算ROUGE分数
  │      ├──CrossEntropy.py                  // 交叉熵损失
  │      ├──data_preprocess.py               // 数据集预处理脚本
  │      ├──generation_utils.py              // 用于帮助生成模型，包含采样等方法
  │      ├──get_config_setting.py            // 获取配置信息
  │      ├──task_utils.py                    // 辅助下游任务的功能脚本
  │      ├──lr_schedule.py                   // 学习率策略脚本
  │      ├──metric_method.py                 // 下游任务的评价指标
  │      ├──tensor_manipulations.py          // 涉及张量操作
  │      ├──tokenization.py                  // 标记化，包含BPE编码和解码
  │      ├──pretrain-data
  │          ├──stopwords.txt                // 用于LAMBADA任务的stopword filter
  ├──create_cbt_data.py                      // 用于CBT任务创建mindrecord
  ├──create_lambada_data.py                  // 用于lambada任务创建mindrecord
  ├──create_lambada_data.py                  // 用于其他任务创建mindrecord
  ├──create_summary_data.py                  // 用于summarization任务创建mindrecord
  ├──download_cnn_dailymail.py               // 下载CNN & Dailymail数据集
  ├──cnn_dataset_sampler.py                  // CNN & Dailymail训练集采样器
  ├──eval_rc_addition_answer.py              // 使用addition_answer评估阅读理解任务
  ├──run_CBT_task.py                         // CBT任务微调&推理API入口
  ├──run_lambada.py                          // LAMBADA任务微调&推理API入口
  ├──run_language_mdoel.py                   // 语言建模任务微调&推理API入口
  ├──run_ReadComprehension.py                // 阅读理解任务微调&推理API入口
  ├──run_summarization.py                    // 摘要生成任务微调&推理API入口
  ├──run_translation.py                      // 翻译任务微调&推理API入口
  ├──task_dataset_preprocess.py              // 各个任务的数据集处理入口
  ├──convert_tf_ckpt
  │      ├──read_weight_tf.py                // 读取tensorflow下的预训练模型
  │      ├──trans_dict.py                    // 模型参数名称字典
  │      ├──save_weight_ms.py                // 生成mindspore ckpt
  ├──third_party
  │      ├──gpt2-merges.txt
  │      ├──gpt2-vocab.json                  // GPT-2预训练词表
  │      ├──bleu.py                          // 辅助bleu值计算的第三方代码

--task:          The GPT-2 downstream task, including [LanguageModeling, CBT, Translation, Lambada, Summarization, ReadingComprehension].
--input_file:    The raw dataset path.
--dataset:       The name of dataset which should be processed, only for LanguageModeling task.
--output_file:   The output dataset path after preprocessing.
--condition:     Process train or test dataset, including [train, test], only for 1BW and CNN & DailyMail dataset.

python task_dataset_preprocess.py --task "LanguageModeling" --input_file /{path}/ptb.test.txt --dataset "ptb" --output_file /{path}/ptb_clean_test.txt --condition "test"

--input_file:      Input raw text file.
--output_file:     Output MindRecord file.
--num_splits:      The MindRecord file will be split into the number of partition.
--max_seq_length:  Maximum sequence length.
--vocab_file:      url of gpt2-vocab.json.
--merge_file:      url of gpt2-merges.txt

python create_lm_data.py --input_file /{path}/ptb.test.txt --output_file /{path}/ptb-test-mindrecord --num_splits 1 --max_length 1024 --vocab_file={path} --merge_file={path}

python task_dataset_preprocess.py --task "CBT" --input_file /{path}/cbtest_CN_valid_2000ex.txt --dataset "cbt" --output_file /{path}/cbt_cn_valid.txt

--input_file:      Input raw text file.
--output_file:     Output MindRecord file.
--num_splits:      The MindRecord file will be split into the number of partition.
--max_seq_length:  Maximum sequence length.
--num_choice:      Number of choices.
--vocab_file:      url of gpt2-vocab.json.
--merge_file:      url of gpt2-merges.txt

python create_cbt_data.py --input_file /{path}/cbt_cn_valid.txt --output_file /{path}/cbt-cn-valid-mindrecord --num_splits 1 --max_length 1024 --num_choice 10 --vocab_file={path} --merge_file={path}

python task_dataset_preprocess.py --task "LAMBADA" --input_file /{path}/lambada_test_plain_text.txt --dataset "LAMBADA" --output_file /{path}/lambada_test_clean.txt

