tfserver 0.1b6

Tensor Flow Model Server

tfserver is an example for serving Tensorflow model with Skitai App Engine.

It can be accessed by gRPC and JSON RESTful API.

This project is inspired by issue #176.

Table of Contents

• Saving Tensorflow Model

• Run Tensorflow Server

• gRPC Client

• RESTful API

• Performance Note Comparing with Proto Buffer and JSON

  • Test Environment

  • Results

• Release History

Saving Tensorflow Model

See tf.saved_model.builder.SavedModelBuilder, but for example:

import tensorflow as tf

# your own neural network
class DNN:
  ...

net = DNN (phase_train=False)

sess = tf.Session()
sess.run (tf.global_variables_initializer())

# restoring checkpoint
saver = tf.train.Saver (tf.global_variables())
saver.restore (sess, "./models/model.cpkt-1000")

# save model with builder
builder = tf.saved_model.builder.SavedModelBuilder ("exported/1/")

prediction_signature = (
  tf.saved_model.signature_def_utils.build_signature_def(
    inputs = {'x': tf.saved_model.utils.build_tensor_info (net.x)},
    outputs = {'y': tf.saved_model.utils.build_tensor_info (net.predict)])},
    method_name = tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
)
# Remember 'x', 'y' for I/O

legacy_init_op = tf.group (tf.tables_initializer (), name = 'legacy_init_op')
builder.add_meta_graph_and_variables(
  sess,
  [ tf.saved_model.tag_constants.SERVING ],
  signature_def_map = {'predict_class': prediction_signature},
  legacy_init_op = legacy_init_op
)
# Remember 'signature_def_name'

builder.save()

Run Tensorflow Server

Example of api.py

import tfserver
import skitai
import tensorflow as tf

pref = skitai.pref ()
pref.max_client_body_size = 100 * 1024 * 1024 # 100 MB

# we want to serve 2 models:
# alias and (model_dir, optional session config)
pref.config.tf_models ["model1"] = "exported/2"
pref.config.tf_models ["model2"] = (
      "exported/3",
      tf.ConfigProto(
        gpu_options=tf.GPUOptions (per_process_gpu_memory_fraction = 0.2),
        log_device_placement = False
  )
)

# If you want to activate gRPC, should mount on '/'
skitai.mount ("/", tfserver, pref = pref)
skitai.run (port = 5000)

And run,

python3 api.py

gRPC Client

Using grpcio library,

from tfserver import cli
from tensorflow.python.framework import tensor_util
import numpy as np

stub = cli.Server ("http://localhost:5000")
problem = np.array ([1.0, 2.0])

resp = stub.predict (
  'model1', #alias for model
  'predict_class', #signature_def_name
  x = tensor_util.make_tensor_proto(problem.astype('float32'), shape=problem.shape)
)
# then get 'y'
resp.y
>> np.ndarray ([-1.5, 1.6])

Using aquests for async request,

import aquests
from tfserver import cli
from tensorflow.python.framework import tensor_util
import numpy as np

def print_result (resp):
  cli.Response (resp.data).y
  >> np.ndarray ([-1.5, 1.6])

stub = aquests.grpc ("http://localhost:5000", callback = print_result)
problem = np.array ([1.0, 2.0])

request = cli.build_request (
  'model_name',
  'signature_def_name',
  x = tensor_util.make_tensor_proto(problem.astype('float32'), shape=problem.shape)
)
stub.Predict (request, 10.0)

aquests.fetchall ()

RESTful API

Using requests,

import requests

problem = np.array ([1.0, 2.0])
api = requests.session ()
resp = api.post (
  "http://localhost:5000/predict",
  json.dumps ({"x": problem.astype ("float32").tolist()}),
  headers = {"Content-Type": "application/json"}
)
data = json.loads (resp.text)
data ["y"]
>> [-1.5, 1.6]

Another,

from aquests.lib import siesta

problem = np.array ([1.0, 2.0])
api = siesta.API ("http://localhost:5000")
resp = api.predict.post ({"x": problem.astype ("float32").tolist()})
resp.data.y
>> [-1.5, 1.6]

Performance Note Comparing with Proto Buffer and JSON

Test Environment

• Input:

  • dtype: Float 32

  • shape: Various, From (50, 1025) To (300, 1025), Prox. Average (100, 1025)

• Output:

  • dtype: Float 32

  • shape: (60,)

• Request Threads: 16

• Requests Per Thread: 100

• Total Requests: 1,600

Results

Average of 3 runs,

• gRPC with Proto Buffer:

  • Use grpcio

  • 11.58 seconds

• RESTful API with JSON

  • Use requests

  • 216.66 seconds

Proto Buffer is 20 times faster than JSON…

Release History

• 0.1b6 (2018. 3.19): found works only grpcio 1.4.0

• 0.1b3 (2018. 2. 4): add @app.umounted decorator for clearing resource

• 0.1b2: remove self.tfsess.run (tf.global_variables_initializer())

• 0.1b1 (2018. 1. 28): Beta release

• 0.1a (2018. 1. 4): Alpha release