def test():
bp = bert_params()
#token_dict = load_vocab(bp.bert_vocab)
t = Tokenizer(bp.bert_vocab)
return t
def test_bert_params():
'''测试bert模型导入数据是否正确'''
'''测试注意事项：with_pool默认为False,在call中注释掉经过最后的Dense层'''
bp = bert_params() #with_pool默认为False
#dg = DataGenerator(bp,batch_size = 2,num_neg = 2,shuffle = True)
my_model = Bert4QA(bp)
tokenizer = Tokenizer(bp.bert_vocab, do_lower_case=True)
token_ids, segment_ids = tokenizer.encode("语言模型")
print("token_ids:", token_ids)
print('\n ===== predicting =====\n')
token_ids = tf.convert_to_tensor([token_ids], dtype=tf.float32)
segment_ids = tf.convert_to_tensor([segment_ids], dtype=tf.float32)
ans = my_model.call([token_ids, segment_ids])
print(ans)
def build_model(with_pool=False, with_nsp=False, with_mlm=False):
bp = bert_params(with_pool=with_pool, with_nsp=with_nsp, with_mlm=with_mlm)
print("bp.with_mlm:", bp.with_mlm)
bert_model = Bert(bp.vocab_size, bp.hidden_size, bp.num_hidden_layers,
bp.num_attention_heads, bp.intermediate_size,
bp.hidden_act, bp.max_position_embeddings,
bp.hidden_dropout_prob, bp.attention_probs_dropout_prob,
bp.with_pool, bp.with_nsp, bp.with_mlm)
inputs = [Input((None, )), Input((None, ))]
outputs = bert_model(inputs)
#如果是预测模型的话，outputs是个列表
#print(outputs.shape)
bert_model.load_weights_from_checkpoint(bp.bert_ckpt)
return bert_model
def main():
#模型训练预测时的参数
mp = model_params()
#构建bert的参数
bp = bert_params(with_pool=True)
datagen = DataGenerator(bp,
batch_size=mp.batch_size,
num_neg=mp.num_neg,
shuffle=mp.shuffle)
#后续再尝试用其它的优化器
optimizer = Adagrad(learning_rate=mp.learning_rate)
my_model = Bert4QA(bp)
#训练主类
t = TrainOrPredict(mp)
#final_model就是训练好的模型
final_model = t.train(my_model, optimizer, datagen)
data = datagen.data_faq
tokenizer = datagen.tokenizer
#训练完成后查看效果
real_query_text = "月球和地球是什么关系?"
question_score = {}
for query_name in data.query_dict.keys():
query_text = data.query_dict[query_name]
token_ids, segment_ids = tokenizer.encode(real_query_text, query_text)
question_score[query_name] = final_model.predict(
[token_ids, segment_ids])
question_score = {k: v.numpy() for k, v in question_score.items()}
qs = dict(sorted(question_score.items(), key=lambda x: x[1], reverse=True))
c = 0
for k, v in qs.items():
c += 1
print(k, data.query_dict[k], v)
if c == 10: break
return final_model
def test():
mp = model_params()
bp = bert_params(with_pool=True)
#其实这里内部不用导入一遍bert源码数据
my_model = Bert4QA(bp)
tokenizer = Tokenizer(bp.bert_vocab, do_lower_case=True)
optimizer = Adagrad(learning_rate=mp.learning_rate)
#注意参数名一定要和保存时的参数名一致
ckpt = tf.train.Checkpoint(optimizer=optimizer, model=my_model)
ckpt.restore(tf.train.latest_checkpoint("./save_checkpoint/"))
#后面的和导入数据没有关系
data_faq = DataBasic(bp.FAQ_file_path)
real_query_text = "月球和地球是什么关系?"
real_query_text = "月球和地球的关系"
real_query_text = "月球是地球的卫星吗"
print("实际查询：", real_query_text)
question_score = {}
for query_name in data_faq.query_dict.keys():
query_text = data_faq.query_dict[query_name]
token_ids, segment_ids = tokenizer.encode(real_query_text, query_text)
question_score[query_name] = my_model.predict([token_ids, segment_ids])
question_score = {k: v.numpy() for k, v in question_score.items()}
qs = dict(sorted(question_score.items(), key=lambda x: x[1], reverse=True))
for k, v in qs.items():
print(k, data_faq.query_dict[k], v)
return qs
bp.with_pool, bp.with_nsp, bp.with_mlm)
inputs = [Input((None, )), Input((None, ))]
outputs = bert_model(inputs)
#如果是预测模型的话，outputs是个列表
#print(outputs.shape)
bert_model.load_weights_from_checkpoint(bp.bert_ckpt)
return bert_model
if __name__ == "__main__":
bp = bert_params()
print(bp.vocab_size)
my_model = Bert(bp.vocab_size, bp.hidden_size, bp.num_hidden_layers,
bp.num_attention_heads, bp.intermediate_size,
bp.hidden_act, bp.max_position_embeddings,
bp.hidden_dropout_prob, bp.attention_probs_dropout_prob,
bp.with_pool, bp.with_nsp, bp.with_mlm)
T = 256
inputs = [Input((None, )), Input((None, ))]
outputs = my_model(inputs)
#如果是预测模型的话，outputs是个列表
#print(outputs.shape)