--input_file:      Input raw text file.
--output_file:     Output MindRecord file.
--num_splits:      The MindRecord file will be split into the number of partition.
--max_seq_length:  Maximum sequence length.
--vocab_file:      url of gpt2-vocab.json.
--merge_file:      url of gpt2-merges.txt

python create_lambada_data.py --input_file /{path}/lambada_test_clean.txt --output_file /{path}/lambada-test-mindrecord --num_splits 1 --max_length 1024 --vocab_file={path} --merge_file={path}

python task_dataset_preprocess.py --task "ReadingComprehension" --input_file /{path}/coqa-dev-v1.0.json --dataset "coqa" --output_file /{path}/coqa_dev.txt

python create_lm_data.py --input_file /{path}/coqa_dev.txt --output_file /{path}/coqa-dev-mindrecord --num_splits 1 --max_length 1024 --vocab_file={path} --merge_file={path}

下载测试集
python download_cnn_dailymail.py --dir ./cnn_dailymail/ --split test

下载训练集
python download_cnn_dailymail.py --dir ./cnn_dailymail/ --split train

GPT-2 small和GPT-2 medium模型的训练集中seq_length=1024, 因此该脚本中设置max_length=1022
python cnn_dataset_sampler.py --input_path="/{path}/cnn_train.txt"
                              --output_path="/{path}/cnn_train_hint_small.txt"
                              --replace_hint="true"
                              --sample="true"
                              --max_length=1022
                              --prob=0.25
                              --max_items=10000
                              --hint="TL;DR:"


GPT-2 large模型的训练集中seq_length=768,因此该脚本中设置max_length=766
python cnn_dataset_sampler.py --input_path="/{path}/cnn_train.txt"
                              --output_path="/{path}/cnn_train_hint_large.txt"
                              --replace_hint="true"
                              --sample="true"
                              --max_length=766
                              --prob=0.25
                              --max_items=10000
                              --hint="TL;DR:"

python create_summary_data.py --input_file /{path}/cnn_dailymail_test.txt --output_file /{path}/cnn_dailymail-test-mindrecord --num_splits 1 --max_length 1024 --vocab_file={path} --merge_file={path} --mode 'cnn_dailymail'

python task_dataset_preprocess.py --task "Translation" --input_file /{path}/test-full --dataset "wmt14" --output_file /{path}/wmt14

python create_lm_data.py --input_file /{path}/wmt14.en_fr.txt --output_file /{path}/en-fr-mindrecord --num_splits 1 --max_length 1024 --vocab_file={path} --merge_file={path}

python create_lm_data.py --input_file /{path}/wmt14.fr_en.txt --output_file /{path}/fr-en-mindrecord --num_splits 1 --max_length 1024 --vocab_file={path} --merge_file={path}

bash scripts/run_language_model.sh   --device_target="Ascend"
                                   --do_train="true"
                                   --do_eval="false"
                                   --epoch_num=1
                                   --train_data_shuffle="true"
                                   --eval_data_shuffle="false"
                                   --save_finetune_ckpt_path={save_finetune_ckpt_path}
                                   --load_pretrain_ckpt_path={load_pretrain_ckpt_path}
                                   --train_data_file_path={train_data_file_path}

bash scripts/run_language_model.sh [--options]

