How to restore a partial graph in tensorflow?

Question

I would like to restore only the part of computation graph in tensorflow. My architecture contains two networks. The output of the first network is the input to the second network. The first network is pretrained and I want to restore from a checkpoint. I don't want to update the parameters of the first network as well. Is there an example that I can follow to achieve this?

Thanks

Answer 1

I don't have exact code for you task, but here is a short guide that may help you:

First you need to parse your network into tf.GraphDef format code should like this:

graph_def = tf.GraphDef()
with tf.gfile.FastGFile("path/to/graphdef") as f:
  s = f.read()
graph_def.ParseFromString(s)

or restore from a checkpoint/saved_mode then convert to GraphDef by:

tf.train.import_meta_graph('checkpoint.meta')
tf.get_default_graph().as_graph_def()

now you have the graph_def

Second, extract subgraph of the graph_def with tf.graph_util.extract_sub_graph, you can specify the dest nodes which are you inputs to the second network as well.

Last, import the subgraph from second step with tf.import_graph_def.

Also, since you don't want to update the parameters for the first network, you can freeze its parameters with tf.graph_util.convert_variables_to_constants