I have experienced problems when using TF_GraphImportGraphDef to import
certain graphs with gradients from Python, and I suspect the function might have a bug.

The simplest example of a problematic graph is:

x = tf.placeholder(tf.float64, [1, 2], name="x")
y = tf.placeholder(tf.float64, [1, 2], name="y")
z = tf.add(x, y, name="z")

tf.gradients(z, [x, y])

When running TF_GraphImportGraphDef, I get the following error message:
Cannot infer multiple unknown dimensions in shape [?,?].

Surprisingly, if I change the shape of one of the inputs ([1, 2] => [None, 2]):

x = tf.placeholder(tf.float64, [None, 2], name="x")
y = tf.placeholder(tf.float64, [1, 2], name="y")
z = tf.add(x, y, name="z")

tf.gradients(z, [x, y])

everything works fine and I can successfully load the graph.
I don't see a reason why I am not allowed to load the graph in the first example.

I've made all the necessary checks and I am sure the graphs I create in Python
are correct (I can run them, and I get correct values for outputs and gradients).
Also, when I run the second graph (that was created in Python, and that I am able to load)
using C API, everything works fine.