2. 调整各种Embedding的实现，使得确保来自dev和test的未发现词使用unk的表示
3. 在Embedding中增加dropout_word的选项，使得可以随机drop掉词语
4. 以及其它若干小的bug

2. 修改embedding.py中的bug
3. ConllReader默认跳过所有的DOCSTART标签
4. 交换bert的heavy lifting到_bert, 将BertEncoder在bert.py中暴露
5. crf中allow_transition的include_end_start修改为false，以与CRF的默认值适配
6. allow_transition与SpanMetric支持BIOES类型的tag
7. datainfo中增加打印格式化输出

2. 重大更新，支持ElmoEmbedding, BertEmbedding

1. 完善 trainer 和 tester 部分的文档
2. 研究注释样例与测试

* delete field.py
* remove logger in all codes
* adjust arguments of Trainer

* rename preprocess.py into utils.py, because nothing about preprocess in it
* anything in loader/ and saver/ is moved directly into io/
* corresponding unit tests are moved to /test/io
* delete fastnlp.py, because we have new and better APIs
* rename Biaffine_parser/run_test.py to Biaffine_parser/main.py; Otherwise, test will fail.
* A looooooooooot of ancient codes to be refined...........

- apply DataSet in Predictor; remove sub-predictors; add "task" argument to specify which task to predict, as how Trainer/Tester did.
- remove Action class
- add helper function for DataSet, to create DataSet easily
- more code comments
- clean up unnecessary codes
- add unit tests for Batch, Predictor, Preprocessor, Trainer, Tester

- update LabelField's to_tensor method to support int & str single label
- update preprocessor's convert_to_dataset method to support single label inputs
- introduce "task" in Trainer/Tester's data_forward, Tester's evaluate and metrics methods
- in cnn_text_classification.py, change the name of the argument of forward
- in sequence_modeling.py, change the name of the argument of forward
- minor adjustments in test codes
- text_classify.py works

- add DataSet, Instance, Field to represent data in different levels
- encapsulate batching method in Batch class
- modify samplers in action.py to fit Batch
- preprocessor.run returns DataSet, instead of list
- Use Batch in Trainer/Tester
- add required_arg "task" in Trainer/Tester
- remove SeqLabelTrainer/SeqLabelTester dependencies successfully. They empty classes to deprecate.
- modify SeqLabeling model, add another argument in forward, in order to compute mask inside model
- test\model\seq_labeling.py works