usage: run_language_model.sh   [--device_target DEVICE_TARGET] [--device_id N]
                               [--metric_method METRIC_METHOD]
                               [--do_train DO_TRAIN] [--do_eval DO_EVAL]
                               [--eval_type EVAL_TYPE] [--epoch_num N]
                               [--train_data_shuffle TRAIN_DATA_SHUFFLE]
                               [--eval_data_shuffle EVAL_DATA_SHUFFLE]
                               [--save_finetune_ckpt_path SAVE_FINETUNE_CKPT_PATH]
                               [--load_pretrain_ckpt_path LOAD_PRETRAIN_CKPT_PATH]
                               [--load_finetune_ckpt_path LOAD_FINETUNE_CKPT_PATH]
                               [--train_data_file_path TRAIN_DATA_FILE_PATH]
                               [--eval_data_file_path EVAL_DATA_FILE_PATH]
options:
    --device_target                   Device type. Default: "Ascend"
    --device_id                       ID of target device
    --metric_method                   The eval method including [PPL]. Default: "PPL"
    --do_train                        Enable train. Default: "false"
    --do_eval                         Enable evaluation. Default: "true"
    --eval_type                       The type of evaluation including [zero-shot, finetuned]. Default: "zero-shot"
    --epoch_num                       Epoch number. Default: 1
    --train_data_shuffle              Enable train data shuffle. Default: "true"
    --eval_data_shuffle               Enable eval data shuffle. Default: "false"
    --save_finetune_ckpt_path         Save the finetuned checkpoint path
    --load_pretrain_ckpt_path         Load the checkpoint file path for train
    --load_finetune_ckpt_path         Load the checkpoint file path for evaluation
    --train_data_file_path            Data path, it is better to use absolute path
    --eval_data_file_path             Data path, it is better to use absolute path

bash scripts/run_language_model.sh   --device_target="Ascend"
                                   --metric_method="PPL"
                                   --do_train="false"
                                   --do_eval="true"
                                   --eval_type="finetuned"
                                   --train_data_shuffle="true"
                                   --eval_data_shuffle="false"
                                   --load_finetune_ckpt_path={load_eval_ckpt_path}
                                   --eval_data_file_path={eval_data_file_path}

bash scripts/run_cbt.sh   --device_target="Ascend"
                        --num_choice=10
                        --metric_method="Accuarcy"
                        --do_train="false"
                        --do_eval="true"
                        --eval_type="zero-shot"
                        --train_data_shuffle="true"
                        --eval_data_shuffle="false"
                        --load_finetune_ckpt_path={load_eval_ckpt_path}
                        --eval_data_file_path={eval_data_file_path}

bash scripts/run_cbt.sh [--options]

usage: run_CBT_task.sh   [--device_target DEVICE_TARGET] [--device_id N][--num_choice N]
                         [--metric_method METRIC_METHOD]
                         [--do_train DO_TRAIN] [--do_eval DO_EVAL]
                         [--eval_type EVAL_TYPE] [--epoch_num N]
                         [--train_data_shuffle TRAIN_DATA_SHUFFLE]
                         [--eval_data_shuffle EVAL_DATA_SHUFFLE]
                         [--save_finetune_ckpt_path SAVE_FINETUNE_CKPT_PATH]
                         [--load_pretrain_ckpt_path LOAD_PRETRAIN_CKPT_PATH]
                         [--load_finetune_ckpt_path LOAD_FINETUNE_CKPT_PATH]
                         [--train_data_file_path TRAIN_DATA_FILE_PATH]
                         [--eval_data_file_path EVAL_DATA_FILE_PATH]
options:
    --device_target                   Device type. Default: "Ascend"
    --device_id                       ID of target device
    --num_choice                      The number of choice in CBT task
    --metric_method                   The eval method including [Accuracy]. Default: "Accuracy"
    --do_train                        Enable train. Default: "false"
    --do_eval                         Enable evaluation. Default: "true"
    --eval_type                       The type of evaluation including [zero-shot, finetuned]. Default: "zero-shot"
    --epoch_num                       Epoch number. Default: 1
    --train_data_shuffle              Enable train data shuffle. Default: "true"
    --eval_data_shuffle               Enable eval data shuffle. Default: "false"
    --save_finetune_ckpt_path         Save the finetuned checkpoint path
    --load_pretrain_ckpt_path         Load the checkpoint file path for train
    --load_finetune_ckpt_path         Load the checkpoint file path for evaluation
    --train_data_file_path            Data path, it is better to use absolute path
    --eval_data_file_path             Data path, it is better to use absolute path

