I have a frozen inference graph stored in a .pb file, which was obtained from a trained Tensorflow model by the freeze_graph function.
Suppose, for simplicity, that I would like to change some of the sigmoid activations in the model to tanh activations (and let's not discuss whether this is a good idea).
How can this be done with access only to the frozen graph in the .pb file, and without the possibility to retrain the model?
I am aware of the Graph Editor library in tf.contrib, which should be able to do this kind of job, but I wasn't able to figure out a simple way to do this in the documentation.
The *.pb file contains a SavedModel protocol buffer. You should be able to load it using a SavedModel loader. You can also inpsect it with the SavedModel CLI. The full documentation on SavedModels is here.
Can you try this:
graph = load_graph(filename)
graph_def = graph.as_graph_def()
# if ReLu op is at node 161
graph_def.node[161].op="tanh"
tf.train.write_graph(graph_def, path2savfrozn, "altered_frozen.pb", False)
Please let know the if it works.
Something along these lines should work:
graph_def = tf.GraphDef()
with open('frozen_inference.pb', 'rb') as f:
graph_def.ParseFromString(f.read())
with tf.Graph().as_default() as graph:
importer.import_graph_def(graph_def, name='')
new_model = tf.GraphDef()
with tf.Session(graph=graph) as sess:
for n in sess.graph_def.node:
if n.op == 'Sigmoid':
nn = new_model.node.add()
nn.op = 'Tanh'
nn.name = n.name
for i in n.input:
nn.input.extend([i])
else:
nn = new_model.node.add()
nn.CopyFrom(n)
