Непоследовательные результаты при оценке сохраненного графика TensorFlow, созданного с помощью Keras

Question

Я наблюдаю противоречивые результаты оценки сохраненного графика TensorFlow, созданного с помощью Keras. Пример кода приведен ниже. Несколько вызовов одного и того же ввода с одним сеансом имеют один и тот же вывод, но с использованием одного и того же ввода в нескольких сеансах (т. Е. Путем выполнения кода), возвращает противоречивые результаты. Пожалуйста, сообщите.

import os
import shutil
import tensorflow as tf
from tensorflow.keras import backend as K
from tensorflow.keras.applications.vgg16 import preprocess_input
from tensorflow.keras.applications.vgg16 import VGG16
from tensorflow.keras.layers import Flatten
from tensorflow.keras.models import Model
from tensorflow.keras.preprocessing.image import img_to_array
from tensorflow.keras.preprocessing.image import load_img
from tensorflow.saved_model import signature_constants
from tensorflow.saved_model import tag_constants
from tensorflow.saved_model.builder import SavedModelBuilder
from tensorflow.saved_model.signature_def_utils import predict_signature_def

export_dir = './export'
input_size = 64


def LoadInput():
  image = load_img('mug.jpg', target_size=(input_size, input_size))
  image = img_to_array(image)
  input_batch = image.reshape((1, image.shape[0], image.shape[1],
                               image.shape[2]))
  input_batch = preprocess_input(input_batch)
  return input_batch


def main():
  vgg16 = VGG16(input_shape=(input_size, input_size, 3), include_top=False)
  output = Flatten()(vgg16.get_output_at(-1))
  model = Model(vgg16.input, output)
  # print(model.summary())

  shutil.rmtree(export_dir)
  builder = SavedModelBuilder(export_dir)
  signature = predict_signature_def(
      inputs={'input': model.inputs[0]}, outputs={
          'output': model.outputs[0]
      })
  builder.add_meta_graph_and_variables(
      sess=K.get_session(),
      tags=[tag_constants.SERVING],
      signature_def_map={
          signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature
      },
      main_op=tf.saved_model.main_op.main_op())
  builder.save()

  with tf.Session() as session:
    tf.saved_model.loader.load(session, ["serve"], export_dir)
    feed_dict = {'input_1:0': LoadInput()}
    output_dict = 'flatten/Reshape:0'
    print(session.run(output_dict, feed_dict))


if __name__ == '__main__':
  main()

Answer 1

Это, как оказалось, вызвано неправильной фиксацией графика.Я выложил рабочую версию ниже:

"""Experimenting with Keras VGG16."""
import tensorflow as tf
from tensorflow.keras import backend as K
from tensorflow.keras.applications.vgg16 import preprocess_input
from tensorflow.keras.applications.vgg16 import VGG16
from tensorflow.keras.layers import Flatten
from tensorflow.keras.models import Model
from tensorflow.keras.preprocessing.image import img_to_array
from tensorflow.keras.preprocessing.image import load_img

from IPython import embed

export_dir = '/tmp/export'
input_size = 64


def LoadInput():
  image = load_img('mug.jpg', target_size=(input_size, input_size))
  image = img_to_array(image)
  input_batch = image.reshape(
      (1, image.shape[0], image.shape[1], image.shape[2]))
  input_batch = preprocess_input(input_batch)
  return input_batch


# stackoverflow.com/questions/45466020/how-to-export-keras-h5-to-tensorflow-pb
def FreezeSession(session,
                  keep_var_names=None,
                  output_names=None,
                  clear_devices=True):
  graph = session.graph
  with graph.as_default():
    freeze_var_names = list(
        set(v.op.name for v in tf.global_variables()).difference(
            keep_var_names or []))
    output_names = output_names or []
    output_names += [v.op.name for v in tf.global_variables()]
    input_graph_def = graph.as_graph_def()
    if clear_devices:
      for node in input_graph_def.node:
        node.device = ''
    frozen_graph = tf.graph_util.convert_variables_to_constants(
        session, input_graph_def, output_names, freeze_var_names)
    return frozen_graph


# medium.com/
# @pipidog/how-to-convert-your-keras-models-to-tensorflow-e471400b886a
def RunModel():
  vgg16 = VGG16(input_shape=(input_size, input_size, 3), include_top=False)
  output = Flatten()(vgg16.get_output_at(-1))
  model = Model(vgg16.input, output)
  print(model.summary())

  output_batch = model.predict(LoadInput())
  print(output_batch)
  # print([v for v in output_batch[0]])

  frozen_graph = FreezeSession(
      K.get_session(), output_names=[out.op.name for out in model.outputs])

  tf.train.write_graph(frozen_graph, '/tmp', 'keras-vgg.pb', as_text=False)

  tf.reset_default_graph()

  with tf.Session() as session:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(open('/tmp/keras-vgg.pb', 'rb').read())
    session.graph.as_default()
    tf.import_graph_def(graph_def, name='')

    # for op in session.graph.get_operations():
    #   print(op.name)

    tensor_input = session.graph.get_tensor_by_name('input_1:0')
    tensor_output = session.graph.get_tensor_by_name('flatten/Reshape:0')
    output_batch = session.run(tensor_output, {tensor_input: LoadInput()})
    print(output_batch)
    # print([v for v in output_batch[0]])


def main():
  RunModel()


if __name__ == '__main__':
  main()