tensorflow - InvalidArgumentError: The node has inputs from different frames

Question

I was playing around with Tensorflow and ran into a problem with this code:

def process_tree_tf(matrix, weights, idxs, name=None):

    with tf.name_scope(name, "process_tree", [tree, weights, idxs]).as scope():
         loop_index = tf.sub(tf.shape(matrix)[0], 1)
         loop_vars = loop_index, matrix, idxs, weights

         def loop_condition(loop_idx, *_):
             return tf.greater(loop_idx, 0)

         def loop_body(loop_idx, mat, idxs, weights):
             x = mat[loop_idx]
             w = weights
             bias = tf.Variable(tf.constant(0.1, [2], dtype=tf.float64)) # Here?

             ...
             return loop_idx-1, mat, idxs, weights

         return tf.while_loop(loop_condition, loop_body, loop_vars, name=scope)[1]

I'm evaluating the function this way:

height = 2
width = 2
nodes = 4
matrix = np.ones((nodes, width+height))
weights = np.ones((width+height, width))/100
idxs = [0,0,1,2]
with tf.Session as sess():
    sess.run(tf.global_variables_initializer()) # Error Here!
    r = process_tree_tf(matrix, weights, idxs)
    print(r.eval())

I'm getting this error:

InvalidArgumentError: The node 'process_tree_tf/Variable/Assign' has inputs from different frames. The input 'process_tree_tf/Const_1' is in frame 'process_tree_tf/process_tree_tf/'. The input 'process_tree_tf/Variable' is in frame ''.

Weirdly, if I restart the kernel in jupyter notebook and run all again, I get this error:

FailedPreconditionError (see above for traceback): Attempting to use uninitialized value bias [[Node: bias/read = IdentityT=DT_FLOAT, _class=["loc:@bias"], _device="/job:localhost/replica:0/task:0/cpu:0"]]

I tried using this instead: bias = tf.get_variable("bias", shape=[2], initializer=tf.constant_initializer(0.1)) but that didn't work either.

I'm very sorry if I'm overlooking something obvious here, but I'd really appreciate it if someone could tell me where I'm going wrong.

Thank you very much!

Answer 1

This is actually a subtle issue with tf.Variable objects in TensorFlow's tf.while_loop(). TensorFlow becomes confused because it appears that the tf.constant() with which you're initializing the variable is a value created inside the loop (even though it's clearly loop invariant), but all variables get hoisted outside the loop. The easiest resolution is to move the creation of the variable outside the loop:

def process_tree_tf(matrix, weights, idxs, name=None):

    with tf.name_scope(name, "process_tree", [tree, weights, idxs]).as scope():
         loop_index = tf.sub(tf.shape(matrix)[0], 1)
         loop_vars = loop_index, matrix, idxs, weights

         # Define the bias variable outside the loop to avoid problems.
         bias = tf.Variable(tf.constant(0.1, [2], dtype=tf.float64)) 

         def loop_condition(loop_idx, *_):
             return tf.greater(loop_idx, 0)

         def loop_body(loop_idx, mat, idxs, weights):
             x = mat[loop_idx]
             w = weights

             # You can still refer to `bias` in here, and the loop body
             # will capture it appropriately.
             ...
             return loop_idx-1, mat, idxs, weights

         return tf.while_loop(loop_condition, loop_body, loop_vars, name=scope)[1]

(Another possible resolution would be to use a tf.constant_initializer() rather than a tf.constant() when creating the variable.)