bash scripts/run_lambada.sh   --device_target="Ascend"
                            --metric_method="Accuarcy"
                            --do_train="false"
                            --do_eval="true"
                            --eval_type="zero-shot"
                            --train_data_shuffle="true"
                            --eval_data_shuffle="false"
                            --generate_length_dynamically="true"
                            --load_finetune_ckpt_path={load_eval_ckpt_path}
                            --eval_data_file_path={eval_data_file_path}
                            --tokenizer_file_path={tokenizer_file_path}
                            --stop_word_file_path={stop_word_file_path}

bash scripts/run_lambada.sh   --device_target="Ascend"
                            --metric_method="PPL"
                            --do_train="false"
                            --do_eval="true"
                            --eval_type="zero-shot"
                            --train_data_shuffle="true"
                            --eval_data_shuffle="false"
                            --load_finetune_ckpt_path={load_eval_ckpt_path}
                            --eval_data_file_path={eval_data_file_path}

bash scripts/run_lambada.sh [--options]

usage: run_lambada.sh   [--device_target DEVICE_TARGET] [--device_id N]
                        [--metric_method METRIC_METHOD]
                        [--do_train DO_TRAIN] [--do_eval DO_EVAL]
                        [--eval_type EVAL_TYPE] [--epoch_num N]
                        [--train_data_shuffle TRAIN_DATA_SHUFFLE]
                        [--eval_data_shuffle EVAL_DATA_SHUFFLE]
                        [--generate_length_dynamically GENERATE_LENGTH_DYNAMICALLY]
                        [--save_finetune_ckpt_path SAVE_FINETUNE_CKPT_PATH]
                        [--load_pretrain_ckpt_path LOAD_PRETRAIN_CKPT_PATH]
                        [--load_finetune_ckpt_path LOAD_FINETUNE_CKPT_PATH]
                        [--train_data_file_path TRAIN_DATA_FILE_PATH]
                        [--eval_data_file_path EVAL_DATA_FILE_PATH]
                        [--tokenizer_file_path TOKENIZER_FILE_PATH]
                        [--stop_word_file_path STOP_WORD_FILE_PATH]
options:
    --device_target                   Device type. Default: "Ascend"
    --device_id                       ID of target device
    --metric_method                   The eval method including [Accuracy, PPL]. Default: "Accuracy"
    --do_train                        Enable train. Default: "false"
    --do_eval                         Enable evaluation. Default: "true"
    --eval_type                       The type of evaluation including [zero-shot, finetuned]. Default: "zero-shot"
    --epoch_num                       Epoch number. Default: 1
    --train_data_shuffle              Enable train data shuffle. Default: "true"
    --eval_data_shuffle               Enable eval data shuffle. Default: "false"
    --generate_length_dynamically     Enable generate_length_Dynamically. Default: "true"
    --save_finetune_ckpt_path         Save the checkpoint path
    --load_pretrain_ckpt_path         Load the checkpoint file path
    --load_finetune_ckpt_path         Load the checkpoint file path
    --train_data_file_path            Data path, it is better to use absolute path
    --eval_data_file_path             Data path, it is better to use absolute path
    --tokenizer_file_path             pretrained vocab and merge file path
    --stop_word_file_path             The stop word file path

bash scripts/run_read_comprehension.sh   --device_target="Ascend"
                                       --metric_method="F1"
                                       --do_train="false"
                                       --do_eval="true"
                                       --eval_type="zero-shot"
                                       --train_data_shuffle="true"
                                       --eval_data_shuffle="false"
                                       --load_finetune_ckpt_path={load_eval_ckpt_path}
                                       --eval_data_file_path={eval_data_file_path}
                                       --tokenizer_file_path={tokenizer_file_path}
                                       --generate_length=55
                                       --top_k=1
                                       --top_p="1.0"
                                       --temperature="1.0"

bash scripts/run_read_comprehension.sh [--options]

