AiSpace provides highly configurable framework for deep learning model development, deployment and conveniently use of pre-trained models (bert, albert, opt, etc.).

Table of Contents

• Features
• Requirements
• Instructions
• Configuration
• Pretrained
• Experiments
• Todos
• Refs

Features

• Highly configurable, we manage all hyperparameters with inheritable Configuration files.
• All modules are registerable, including models, dataset, losses, optimizers, metrics, callbacks, etc.
• Standardized process
• Multi-GPU Training
• K-fold cross validation training
• Integrate lr finder
• Integrate multiple pre-trained models, including chinese
• Simple and fast deployment using BentoML
• Integrated Chinese benchmarks CLUE

Requirements

git clone https://github.com/yingyuankai/AiSpace.git
cd AiSpace && pip install -r requirements

Instructions

Training

python -u aispace/trainer.py \
    --schedule train_and_eval \
    --config_name CONFIG_NAME \
    --config_dir CONFIG_DIR \
    [--experiment_name EXPERIMENT_NAME] \
    [--model_name MODEL_NAME] \
    [--gpus GPUS] 

Output file structure

The default output path is save, which may has multiple output directories under name as:

{experiment_name}_{model_name}_{dataset_name}_{random_seed}_{id}

Where id indicates the sequence number of the experiment for the same task, increasing from 0.

Take the text classification task as an example, the output file structure is similar to the following:

experiment_name: test

model_name: bert_for_classification

dataset_name: glue_zh/tnews

random_seed: 119

id: 0

test_bert_for_classification_glue_zh__tnews_119_0
├── checkpoint                  # 1. checkpoints
│   ├── checkpoint
│   ├── ckpt_1.data-00000-of-00002
│   ├── ckpt_1.data-00001-of-00002
│   ├── ckpt_1.index
|   ...
├── deploy                      # 2. Bentoml depolyment directory
│   └── BertTextClassificationService
│       └── 20191208180211_B6FC81
├── hparams.json                # 3. Json file of all hyperparameters
├── logs                        # 4. general or tensorboard log directory
│   ├── errors.log              # error log file
│   ├── info.log                # info log file
│   ├── train                
│   │   ├── events.out.tfevents.1574839601.jshd-60-31.179552.14276.v2
│   │   ├── events.out.tfevents.1574839753.jshd-60-31.profile-empty
│   └── validation
│       └── events.out.tfevents.1574839787.jshd-60-31.179552.151385.v2
├── model_saved                 # 5. last model saved
│   ├── checkpoint
│   ├── model.data-00000-of-00002
│   ├── model.data-00001-of-00002
│   └── model.index
└── reports                     # 6. Eval reports for every output or task
    └── output_1_classlabel     # For example, text classification task
        ├── confusion_matrix.txt
        ├── per_class_stats.json
        └── stats.json

Training with resumed model

python -u aispace/trainer.py \
    --schedule train_and_eval \
    --config_name CONFIG_NAME \
    --config_dir CONFIG_DIR \
    --model_resume_path MODEL_RESUME_PATH \
    [--experiment_name EXPERIMENT_NAME] \
    [--model_name MODEL_NAME] \
    [--gpus GPUS] 

--model_resume_path is a path to initialization model.

lr finder

Firstly, use optimizer adma and open lr_finder callback.

policy:
    name: "base"
        
optimizer:
  name: adam
    
callbacks:
    lr_finder:
      switch: true

Then run training policy as base.

Lastly, you can find lr_finder.jpg in you workspace.

K-fold cross validation training

Firstly, Replace training default policy form base to:

training:
  policy:
    name: "k-fold"
    config:
      k: 5

The k is the number of fold. Your can refer to the configuration file in:

./confis/glue_zh/tnews_k_fold.yml

Then run training script as usual.

