pip install keras-bert

./code/gen_text.py：

split the sentences and write into file ./data人教+北师大中小学

./data/train.txt&valid.txt

end of sentence: "##",now useless

./code/sample.py：

sample train&valid(8:2) vectors from bert.npy,labels.npy

train:[82, 82, 105, 101, 103, 106] valid:[16, 35, 22, 25, 30, 17]

./code/get_bert_CLS.py

load Bert and predict CLS vector of each sentence

./code/model.py

keras model for bert-pre-generate-CLS-vector bi-LSTM classfier using RMSprop, learning rate=0.0005, batch_size = 15 if just want generate model and visualize the model picture $summry, run follow code

    cd ./code/../
    python model.py

./cdoe/train.py

load and train model

    cd ./code
    python train.py

---

./data/

人教+北师大中小学:

Max senteces numers:284 Max senteces length:220 sizes of each grade: [98, 117, 127, 126, 133, 123, 69, 68, 59]

人教+北师大小学:

Max senteces numers:169 Max senteces length:153

---

on fine version of txtCNN

kl_size=[2,3,4],kl_dim = 50. reach 0.47 - 0.50 acc on primary school dataset