usage: run_read_comprehension.sh   [--device_target DEVICE_TARGET] [--device_id N]
                                   [--metric_method METRIC_METHOD]
                                   [--do_train DO_TRAIN] [--do_eval DO_EVAL]
                                   [--eval_type EVAL_TYPE] [--epoch_num N]
                                   [--train_data_shuffle TRAIN_DATA_SHUFFLE]
                                   [--eval_data_shuffle EVAL_DATA_SHUFFLE]
                                   [--save_finetune_ckpt_path SAVE_FINETUNE_CKPT_PATH]
                                   [--load_pretrain_ckpt_path LOAD_PRETRAIN_CKPT_PATH]
                                   [--load_finetune_ckpt_path LOAD_FINETUNE_CKPT_PATH]
                                   [--train_data_file_path TRAIN_DATA_FILE_PATH]
                                   [--eval_data_file_path EVAL_DATA_FILE_PATH]
                                   [--tokenizer_file_path TOKENIZER_FILE_PATH]
                                   [--generate_length N] [--top_k N] [--top_p TOP_P]
                                   [--temperature TEMPERATURE]
options:
    --device_target                   Device type. Default: "Ascend"
    --device_id                       ID of target device
    --metric_method                   The eval method including [F1]. Default: "F1"
    --do_train                        Enable train. Default: "false"
    --do_eval                         Enable evaluation. Default: "false"
    --eval_type                       The type of evaluation including [zero-shot, finetuned]. Default: "zero-shot"
    --epoch_num                       Epoch number. Default: 1
    --train_data_shuffle              Enable train data shuffle. Default: "true"
    --eval_data_shuffle               Enable eval data shuffle. Default: "false"
    --save_finetune_ckpt_path         Save the checkpoint path
    --load_pretrain_ckpt_path         Load the checkpoint file path
    --load_finetune_ckpt_path         Load the checkpoint file path
    --train_data_file_path            Data path, it is better to use absolute path
    --eval_data_file_path             Data path, it is better to use absolute path
    --tokenizer_file_path             pretrained vocab and merge file path
    --generate_length                 The generation length of answer sentence
    --top_k                           Parameter for Top-K sampling
    --top_p                           Parameter for Top-P sampling
    --temperature                     Parameter for generation, greater if generation more diverse

bash scripts/run_summarization.sh   --device_target="Ascend"
                                  --do_train="true"
                                  --do_eval="false"
                                  --metric_method="Rouge"
                                  --train_data_shuffle="true"
                                  --eval_data_shuffle="false"
                                  --generate_length=100
                                  --top_k=2
                                  --top_p="1.0"
                                  --temperature="1.0"
                                  --eval_type="finetuned"
                                  --load_finetune_ckpt_path={load_eval_ckpt_path}
                                  --eval_data_file_path={eval_data_file_path}
                                  --tokenizer_file_path={tokenizer_file_path}

bash scripts/run_summarization.sh   --device_target="Ascend"
                                  --do_train="false"
                                  --do_eval="true"
                                  --metric_method="Rouge"
                                  --train_data_shuffle="true"
                                  --eval_data_shuffle="false"
                                  --generate_length=100
                                  --top_k=2
                                  --top_p="1.0"
                                  --temperature="1.0"
                                  --eval_type="finetuned"
                                  --load_finetune_ckpt_path={load_eval_ckpt_path}
                                  --eval_data_file_path={eval_data_file_path}
                                  --tokenizer_file_path={tokenizer_file_path}

bash scripts/run_summarization.sh [--options]

