KZKY memo

自分用メモ．

TensorFlow: Graphをもう少し理解

tensorflow python deeplearning

Graphをもう少し理解する．

  1. Default graph
  2. Create another graph in this thread (main thread)
  3. Graph in multi thread
  4. Write graph as protbuf to disk
  5. Read graph from disk and to Graph

の5通りのサンプル．チェックポイントからの復帰はやってない.

チェックポイントからの復帰は，これを参考にすればできるはず．

tf.import_graph_def(graph_def)でimportしたときには，opsは，default_graphに追加される．別のgraphに追加したいなら，g = tf.Graph()して，g.as_default()して，そのコンテキストの中で，tf.import_graph_def(graph_def)すること．

#!/usr/bin/env python

import tensorflow as tf
import threading
import numpy as np

class GraphWorker(threading.Thread):
    """
    """
    
    def __init__(self, index):
        """
        """
        super(GraphWorker, self).__init__()
        self.index = index
        
        pass

    def run(self):

        g = tf.get_default_graph()
        #print "Graph in main thread", g

        g_local = tf.Graph()
        #print "Graph in this thread", g_local

def main():
    # Default graph
    """
    Confirm that a default Graph is always registered, and accessible by calling tf.get_default_graph(). To add an operation to the default graph, simply call one of the functions that defines a new Operation:
    """
    c = tf.constant(4.0)
    assert c.graph is tf.get_default_graph()

    # Create another graph in this thread (main thread)
    """
    Confirm that tf.Graph.as_default() method should be used if you want to create multiple graphs in the same process.
    """
    with tf.Graph().as_default() as g:
        c = tf.constant(30.0)
        assert c.graph is g
        d = tf.constant(40.0)
        assert d.graph is g
        e = tf.Variable(np.random.rand(10, 10), name="e")
        
    assert tf.get_default_graph() != g

    # Graph in multi thread
    """
    The default graph is a property of the current thread. If you create a new thread, and wish to use the default graph in that thread, you must explicitly add a with g.as_default(): in that thread's function.
    """
    n = 4
    graph_workers = []
    for i in xrange(n):
        worker = GraphWorker(i)
        worker.start()
        graph_workers.append(worker)

    for i in xrange(n):
        graph_workers[i].join()

    # Write graph as protbuf to disk
    print g.get_operations()
    print len(g.get_operations())
    tf.train.write_graph(g.as_graph_def(), "./graph_dir_text", "./graph.pbtxt")
    tf.train.write_graph(g.as_graph_def(), "./graph_dir", "./graph.pb", as_text=False)

    # Read graph from disk and to Graph
    print tf.get_default_graph()
    with open("./graph_dir/graph.pb", "rb") as fp:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(fp.read())
        tf.import_graph_def(graph_def)
    print tf.get_default_graph()

    print "----- ops in g -----"
    for op in g.get_operations():
        print op.name

    print "----- ops in default graph after import_graph_def -----"
    for op in tf.get_default_graph().get_operations():
        print op.name
    
    pass


if __name__ == '__main__':
    main()

参考