Investigating unification for TensorFlow graphs/objects

investigate_tf_unify.py

import tensorflow as tf
import tensorflow_probability as tfp

from tensorflow.python.framework import ops

ops.disable_eager_execution()

x_tf = tf.constant(1, name='x', dtype=tf.float64)

# An `Operation`
x_op = x_tf.op
x_op.type
#
# Where's the value (i.e. 1), name and dtype in the Op inputs?
#
list(x_op.inputs)
x_op.outputs

# ...
# the REAL operator is an `OpDef` (wtf?)
# yes, it's one of those protobuf objects
x_op_def = x_op.op_def

x_op_def.name == x_op.type

# is it `x_op.inputs`?
x_op_def.input_arg

# we've found our missing parameters/inputs:
list(x_op_def.attr)

# but those are just the "definition" of an Op, the
# actual instantiated Op object is `x_op` above.
# now, we know what to ask for in the protobuf `x_op` object.
x_op.get_attr("value")
x_op.get_attr("dtype")



#
# Let's use this to unify TF objects/graphs.
#
from unification import unify, reify, var, variables


y_tf = tf.constant(2, name='y', dtype=tf.float64)
z_tf = tf.add(x_tf, y_tf, name='z')

# How?
# unify(z_tf, ('add', var('x'), 2), {})


unify((z_tf.op.type, z_tf.op.inputs[0], z_tf.op.inputs[1]),
      ('Add', var('x'), 2),
      {})

unify(var('x'), z_tf.op.inputs[0], {})

unify(z_tf, ..., {})