usage: run_summarization.sh   [--device_target DEVICE_TARGET] [--device_id N][--num_choice N]
                              [--metric_method METRIC_METHOD]
                              [--do_train DO_TRAIN] [--do_eval DO_EVAL]
                              [--eval_type EVAL_TYPE] [--epoch_num N]
                              [--train_data_shuffle TRAIN_DATA_SHUFFLE]
                              [--eval_data_shuffle EVAL_DATA_SHUFFLE]
                              [--save_finetune_ckpt_path SAVE_FINETUNE_CKPT_PATH]
                              [--load_pretrain_ckpt_path LOAD_PRETRAIN_CKPT_PATH]
                              [--load_finetune_ckpt_path LOAD_FINETUNE_CKPT_PATH]
                              [--train_data_file_path TRAIN_DATA_FILE_PATH]
                              [--eval_data_file_path EVAL_DATA_FILE_PATH]
options:
    --device_target                   Device type. Default: "Ascend"
    --device_id                       ID of target device
    --do_train                        Enable train. Default: false.
    --do_eval                         Enable evaluation. Default: false.
    --metric_method                   The eval method including [Rouge(Rouge1,Rouge2,RougeL,Rouge Avg)]. Default: Rouge. Default: "false"
    --epoch_num                       Epoch number. Default: 2.
    --train_data_shuffle              Enable train data shuffle. Default: true.
    --eval_data_shuffle               Enable eval data shuffle. Default: false.
    --save_finetune_ckpt_path         Save the checkpoint path.
    --load_pretrain_ckpt_path         Load the checkpoint file path.
    --load_finetune_ckpt_path         Load the checkpoint file path.
    --train_data_file_path            Data path, it is better to use absolute path.
    --eval_data_file_path             Data path, it is better to use absolute path.
    --eval_type                       The type of evaluation including [zero-shot, finetuned]. Default: zero-shot.
    --top_k                           Top k tokens chosen for sampling.
    --top_p                           Top p accumulated probability threshold for logit to be counted.
    --generate_length                 The number of generated tokens.
    --temperature                     Temperature on logits for sampling.
    --tokenizer_file_path             Vocab & merge file path.

bash scripts/run_translation.sh   --device_target="Ascend"
                                --metric_method="BLEU"
                                --do_train="false"
                                --do_eval="true"
                                --eval_type="zero-shot"
                                --train_data_shuffle="true"
                                --eval_data_shuffle="false"
                                --load_finetune_ckpt_path={load_eval_ckpt_path}
                                --eval_data_file_path={eval_data_file_path}
                                --tokenizer_file_path={tokenizer_file_path}
                                --generate_length=100
                                --top_k=1
                                --top_p="1.0"
                                --temperature="1.0"

bash scripts/run_translation.sh [--options]

usage: run_translation.sh   [--device_target DEVICE_TARGET] [--device_id N]
                            [--metric_method METRIC_METHOD]
                            [--do_train DO_TRAIN] [--do_eval DO_EVAL]
                            [--eval_type EVAL_TYPE] [--epoch_num N]
                            [--train_data_shuffle TRAIN_DATA_SHUFFLE]
                            [--eval_data_shuffle EVAL_DATA_SHUFFLE]
                            [--save_finetune_ckpt_path SAVE_FINETUNE_CKPT_PATH]
                            [--load_pretrain_ckpt_path LOAD_PRETRAIN_CKPT_PATH]
                            [--load_finetune_ckpt_path LOAD_FINETUNE_CKPT_PATH]
                            [--train_data_file_path TRAIN_DATA_FILE_PATH]
                            [--eval_data_file_path EVAL_DATA_FILE_PATH]
                            [--tokenizer_file_path TOKENIZER_FILE_PATH]
                            [--generate_length N] [--top_k N] [--top_p TOP_P]
                            [--temperature TEMPERATURE]
options:
    --device_target                   Device type. Default: "Ascend"
    --device_id                       ID of target device
    --metric_method                   The eval method including [BLEU]. Default: "BLEU"
    --do_train                        Enable train. Default: "false"
    --do_eval                         Enable evaluation. Default: "true"
    --eval_type                       The type of evaluation including [zero-shot, finetuned]. Default: "zero-shot"
    --epoch_num                       Epoch number. Default: 1
    --train_data_shuffle              Enable train data shuffle. Default: "true"
    --eval_data_shuffle               Enable eval data shuffle. Default: "false"
    --save_finetune_ckpt_path         Save the checkpoint path
    --load_pretrain_ckpt_path         Load the checkpoint file path
    --load_finetune_ckpt_path         Load the checkpoint file path
    --train_data_file_path            Data path, it is better to use absolute path
    --eval_data_file_path             Data path, it is better to use absolute path
    --tokenizer_file_path             pretrained vocab and merge file path
    --generate_length                 The generation length of translation sentence
    --top_k                           Parameter for Top-K sampling
    --top_p                           Parameter for Top-P sampling
    --temperature                     Parameter for generation, greater if generation more diverse

math
numpy
copy
collections
re
rouge 1.0.0
datasets >=0.4.0
json
tensorflow