Tornado multiprocess does not work for tensorflow

[alexwongdl]

I try to write a simple frame for tornado + tensorflow server, code is as following:

import tornado.httpserver
import tornado.ioloop
import tornado.options
import tornado.web
from tornado.options import define
import os
os.environ["CUDA_VISIBLE_DEVICES"] = ""

define("port", default=18015, help="run on the given port", type=int)

import tensorflow as tf
a = tf.placeholder(tf.int32, shape=(), name="input")
asquare = tf.multiply(a, a, name="output")
config = tf.ConfigProto(device_count={"CPU": 20}, inter_op_parallelism_threads=10, intra_op_parallelism_threads=10)
sess = tf.Session(config=config)

class IndexHandler(tornado.web.RequestHandler):
    def initialize(self, thesess):
        self.sess = thesess

    def get(self):
        num = self.get_argument('num', 2)
        for i in range (100):
            ret = self.sess.run([asquare], feed_dict={a: num})
        self.write( 'result:' + str(ret))

if __name__ == "__main__":
    tornado.options.parse_command_line()
    app = tornado.web.Application(handlers=[(r"/", IndexHandler,dict(thesess=sess))])
    http_server = tornado.httpserver.HTTPServer(app)
    # http_server.listen(18824)
    # tornado.ioloop.IOLoop.current().start()

    http_server.bind(18825)
    http_server.start(10)
    tornado.ioloop.IOLoop.current().start()

subprocess starts succeed, however, the program is stuck and does not process any request. How could I make it work please. When I use simple process or multithread, the program also works.

Environment:

1. tensorflow 1.2
2. tornado 4.4.1
3. python 3.5.2

[bdarnell]

You may only use fork-safe libraries before entering multi-process mode (by calling http_server.start(N)). Many complex libraries are not fork-safe. You must either move the initialization of tensorflow.Session() after the HTTP server is started (in which case you will have 10 sessions, one in each process), or you can create a separate tensorflow server and connect to it using the target argument of tensorflow.Session (sessions with a target are fork-safe). The latter option is described in tensorflow/tensorflow#2448