Average checkpoints

python -u aispace/trainer.py \
    --schedule avg_checkpoints \
    --config_name CONFIG_NAME \
    --config_dir CONFIG_DIR \
    --prefix_or_checkpoints PREFIX_OR_CHECKPOINGS \
    [--ckpt_weights CKPT_WEIGHTS] \
    [--experiment_name EXPERIMENT_NAME] \
    [--model_name MODEL_NAME] \
    [--gpus GPUS] 

--prefix_or_checkpoints is paths to multiple checkpoints separated by comma.

--ckpt_weights is weights same order as the prefix_or_checkpoints.

Deployment

Generate deployment files before deployment, you need to specify the model path (--model_resume_path) to be deployed like following.

python -u aispace/trainer.py \
    --schedule deploy \
    --config_name CONFIG_NAME \
    --config_dir CONFIG_DIR \
    --model_resume_path MODEL_RESUME_PATH \
    [--experiment_name EXPERIMENT_NAME] \
    [--model_name MODEL_NAME] \
    [--gpus GPUS] 

We use BentoML as deploy tool, so your must implement the deploy function in your model class.

Configuration

All the configurations are in configs, in which base (./configs/default/base.yml) is the most basic, any configuration downstream includes this configuration directly or indirectly. Before you start, it is best to read this configuration carefully.

Your can use includes field to load other configurations, then the current configuration inherits the configurations and overrides the same configuration fields. Just like class inheritance, a function of the same name in a subclass overrides a function of the parent class.

The syntax is like this:

merge configuration of bert_huggingface into current.

includes:
  - "../pretrain/bert_huggingface.yml"     # relative path

Datasets

Dataset	Info	Ref
glue_zh/afqmc	Ant Financial Question Matching Corpus(蚂蚁金融语义相似度)	https://github.com/CLUEbenchmark/CLUE
glue_zh/tnews	TNEWS 今日头条中文新闻（短文）分类	https://github.com/CLUEbenchmark/CLUE
glue_zh/iflytek	IFLYTEK' 长文本分类	https://github.com/CLUEbenchmark/CLUE
glue_zh/cmnli	CMNLI 语言推理任务	https://github.com/CLUEbenchmark/CLUE
glue_zh/copa	COPA 因果推断-中文版	https://github.com/CLUEbenchmark/CLUE
glue_zh/wsc	WSC Winograd模式挑战中文版	https://github.com/CLUEbenchmark/CLUE
glue_zh/csl	CSL 论文关键词识别	https://github.com/CLUEbenchmark/CLUE
glue_zh/cmrc2018	Reading Comprehension for Simplified Chinese 简体中文阅读理解任务	https://github.com/CLUEbenchmark/CLUE
glue_zh/drcd	繁体阅读理解任务	https://github.com/CLUEbenchmark/CLUE
glue_zh/chid	成语阅读理解填空 Chinese IDiom Dataset for Cloze Test	https://github.com/CLUEbenchmark/CLUE
glue_zh/c3	中文多选阅读理解	https://github.com/CLUEbenchmark/CLUE
Dureader/robust	首个关注阅读理解模型鲁棒性的中文数据集	https://aistudio.baidu.com/aistudio/competition/detail/49
Dureader/yesno	一个以观点极性判断为目标任务的数据集	https://aistudio.baidu.com/aistudio/competition/detail/49
LSTC_2020/DuEE_trigger	从自然语言文本中抽取事件并识别事件类型	https://aistudio.baidu.com/aistudio/competition/detail/32
LSTC_2020/DuEE_role	从自然语言文本中抽取事件元素	https://aistudio.baidu.com/aistudio/competition/detail/32

Pretrained

We have integrated multiple pre-trained language models and are constantly expanding。

Model	#Model	#Chinese model	Download manually？	Refs	Status
bert	13	1	no	transformers	Done
albert	8	0	no	transformers	Done
albert_chinese	9	9	yes	albert_zh	Done
bert_wwm	4	4	yes	Chinese-BERT-wwm	Done
xlnet	2	0	no	transformers	Processing
xlnet_chinese	2	2	yes	Chinese-PreTrained-XLNets	Done
ernie	4	2	yes	ERNIE	Done
NEZHA	4	4	yes	NEZHA	Done
TinyBERT	-	-	-	TinyBERT	Processing
electra_chinese	4	4	yes	Chinese-ELECTR	Done

For those models that need to be downloaded manually, download, unzip them and modify the path in the corresponding configuration.

