Python 多线程Flask应用程序使用MRCNN引发错误

我有一个flask应用程序正在线程真实模式下运行

if __name__ == '__main__':
    setup.check_director()
    web_app.secret = "image-service"
    web_app.run(host='0.0.0.0', port=8084, debug=True, threaded=True)

我正在使用api调用触发mrcnn（）任务，Predictor类如下所示

class Predictor(object):
    def __init__(self, config=None, weights=WEIGHTS_PATH, device=DEVICE):
        self.weights=weights
        self.config = ElementConfig() if config is None else config
        # self.el = Element_linker()

        self.logger.info("Loading Mask R-CNN model")
        print("Loading Mask R-CNN model and weights")
        start = time.time()

        self._graph = tf.Graph()

        with self._graph.as_default():
            self.model = modellib.MaskRCNN(mode="inference", model_dir=MODEL_DIR, config=self.config)
            self.logger.info("Loading saved weights from {}".format(weights))
            self.model.load_weights(weights, by_name=True)
            self.init_g = tf.global_variables_initializer()
            self.init_l = tf.local_variables_initializer()
            print("Loading completed in %s seconds" % ((time.time() - start)))

    ....
    ....
    ....

    def predict(self):
        with tf.Session(graph=self._graph) as _session:
           _session.run(self.init_g)
           _session.run(self.init_l)
           results = _session.run(self.model.detect([image], verbose=0))

    ....
    ....
    ....
    ....

当一次只有一个api调用时，代码工作正常。当我同时命中多个api时，代码抛出错误，如下所示：

ValueError: Tensor Tensor("Placeholder:0", shape=(7, 7, 3, 64), dtype=float32) is not an element of this graph.

详细日志：

我还尝试使用不同的方式组织会话和图表。我们也尝试了mrcnn github问题页面上的所有解决方案，但没有任何帮助

线程或上述方法是否存在问题？