解决falsk和TensorFlow兼容性问题

 其实我们只需要添加三个代码就可以解决，这个问题的原因是因为TF默认图和你自己的图不在一个session中，会提示各种错误。

# ----------------------------解决TF和flask兼容问题----------------------------

# 在代码前加上这三行
sess = tf.Session()
# sess = tf.compat.v1.Session()
graph = tf.get_default_graph()

# ----------------------------解决TF和flask兼容问题----------------------------
# 在model加载前添加set_session
set_session(sess)

# ----------------------------解决TF和flask兼容问题----------------------------

 # 在预测前，加入下面这段代码

global sess
global graph
with graph.as_default():
    set_session(sess)
    qyery_emb = sent2vec_312(sentence_list)
    # 从封装形式改为代码直接处理的形式
    print(qyery_emb.shape)

下面是一个项目完整的代码

from flask import Flask, request, jsonify
from log import Logger
import time
import json
import os
# from bert4vec import Bert4Vec
from pyhive import hive
import pandas as pd
pd.set_option('display.max_columns', None)
import faiss
import pickle
import datetime

import warnings
warnings.filterwarnings('ignore')

import numpy as np
from collections import Counter
from bert4keras.backend import keras, K
from bert4keras.models import build_transformer_model
from bert4keras.tokenizers import Tokenizer
from bert4keras.snippets import sequence_padding
from bert4keras.snippets import uniout
from keras.models import Model
from faiss_index import faissIndex
import tensorflow as tf
from tensorflow.python.keras.backend import set_session

# self_graph = tf.get_default_graph()

# define logging configuration
logger = Logger(filename='./logs/semantic_matching_service.log', level='debug', when='midnight', back_count=14).logger

# ----------------------------解决TF和flask兼容问题----------------------------
sess = tf.Session()
# sess = tf.compat.v1.Session()
graph = tf.get_default_graph()

# ----------------------------解决TF和flask兼容问题----------------------------
# 在model加载前添加set_session
set_session(sess)

# with self_graph.as_default():
# 准备bert的相关环境
maxlen = 64
# bert配置
config_path = '/home/models/chinese_simbert_L-4_H-312_A-12/bert_config.json'
checkpoint_path = '/home/models/chinese_simbert_L-4_H-312_A-12/bert_model.ckpt'
dict_path = '/home/models/chinese_simbert_L-4_H-312_A-12/vocab.txt'

# 建立分词器
tokenizer = Tokenizer(dict_path, do_lower_case=True)  # 建立分词器
# 建立加载模型
bert = build_transformer_model(
    config_path,
    checkpoint_path,
    with_pool='linear',
    application='unilm',
    return_keras_model=False,
)
encoder = keras.models.Model(bert.model.inputs, bert.model.outputs[0])

print("加载完毕。。。。。。。。。。。。。。。。。。")

# 解决Flask和TensorFlow的兼容问题，解决方案： 
# 1. 初始化后加 self.graph = tf.get_default_graph()
# 2. predict时加 graph = self.graph




# 输入一个句子的list ,list长度为100
# 输出一个向量，大小为100x312
def sent2vec_312(sent_list):
    # 测试相似度效果
    a_token_ids = []
    for d in sent_list:  
        token_id = tokenizer.encode(d, max_length=maxlen)[0]
        a_token_ids.append(token_id) # 2d-list

    a_token_ids = sequence_padding(a_token_ids) # 统一句子长度
    # with self_graph.as_default():
    a_vecs = encoder.predict([a_token_ids, np.zeros_like(a_token_ids)], verbose=True)
    return a_vecs


"""
传参接收一个query
运行simbert进行编码成向量
加载faiss进行检索
返回resource_type和resource_id
"""
# define flask app
app = Flask(__name__)
@app.route("/semantic_matching_service", methods=["GET","POST"])
def search_rank_ctr():
    print("semantic_matching_service 程序开始运行.....")

    total_start_time = time.time()
    # --------------------------解析参数--------------------------
    query = request.args.get("query")
    print("用户请求的query = %s"%query)
    # --------------------------解析参数--------------------------

    # --------------------------运行simbert进行编码成向量--------------------------
    sentence_list = []
    sentence_list.append(query)
    # with self_graph.as_default():
    global sess
    global graph
    with graph.as_default():
        set_session(sess)
        qyery_emb = sent2vec_312(sentence_list)
        # 从封装形式改为代码直接处理的形式
        print(qyery_emb.shape)
    # --------------------------运行simbert进行编码成向量--------------------------


    total_end_time = time.time()
    print("semantic_matching_service cost total time = %s ms"%int((total_end_time-total_start_time)*1000))
    return "你好"


if __name__ == '__main__':

    # 模型部署
    app.run(host='0.0.0.0', threaded=True, port=5015)