duplicate a tensorflow graph

Question

What is the best way of duplicating a TensorFlow graph and keep it uptodate?

Ideally I want to put the duplicated graph on another device (e.g. from GPU to CPU) and then time to time update the copy.

Answer 1

I'll try to go with a pretty simplified answer, to see if the general approach is what OP is describing:

I'd implement it via the tf.train.Saver object.

Suppose you have your weights in a variable W1, W2, and b1

mysaver = tf.train.Saver(({'w1': W1, 'w2': W2, 'b1': b1}))

In the train loop you can add, every n iterations:

saver.save(session_var, 'model1', global_step=step)

And then in the loading instance, when needed, you run:

tf.train.Saver.restore(other_session_object, 'model1')

Hope this is similar to the solution you are asking.

Answer 2

Short answer: You probably want checkpoint files (permalink).

---

Long answer:

Let's be clear about the setup here. I'll assume that you have two devices, A and B, and you are training on A and running inference on B. Periodically, you'd like to update the parameters on the device running inference with new parameters found during training on the other. The tutorial linked above is a good place to start. It shows you how tf.train.Saver objects work, and you shouldn't need anything more complicated here.

Here is an example:

import tensorflow as tf

def build_net(graph, device):
  with graph.as_default():
    with graph.device(device):
      # Input placeholders
      inputs = tf.placeholder(tf.float32, [None, 784])
      labels = tf.placeholder(tf.float32, [None, 10])
      # Initialization
      w0 = tf.get_variable('w0', shape=[784,256], initializer=tf.contrib.layers.xavier_initializer())
      w1 = tf.get_variable('w1', shape=[256,256], initializer=tf.contrib.layers.xavier_initializer())
      w2 = tf.get_variable('w2', shape=[256,10], initializer=tf.contrib.layers.xavier_initializer())
      b0 = tf.Variable(tf.zeros([256]))
      b1 = tf.Variable(tf.zeros([256]))
      b2 = tf.Variable(tf.zeros([10]))
      # Inference network
      h1  = tf.nn.relu(tf.matmul(inputs, w0)+b0)
      h2  = tf.nn.relu(tf.matmul(h1,w1)+b1)
      output = tf.nn.softmax(tf.matmul(h2,w2)+b2)
      # Training network
      cross_entropy = tf.reduce_mean(-tf.reduce_sum(labels * tf.log(output), reduction_indices=[1]))
      optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)    
      # Your checkpoint function
      saver = tf.train.Saver()
      return tf.initialize_all_variables(), inputs, labels, output, optimizer, saver

The code for the training program:

def programA_main():
  from tensorflow.examples.tutorials.mnist import input_data
  mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
  # Build training network on device A
  graphA = tf.Graph()
  init, inputs, labels, _, training_net, saver = build_net(graphA, '/cpu:0')
  with tf.Session(graph=graphA) as sess:
    sess.run(init)
    for step in xrange(1,10000):
      batch = mnist.train.next_batch(50)
      sess.run(training_net, feed_dict={inputs: batch[0], labels: batch[1]})
      if step%100==0:
        saver.save(sess, '/tmp/graph.checkpoint')
        print 'saved checkpoint'

...and code for an inference program:

def programB_main():
  from tensorflow.examples.tutorials.mnist import input_data
  mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
  # Build inference network on device B
  graphB = tf.Graph()
  init, inputs, _, inference_net, _, saver = build_net(graphB, '/cpu:0')
  with tf.Session(graph=graphB) as sess:
    batch = mnist.test.next_batch(50)

    saver.restore(sess, '/tmp/graph.checkpoint')
    print 'loaded checkpoint'
    out = sess.run(inference_net, feed_dict={inputs: batch[0]})
    print out[0]

    import time; time.sleep(2)

    saver.restore(sess, '/tmp/graph.checkpoint')
    print 'loaded checkpoint'
    out = sess.run(inference_net, feed_dict={inputs: batch[0]})
    print out[1]

If you fire up the training program and then the inference program, you'll see the inference program produces two different outputs (from the same input batch). This is a result of it picking up the parameters that the training program has checkpointed.

Now, this program obviously isn't your end point. We don't do any real synchronization, and you'll have to decide what "periodic" means with respect to checkpointing. But this should give you an idea of how to sync parameters from one network to another.

One final warning: this does not mean that the two networks are necessarily deterministic. There are known non-deterministic elements in TensorFlow (e.g., this), so be wary if you need exactly the same answer. But this is the hard truth about running on multiple devices.

Good luck!