Hello!
I am trying to convert a TF2 saved model into a frozen graph so that I can load it into TensorBoard in order to figure out what the input/output node names are. Everything that has to do with conversions to frozen graphs seems to be deprecated. And when I try to load the SavedModel file into tensorboard it is overly complicated and doesn’t show input/output node names. Would anyone be able to help with this please?
Thank,
Ahmad
1 Like
I guess this is not possible anymore. The SavedModel serialization format is the only format nowadays supported.
You can use the saved_model_cli
tool to inspect the content of the saved model and understand what graphs are stored inside and what are the inputs and outputs.
The typical usage of saved_model_cli
is as follows
saved_model_cli show --all --dir saved_model_folder
This should show you the content of the saved model.
2 Likes
Thank you sir! This was really helpful.
Do you have any knowledge of loading SavedModels into Tensorboard? I am able to do so but it results in very complicated graphs that aren’t very useful to what I need them for.
1 Like
I guess you can load the SavedModel as a Keras model
model = tf.keras.models.load_model(saved_model_path)
and once you have it, you can follow the answer I wrote on StackOverflow about how to graph any graph (the call
method of a Keras model is always decorated with tf.function
hence the same answer applies).
Here’s the answer on SO: python 3.x - How to graph tf.keras model in Tensorflow-2.0? - Stack Overflow
Hope it helps
2 Likes
Thank you so much for your help; I really appreciate it! I’ll give it a try.
2 Likes