Flask部署keras/tensorflow训练好模型进行预测：ValueError: Tensor Tensor is not an element of this graph.

flask接口部署keras、tensorflow服务，报如下错误：

1，ValueError: Tensor Tensor("dense_7/Softmax:0", shape=(?, 3), dtype=float32) is not an element of this graph.

2，tensorflow.python.framework.errors_impl.FailedPreconditionError: Error while reading resource variable Embedding-Norm/beta from Container: localhost. This could mean that the variable was uninitialized. Not found: Container localhost does not exist. (Could not find resource: localhost/Embedding-Norm/beta)
     [[{{node Embedding-Norm/add_1/ReadVariableOp}}]]

解决办法：

1，创建全局的graph、session

sess = tf.Session()
graph = tf.get_default_graph()
# tf2： graph = tf.compat.v1.get_default_graph()
set_session(sess)

2，在全局graph、Session中，model加载前添加set_session，进行model的预测

    （全局变量最好放在函数体外）

    global sess
    global graph
    with graph.as_default():
        set_session(sess)

具体代码如下：

import tensorflow as tf
from keras.backend import set_session
app = Flask(__name__)

#global全局变量，最好放在函数外，避免出错
global sess
global graph
sess = tf.Session()
graph = tf.get_default_graph()
set_session(sess)
model.load_weights('best_model.weights')

@app.route('/judge', methods=['GET'])
def handler():
    keyword = request.args.get('keyword')
    sentence = request.args.get('sentence')
    judge = {}

    with graph.as_default():
        set_session(sess)
        if keyword in sentence:
            judge['result'] = str(evaluate(test_generator))
        else:
            judge['result'] = '0'
    return json.dumps(judge)


if __name__ == '__main__':
    app.run(host='0.0.0.0', port=65534, threaded=True)