Some pre-trained models don't have tensorflow versions, I converted them and made them available for download。

Model	Refs	tf version
ERNIE_Base_en_stable-2.0.0	ERNIE	baidu yun
ERNIE_stable-1.0.1	ERNIE	baidu yun
ERNIE_1.0_max-len-512	ERNIE	baidu yun
ERNIE_Large_en_stable-2.0.0	ERNIE	baidu yun

Examples

We have implemented some tasks in the CLUE (The Chinese General Language Understanding Evaluation (GLUE) benchmark).

Please refer to Examples Doc.

Take glue_zh/tnews as an example:

Tnews is a task of Chinese GLUE, which is a short text classification task from ByteDance.

Run Tnews classification

python -u aispace/trainer.py \
    --experiment_name test \
    --model_name bert_for_classification \
    --schedule train_and_eval \
    --config_name tnews \
    --config_dir ./configs/glue_zh \
    --gpus 0 1 2 3  \

Specify different pretrained model, please change includes and pretrained.name in config file.

Model	Accuracy	Macro_precision	Macro_recall	Macro_f1
bert-base-chinese-huggingface	65.020	64.987	62.484	63.017
albert_base_zh	62.160	62.514	59.267	60.377
albert_base_zh_additional_36k_steps	61.760	61.723	58.534	59.273
albert_small_zh_google	62.620	63.819	58.992	59.387
albert_large_zh	61.830	61.980	59.843	60.200
albert_tiny	60.110	57.118	55.559	56.077
albert_tiny_489k	61.130	57.875	57.200	57.332
albert_tiny_zh_google	60.860	59.500	57.556	57.702
albert_xlarge_zh_177k	63.380	63.603	60.168	60.596
albert_xlarge_zh_183k	63.210	67.161	59.220	59.599
chinese_wwm	64.000	62.747	64.509	63.042
chinese_wwm_ext	65.020	65.048	62.017	62.688
chinese_roberta_wwm_ext	64.860	64.819	63.275	63.591
chinese_roberta_wwm_large_ext	65.700	62.342	61.527	61.664
ERNIE_stable-1.0.1	66.330	66.903	63.704	64.524
ERNIE_1.0_max-len-512	66.010	65.301	62.230	62.884
chinese_xlnet_base	65.110	64.377	64.862	64.169
chinese_xlnet_mid	66.000	66.377	63.874	64.708
chinese_electra_small	60.370	60.223	57.161	57.206
chinese_electra_small_ex	59.900	58.078	55.525	56.194
chinese_electra_base	60.500	60.090	58.267	58.909
chinese_electra_large	60.500	60.362	57.653	58.336
nezha-base	58.940	57.909	55.650	55.630
nezha-base-wwm	58.800	60.060	54.859	55.831

NOTE: The hyper-parameters used here have not been fine-tuned.

Todos

• More complete and detailed documentation;
• More pretrained models;
• More evaluations of CLUE;
• More Chinese dataset;
• Support Pytorch;
• Improve the tokenizer to make it more versatile;
• Build AiSpace server, make it can train and configure using UI.

Contact Author

Mail: yingyuankai@aliyun.com

Wechat: woshimoming1991

Citing

@misc{AiSpace,
  author = {yuankai ying},
  title = {AiSpace: Highly configurable framework for deep learning model development and deployment},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/yingyuankai/AiSpace}},
}

Refs

• transformers

• UER-py

• Huawei pretrained models