Feed both tf.data.Dataset and NumPy array to model

Question

I have a class representing a model that is set up as follows:

class Model:
  def __init__(self):
    self.setup_graph()

  def setup_graph():
    # sets up the model
    ....

  def train(self, dataset):
    # dataset is a tf.data.Dataset iterator, from which I can get 
    # tf.Tensor objects directly, which become part of the graph
    ....

  def predict(self, sample):
    # sample is a single NumPy array representing a sample,
    # which could be fed to a tf.placeholder using feed_dict
    ....

During training I want to make use of the efficiency of TensorFlow's tf.data.Dataset, but I still want to be able to get the output of the model on a single sample. It seems to me that this requires recreating the graph for prediction. Is this true, or can I create a TF graph where I can either run with a sample from a tf.data.Dataset, or with a given sample I feed to a tf.placeholder?

javidcf · Accepted Answer · 2018-10-02 14:51:33Z

1

You can create your model with a dataset, iterator, etc as usual. Then, if you want to pass some custom data with feed_dict, you can just by passing values to the tensors produced by get_next():

import tensorflow as tf
import numpy as np

dataset = (tf.data.Dataset
    .from_tensor_slices(np.ones((100, 3), dtype=np.float32))
    .batch(5))
iterator = dataset.make_one_shot_iterator()
batch = iterator.get_next()

output = 2 * batch

with tf.Session() as sess:
    print('From iterator:')
    print(sess.run(output))
    print('From feed_dict:')
    print(sess.run(output, feed_dict={batch: [[1, 2, 3]]}))

Output:

From iterator:
[[2. 2. 2.]
 [2. 2. 2.]
 [2. 2. 2.]
 [2. 2. 2.]
 [2. 2. 2.]]
From feed_dict:
[[2. 4. 6.]]

In principle you could achieve the same effect with initializable, reinitializable or feedable iterators, but if you really just want to test single samples of data I think this is the quickest and less intrusive way.

edited Oct 2, 2018 at 14:51

answered Oct 2, 2018 at 14:46

javidcf

59.9k7 gold badges87 silver badges134 bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

Moos Hueting Over a year ago

Great, jdehesa, thank you! Is there a way I can do this the other way around, i.e. can I set it up so that I can pass self.train() a tf.data.Dataset object?

javidcf Over a year ago

@MoosHueting You cannot pass a dataset (or an iterator, or its output) in the feed_dict. You can create different datasets and switch between them with a reinitializable or a feedable iterator (see Creating an iterator), or you can switch between a dataset and a placeholder or something else with tf.cond, or something like that. I am not sure what should be the best option for your use case.

Moos Hueting Over a year ago

Thanks @jdehesa. Feedable iterators turned out to be the solution for me.

Collectives™ on Stack Overflow

Feed both tf.data.Dataset and NumPy array to model

1 Answer 1

3 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

3 Comments

Your Answer

Sign up or log in

Post as a guest

Related