tensorflow map_fn TensorArray has inconsistent shapes

Question

I am playing around with the map_fn function, and noticed that it outputs a TensorArray, which should mean it is capable of outputting "jagged" tensors (where the tensors on the inside have different first dimensions.

I tried to see this in action with this code:

import tensorflow as tf
import numpy as np

NUM_ARRAYS = 1000
MAX_LENGTH = 1000

lengths = tf.placeholder(tf.int32)
tArray = tf.map_fn(lambda x: tf.random_normal((x,), 0, 1),
                   lengths,
                   dtype=tf.float32) # Should return a TensorArray.

# startTensor =  tf.random_normal((tf.reduce_sum(lengths),), 0, 1)
# tArray = tf.TensorArray(tf.float32, NUM_ARRAYS)
# tArray = tArray.split(startTensor, lengths)
# outArray = tArray.concat()


with tf.Session() as sess:
    outputArray, l = sess.run(
        [tArray, lengths],
        feed_dict={lengths: np.random.randint(MAX_LENGTH, size=NUM_ARRAYS)})
    print outputArray.shape, l

however got the error:

"TensorArray has inconsistent shapes. Index 0 has shape: [259] but index 1 has shape: [773]"

This of course comes as a surprise to me since I am under the impression that TensorArrays should be able to handle it. Am I wrong?

Answer 1

While the tf.map_fn() does use tf.TensorArray objects internally, and a tf.TensorArray can hold objects of different size, this program won't work as-is because tf.map_fn() converts its tf.TensorArray result back to a tf.Tensor by stacking the elements together, and it is this operation that fails.

You can however implement the tf.TensorArray-based using the lower-lever tf.while_loop() op instead:

lengths = tf.placeholder(tf.int32)
num_elems = tf.shape(lengths)[0]
init_array = tf.TensorArray(tf.float32, size=num_elems)

def loop_body(i, ta):
  return i + 1, ta.write(i, tf.random_normal((lengths[i],), 0, 1))

_, result_array = tf.while_loop(
    lambda i, ta: i < num_elems, loop_body, [0, init_array])

Answer 2

Building upon mrry's answer, some more examples that can be run under TF2.x

import tensorflow as tf
# ================= example 1 ==================
num_elems = 5
init_array = tf.TensorArray(tf.float32, size=num_elems, infer_shape=False)
lengths = tf.range(0, 5)
def loop_body(i, ta):
  return i + 1, ta.write(i, tf.random.normal((lengths[i],), 0, 1))

_, result_array = tf.while_loop(
    lambda i, ta: i < num_elems, loop_body, [0, init_array])

for i in range(num_elems):
    print(result_array.read(i))

# ================== example 2 ==================
# TensorArray whose size is known at run time and shapes of elements
# are not necessarily the same
ta = tf.TensorArray(tf.float32, size=0, dynamic_size=True, infer_shape=False)

# init ta with some mock data
ta = ta.write(0, 0.0)
ta = ta.write(1, 1.0)
ta = ta.write(2, tf.constant([2.0, 2.0]))

# loop body
def loop_body(i, t):
    val = t.read(i)
    # do something
    t = t.write(i, tf.multiply(2.0, val))
    return i+1, t

# stop condition for while loop
index = tf.constant(0)
cond = lambda i, t: tf.less(i, t.size())

# results
i = tf.constant(0)
_, result_array = tf.while_loop(cond, loop_body, [i, ta])

for i in range(result_array.size()):
    print(result_array.read(i))