|
|
|
""" |
|
Created on Thu Mar 26 09:04:13 2020 |
|
|
|
@author: luol2 |
|
""" |
|
|
|
import time |
|
import sys |
|
import numpy as np |
|
import keras |
|
from src.nn_represent import CNN_RepresentationLayer,BERT_RepresentationLayer |
|
from keras.layers import * |
|
from keras.models import Model |
|
from keras_bert import load_trained_model_from_checkpoint |
|
|
|
|
|
|
|
|
|
class bioTag_CNN(): |
|
def __init__(self, model_files): |
|
self.model_type='cnn' |
|
model_test_type='cnn' |
|
self.fea_dict = {'word': 1, |
|
'char': 1, |
|
'lemma':0, |
|
'pos':0} |
|
|
|
self.hyper = {'sen_max' :20, |
|
'word_max' :40, |
|
'charvec_size' :50, |
|
'pos_size' :50} |
|
|
|
self.w2vfile=model_files['w2vfile'] |
|
self.charfile=model_files['charfile'] |
|
self.labelfile=model_files['labelfile'] |
|
self.posfile=model_files['posfile'] |
|
|
|
vocab={'char':self.charfile,'label':self.labelfile,'pos':self.posfile} |
|
print('loading w2v model.....') |
|
self.rep = CNN_RepresentationLayer(self.w2vfile,vocab_file=vocab, frequency=400000) |
|
|
|
print('building model......') |
|
all_fea = [] |
|
fea_list = [] |
|
|
|
if self.fea_dict['word'] == 1: |
|
word_input = Input(shape=(self.hyper['sen_max'],), dtype='int32', name='word_input') |
|
all_fea.append(word_input) |
|
word_fea = Embedding(self.rep.vec_table.shape[0], self.rep.vec_table.shape[1], weights=[self.rep.vec_table], trainable=True,mask_zero=False, input_length=self.hyper['sen_max'], name='word_emd')(word_input) |
|
fea_list.append(word_fea) |
|
|
|
if self.fea_dict['char'] == 1: |
|
char_input = Input(shape=(self.hyper['sen_max'],self.hyper['word_max']), dtype='int32', name='char_input') |
|
all_fea.append(char_input) |
|
char_fea = TimeDistributed(Embedding(self.rep.char_table_size, self.hyper['charvec_size'], trainable=True,mask_zero=False), name='char_emd')(char_input) |
|
char_fea = TimeDistributed(Conv1D(self.hyper['charvec_size']*2, 3, padding='same',activation='relu'), name="char_cnn")(char_fea) |
|
char_fea_max = TimeDistributed(GlobalMaxPooling1D(), name="char_pooling_max")(char_fea) |
|
fea_list.append(char_fea_max) |
|
|
|
if self.fea_dict['lemma'] == 1: |
|
lemma_input = Input(shape=(self.hyper['sen_max'],), dtype='int32', name='lemma_input') |
|
all_fea.append(lemma_input) |
|
lemma_fea = Embedding(self.rep.vec_table.shape[0], self.rep.vec_table.shape[1], weights=[self.rep.vec_table], trainable=True,mask_zero=False, input_length=self.hyper['sen_max'], name='lemma_emd')(lemma_input) |
|
fea_list.append(lemma_fea) |
|
|
|
if self.fea_dict['pos'] == 1: |
|
pos_input = Input(shape=(self.hyper['sen_max'],), dtype='int32', name='pos_input') |
|
all_fea.append(pos_input) |
|
pos_fea = Embedding(self.rep.pos_table_size, self.hyper['pos_size'], trainable=True,mask_zero=False, input_length=self.hyper['sen_max'], name='pos_emd')(pos_input) |
|
fea_list.append(pos_fea) |
|
|
|
if len(fea_list) == 1: |
|
concate_vec = fea_list[0] |
|
else: |
|
concate_vec = Concatenate()(fea_list) |
|
|
|
concate_vec = Dropout(0.4)(concate_vec) |
|
|
|
|
|
if model_test_type=='cnn': |
|
cnn = Conv1D(1024, 1, padding='valid', activation='relu',name='cnn1')(concate_vec) |
|
cnn = GlobalMaxPooling1D()(cnn) |
|
elif model_test_type=='lstm': |
|
bilstm = Bidirectional(LSTM(200, return_sequences=True, implementation=2, dropout=0.4, recurrent_dropout=0.4), name='bilstm1')(concate_vec) |
|
cnn = GlobalMaxPooling1D()(bilstm) |
|
|
|
|
|
dense = Dense(1024, activation='relu')(cnn) |
|
dense= Dropout(0.4)(dense) |
|
output = Dense(self.rep.label_table_size, activation='softmax')(dense) |
|
self.model = Model(inputs=all_fea, outputs=output) |
|
def load_model(self,model_file): |
|
self.model.load_weights(model_file) |
|
|
|
print('load cnn model done!') |
|
|
|
class bioTag_BERT(): |
|
def __init__(self, model_files): |
|
self.model_type='bert' |
|
self.maxlen = 64 |
|
config_path = model_files['config_path'] |
|
checkpoint_path = model_files['checkpoint_path'] |
|
vocab_path = model_files['vocab_path'] |
|
self.label_file=model_files['labelfile'] |
|
|
|
self.rep = BERT_RepresentationLayer( vocab_path, self.label_file) |
|
|
|
|
|
bert_model = load_trained_model_from_checkpoint(config_path, checkpoint_path, training=False, trainable=True,seq_len=self.maxlen) |
|
|
|
x1_in = Input(shape=(None,)) |
|
x2_in = Input(shape=(None,)) |
|
x = bert_model([x1_in, x2_in]) |
|
x = Lambda(lambda x: x[:, 0])(x) |
|
outputs = Dense(self.rep.label_table_size, activation='softmax')(x) |
|
|
|
self.model = Model(inputs=[x1_in,x2_in], outputs=outputs) |
|
|
|
def load_model(self,model_file): |
|
self.model.load_weights(model_file) |
|
|
|
|
|
class bioTag_Bioformer(): |
|
def __init__(self, model_files): |
|
self.model_type='bioformer' |
|
self.maxlen = 32 |
|
config_path = model_files['config_path'] |
|
checkpoint_path = model_files['checkpoint_path'] |
|
vocab_path = model_files['vocab_path'] |
|
self.label_file=model_files['labelfile'] |
|
|
|
self.rep = BERT_RepresentationLayer( vocab_path, self.label_file) |
|
|
|
|
|
bert_model = load_trained_model_from_checkpoint(config_path, checkpoint_path, training=False, trainable=True,seq_len=self.maxlen) |
|
|
|
x1_in = Input(shape=(None,)) |
|
x2_in = Input(shape=(None,)) |
|
x = bert_model([x1_in, x2_in]) |
|
x = Lambda(lambda x: x[:, 0])(x) |
|
outputs = Dense(self.rep.label_table_size, activation='softmax')(x) |
|
|
|
self.model = Model(inputs=[x1_in,x2_in], outputs=outputs) |
|
|
|
def load_model(self,model_file): |
|
self.model.load_weights(model_file) |
|
|
|
print('load bioformer model done!') |
|
|
|
|
