CODEWITHSUNDEEP
pythontensorflowtensorflow-serving
Carpet4
Carpet4

Reputation: 610

using multiple separate neural nets on the same tensor-flow import

I have built a generic python class for interacting with trained neural networks that are saved using "tf.saved_model.builder.SavedModelBuilder".

when I inherit from the class once with a given neural net, everything works correctly. however, when i inherit once more with a second neural net with different architecture, tensor flow throws an error that the shape doesn't fit: "Assign requires shapes of both tensors to match. lhs shape= [100,2] rhs shape= [400,4]"

these shapes are of the two different neural nets, but I don't see why would tensor flow remember about the first net.

Is there an easy way to fix this? and if not, what is the correct way of using multiple neural networks in a project?

here's the class code:

import tensorflow as tf


# prevents tensorflow from using GPU
config = tf.ConfigProto(
  device_count={'GPU': 0}
)


class TFService():

  def __init__(self, netName, inputName, outputName):
    # opens a tensorflow session to use continously
    self.session = tf.Session(config=config)

    # loads the trained neural net
    importDir = 'ocr/neural_nets/{}'.format(netName)

    tf.saved_model.loader.load(
      self.session,
      [tf.saved_model.tag_constants.SERVING],
      importDir
    )

    # saves the input and output tensors for the net
    self.x = tf.get_default_graph().get_tensor_by_name(inputName)
    self.y_pred = tf.get_default_graph().get_tensor_by_name(outputName)

  def getPredictions(self, inputData):
    # the object to feed the neural net
    feed_dict = {self.x: inputData}

    # runs the neural net and returns an array with the predictions
    results = self.session.run(self.y_pred, feed_dict=feed_dict)

    return results

Upvotes: 2

Views: 536

Answers (1)

nessuno
nessuno

Reputation: 27042

Use different graphs for different nets.

You can do something like:

def __init__(self, netName, inputName, outputName):
    self.graph = tf.Graph()
    # opens a tensorflow session to use continously
    # use self.graph as graph the the session
    self.session = tf.Session(config=config, graph=self.graph)


    tf.saved_model.loader.load(
      self.session,
      [tf.saved_model.tag_constants.SERVING],
      importDir
    )

    # saves the input and output tensors for the net
    self.x = self.graph.get_tensor_by_name(inputName)
    self.y_pred = self.graph.get_tensor_by_name(outputName)

Upvotes: 2

Related Questions