tfserver 0.1a8

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.

Saving Tensoflw Model

See tf.saved_model.builder.SavedModelBuilder

Tensorflow Server

Example of api.py

import tfserver
import skitai
import dnn
import tensorflow as tf

pref = skitai.pref ()

pref.debug = True
pref.use_reloader = True

tf.reset_default_graph()
net = dnn.make_mlp_network (phase_train=False)

pref.config.tf_config = tf.ConfigProto(
  gpu_options=tf.GPUOptions (per_process_gpu_memory_fraction = 0.2),
  log_device_placement = False
)
pref.config.tf_model_dir = "./exported/2"
pref.config.tf_predict_op = net ["pred"]
pref.config.tf_x = net ["x"]

skitai.mount ("/", tfserver, pref = pref)
skitai.run (port = 5000)

And run,

python api.py

gRPC Client

Using grpc,

from tfserver import cli
from tensorflow.python.framework import tensor_util

stub = cli.Proxy ("localhost", 5000)
x = np.array ([1.0, 2.0])

resp = stub.predict (
  'model_name',
  'signature_name',
  tensor_util.make_tensor_proto(x.astype('float32'), shape=x.shape)
)
resp.y
>> [-1.5, 1.6]

Using aquests,

from tfserver import predict_pb2, cli
import aquests
from tensorflow.python.framework import tensor_util

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

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

request = predict_pb2.PredictRequest()
request.model_spec.name = 'model_name'
request.model_spec.signature_name = 'signature_name'
request.inputs ["x"].CopyFrom(tensor_util.make_tensor_proto(fftseq.astype('float32'), shape=fftseq.shape))
stub.Predict (request, 10.0)

aquests.fetchall ()

But aquests’ grpc is not stable yet.

REST API

Using requests,

import requests

api = requests.session ()
resp = api.post (
  "http://localhost:5000/predict",
  json.dumps ({"x": getone ().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

x = np.array ([1.0, 2.0])

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