[tensorflow] modelのsave,load,run

modelのsave

tensorflowでmodelをsaveする方法は二つある。check_pointとsaved_model。

check_point

check_pointはEstimatorにRunconfigを渡すことで可能。何分でcheck_pointを取るか設定可能。train途中に中止してもcheck_pointを読み込むことでtrainを続けることが可能。定義しなければdefaultの設定に従う。saved_modelを使った方が簡単になるし、saved_model_cliを使うことも可能。

saved_model

saved_modelはsession内でSavedModelBuilderを使って自分でbuildするか、Estimatorの関数export_savedmodelを使う。以下はexport_savedmodelの例。

example

def model_fn(features, labels, mode, params):
  model = Model(params['data_format'])
  image = features
  # saved_modelをloadしてserveする場合
  if isinstance(image, dict):
    image = features['image']
  if mode == tf.estimator.ModeKeys.PREDICT:
    logits = model(image, training=False)
  ...
mnist_classifier = tf.estimator.Estimator(
    model_fn=model_fn,
    model_dir=model_dir,
    params={
        'data_format': 'channels_last'
    })
# input
image = tf.placeholder(tf.float32, shape=[None, 28, 28], name='image')
# input receiver
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
    'image': image,
})
# export model
mnist_classifier.export_savedmodel(model_dir, input_fn)

modelのloadとrun

saved_model_cli

saved_modelの中身を見るコマンド

command

saved_model_cli show --dir ./mnist_model/1521441078 --all

result

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['classify']:
The given SavedModel SignatureDef contains the following input(s):
inputs['image'] tensor_info:
    dtype: DT_FLOAT
    shape: (-1, 28, 28)
    name: Placeholder:0
The given SavedModel SignatureDef contains the following output(s):
outputs['classes'] tensor_info:
    dtype: DT_INT64
    shape: (-1)
    name: ArgMax:0
outputs['probabilities'] tensor_info:
    dtype: DT_FLOAT
    shape: (-1, 10)
    name: Softmax:0
Method name is: tensorflow/serving/predict

signature_def['serving_default']:
The given SavedModel SignatureDef contains the following input(s):
inputs['image'] tensor_info:
    dtype: DT_FLOAT
    shape: (-1, 28, 28)
    name: Placeholder:0
The given SavedModel SignatureDef contains the following output(s):
outputs['classes'] tensor_info:
    dtype: DT_INT64
    shape: (-1)
    name: ArgMax:0
outputs['probabilities'] tensor_info:
    dtype: DT_FLOAT
    shape: (-1, 10)
    name: Softmax:0
Method name is: tensorflow/serving/predict

上のnameに注目、Placeholder:0がinput test image、ArgMax:0がouput classes

saved_modelを使ってnpy(numpy saved data)ファイルのテストデータのラベルを予測する。

command

saved_model_cli run --dir ./mnist_model/1521441078 --tag_set serve --signature_def classify --inputs image=./example2.npy

result

Result for output key classes:
[7 2 1 ... 9 8 6]
Result for output key probabilities:
[[5.71908802e-03 5.26234088e-03 5.65170124e-03 ... 8.65403712e-01
  2.58122981e-02 4.70778644e-02]
 [3.00555006e-02 1.64930541e-02 6.53899252e-01 ... 9.86633589e-04
  3.90466303e-02 4.75284224e-03]
 [6.55172905e-03 7.44415641e-01 3.45991701e-02 ... 2.27554981e-02
  5.24941944e-02 3.27617601e-02]
 ...
 [6.86948374e-03 3.37781794e-02 1.31903710e-02 ... 1.70341298e-01
  1.24142714e-01 3.21431905e-01]
 [6.56285435e-02 3.99611443e-02 3.55339721e-02 ... 8.27240124e-02
  2.49543816e-01 1.30154327e-01]
 [1.08888358e-01 8.15967447e-04 1.35427341e-01 ... 3.61581246e-04
  7.29427906e-03 4.25529340e-03]]

Result for output key classes = ArgMax:0
Result for output key probabilities = Softmax:0

saved_modelをsessionで読み込み、pythonコードで結果をみる例

example

import tensorflow as tf
import numpy as np

export_dir = './mnist_model/1521510868'

te = np.load('example2.npy')
te0 = np.reshape(te[0,:], (-1, 28, 28))

with tf.Session(graph=tf.Graph()) as sess:
  # saved_model load
  tf.saved_model.loader.load(sess, [tf.saved_model.tag_constants.SERVING], export_dir)
  # input
  i = sess.graph.get_tensor_by_name("image:0")
  # output
  r = sess.graph.get_tensor_by_name("ArgMax:0")
  print(sess.run(r, feed_dict={i:te0}))