tenserflow serving test

Dockerfile

# 2019/02/10現在、Ruby 2.6ではgoogle-protobufが動かない(google-protobufの3.7.0で治る)
FROM ruby:2.5.3

RUN gem install grpc grpc-tools

fmnist.py

import tensorflow as tf
from tensorflow import keras

export_path = './fmnist_model/1'

fashion_mnist = keras.datasets.fashion_mnist
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()
train_images = train_images / 255.0

model = keras.Sequential([
    keras.layers.Flatten(input_shape=(28, 28), name='inputs'),
    keras.layers.Dense(128, activation=tf.nn.relu),
    keras.layers.Dense(10, activation=tf.nn.softmax)
])

model.compile(optimizer='adam', 
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

model.fit(train_images, train_labels, epochs=5)

with tf.keras.backend.get_session() as sess:
    tf.saved_model.simple_save(
        sess,
        export_path,
        inputs={'inputs': model.input},
        outputs={'outputs': model.output})

serving.rb

lib_dir = File.join('/work/proto_ruby')
$LOAD_PATH.unshift(lib_dir) unless $LOAD_PATH.include?(lib_dir)
Dir.glob('/work/proto_ruby/**/*.rb'){ |path| require_relative(path) unless File.directory? path }

require 'json'

data = JSON.load(open('/work/ruby_data.json'))

request = Tensorflow::Serving::PredictRequest.new
request.model_spec = Tensorflow::Serving::ModelSpec.new(name: "fmnist_model", signature_name: data["signature_name"])

images_proto = Tensorflow::TensorProto.new
images_proto.dtype = :DT_FLOAT

shape = Tensorflow::TensorShapeProto.new
shape.dim << Tensorflow::TensorShapeProto::Dim.new(size: 1)
shape.dim << Tensorflow::TensorShapeProto::Dim.new(size: 28)
shape.dim << Tensorflow::TensorShapeProto::Dim.new(size: 28)
shape.dim << Tensorflow::TensorShapeProto::Dim.new(size: 1)

images_proto.tensor_shape = shape

data["inputs"][0].each { |line| line.each { |dot| images_proto.float_val << dot } }

request.inputs['inputs'] = images_proto

require 'grpc'

stub = Tensorflow::Serving::PredictionService::Stub.new('host.docker.internal:8500', :this_channel_is_insecure)
ret = stub.predict(request)
vals = ret.outputs['outputs'].float_val
vals.index(vals.max)
# => 2

shell.sh

cd /work
mkdir proto_ruby
export PROTO_DIR=/work/proto_ruby/

find ./serving/tensorflow_serving -name *.proto | xargs grpc_tools_ruby_protoc -I=serving -I=serving/tensorflow --ruby_out=$PROTO_DIR --grpc_out=$PROTO_DIR --plugin=protoc-gen-grpc=`which grpc_tools_ruby_protoc_plugin` 

grpc_tools_ruby_protoc -I serving/tensorflow --ruby_out=$PROTO_DIR --grpc_out=$PROTO_DIR --plugin=protoc-gen-grpc=`which grpc_tools_ruby_protoc_plugin` serving/tensorflow/tensorflow/core/{framework,example,protobuf}/*.proto

grpc_tools_ruby_protoc -I serving/tensorflow --ruby_out=$PROTO_DIR --grpc_out=$PROTO_DIR --plugin=protoc-gen-grpc=`which grpc_tools_ruby_protoc_plugin` serving/tensorflow/tensorflow/core/lib/core/error_codes.proto

test_data.py

import json

d = {
    "signature_name": 'serving_default',
    "inputs": [test_images[0].tolist()]
}

with open('./test_data.json', mode='w') as f:
    f.write(json.dumps(d))
    
d = {
    "signature_name": 'serving_default',
    "inputs": [test_images[1].tolist()]
}

with open('./ruby_data.json', mode='w') as f:
    f.write(json.dumps(d))
    
print("{}, {}".format(test_labels[0], test_labels[1]))
# => 9, 2