Relay TF frontend nested switch vs. tf.graph_util.convert_variables_to_constants

Hi folks,

It seems that tf.graph_util.convert_variables_to_constants will mess with relay TF FE. Specifically, since switch nodes get removed from the graph by the cmd, TF FE will fail to match merge nodes with corresponding switch node (right now it seems they’re matched by name here)

This is observed with a simple test model with nested branches. Attached is a python script to reproduce:

import tensorflow as tf

def gen_nested_cond():
    a = tf.constant(2., dtype = tf.float32)
    b = tf.constant(3., dtype = tf.float32)

    def fn1(a, b):
        def nest_fn1(a, b):
            return tf.add(a, b)

        def nest_fn2(a, b):
            return tf.multiply(a, b)

        res = tf.cond(tf.less(a, b), lambda: nest_fn1(a, b), lambda: nest_fn2(a, b))
        return res

    def fn2(a, b):
        return tf.add(a, b)

    r = tf.cond(tf.less(a, b), lambda: fn1(a, b), lambda: fn2(a, b))
    with tf.Session() as sess:
        print("start save model:")

        graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
        graph_def = tf.graph_util.convert_variables_to_constants(sess, graph_def, ['cond_1/Merge'])  # This line breaks TF FE
        with"./nested_cond_new_const.pb", "wb") as f:

Is this expected behavior, or am I using/understanding it wrong?

I don’t think we have ever tried convert_variables_to_constants. Is TFParser not sufficient for model loading?

BTW, it looks like convert_variables_to_constants is deprecated in tf.

I mean some models are saved after this transformation, and TFParser will throw out error msg when handling merge nodes. E.g

  File "workplace/incubator-tvm/python/tvm/relay/frontend/", line 2941, in _convert_control_flow_operator
    branch = self._branches[node_name_prefix]
KeyError: 'cond/cond'