(升级)模型蒸馏-分类任务

.
├── data
│   ├── download_general_data.sh
│   ├── download_task_data.sh
│   ├── predict_demo
│   ├── wiki_wenxin
│   └── iflytek
├── examples
│   ├── iflytek
├── inference
│   ├── custom_inference.py
├── model
│   ├── base_cls.py
│   ├── ernie_classification.py
│   ├── general_enhanced_distill_minilm.py
│   ├── task_distill_minilm.py
├── reader
│   ├── cls_field_reader.py
│   ├── ernie_classification_dataset_reader.py
├── run_infer.py
├── run_trainer.py
└── trainer
    ├── custom_dynamic_trainer.py

 cd ./wenxin_appzoo/tasks/model_distillation_cls/data
 bash download_task_data.sh
 bash download_general_data.sh
 ```

### 分阶段运行

1.  准备预训练教师模型和学生模型

 ```shell

 cd wenxin_appzoo/models_hub
 bash download_ernie_3.0_large_distill_ch.sh
 bash download_ernie_3.0_tiny_ch.sh
 cd wenxin_appzoo/tasks/model_distillation_cls

  ```

2.  fine-tune 教师模型

 ```shell

 python run_trainer.py --param_path ./examples/iflytek/ernie_3.0_large_ft.json

  ```

3.  任务蒸馏通用数据增强阶段：GED

 - 首先需要根据评估指标选择最优的fine-tuned教师模型，然后运行：

 ```shell

 python run_trainer.py --param_path ./examples/iflytek/minilm_ged.json

 ```

4.  任务蒸馏二阶段：TD1

 ```shell

 python run_trainer.py --param_path ./examples/iflytek/minilm_td1.json

 ```

5.  任务蒸馏二阶段：TD2

```shell

 python run_trainer.py --param_path ./examples/iflytek/minilm_td2.json