3

My question is related to this one here about persisting graphs: I wonder if it is possible to only export a certain part of a graph, e.g. a subgraph prefixed by a given scope, using TensorFlow 0.12 or newer.

I'm currently using a combination of tf.train.import_meta_graph(), graph_util.convert_variables_to_constants() and tf.train.write_graph() to export ("freeze") a graph into a protocol buffer file which I then can load back using tf.import_graph_def(). During the export I can specify which nodes are considered required outputs of the graph, so no upstream nodes are thrown away, while during import I can rewire certain parts of the graph to other operations using input_map.

This all works fine, but it is missing the notion of unnecessary inputs and the problem is that by doing so, the entire upstream of the output_nodes is written to the file as well, i.e. everything input and preprocessing related.

Currently, exporting looks like this:

output_nodes = ['subgraph/y', 'subgraph/dropout'] checkpoint = tf.train.latest_checkpoint(input_path) importer = tf.train.import_meta_graph(checkpoint + '.meta', clear_devices=True) graph = tf.get_default_graph() gd = graph.as_graph_def() with tf.Session() as sess: importer.restore(sess, checkpoint) output_graph_def = graph_util.\ convert_variables_to_constants(sess, gd, output_nodes) tf.train.write_graph(output_graph_def, 'dir', 'file.pb', as_text=False)

While importing looks like this:

with tf.gfile.GFile(input_path, "rb") as f: graph_def = tf.GraphDef() graph_def.ParseFromString(f.read()) elems = tf.import_graph_def( graph_def, input_map=None, return_elements=output_nodes, name='imported' )

Is there a way to filter and/or remove the parts I do not require during or before exporting? To be clear, I known I can ignore them after loading, but I do not to export them to begin with. Would using collections help at some point?

Improve this question

1 Answer 1

Reset to default
1

The built in graph saving tools need to be reworked in my opinion.

The best solution I've found is to create a list of the variables you want saved then save them during the session. Afterwards you have the freedom to reload the weights and change the graph as needed.

parameters = [] parameters += [wieghts, biases] ... def load_weights(self, weight_file, sess): weights = np.load(weight_file) keys = sorted(weights.keys()) for i, k in enumerate(keys): print i, k, np.shape(weights[k]) sess.run(self.parameters[i].assign(weights[k]))

credit for idea goes to Davi Frossard here

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

2 Comments

Thanks so far. Maybe I'm missing something, but doesn't that require me to know how to construct the graph? What I want is one single, deployable file containing both weights and operations (without all the pre- and postprocessing pieces).
The only thing I've seen is if you can literally isolate that part of the code and still have it run in tf then you can export like you want. I haven't found any clean way to do what you've suggested. Maybe someone else knows a better way and I'd also like to know.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.