parallel_for: use absl wrapper and add converter for SoftplusGrad

[artsobolev]

Consider the following snippet

import tensorflow as tf
from tensorflow.python.ops.parallel_for import gradients

x = tf.placeholder(tf.float32, (None,))
y = tf.nn.softplus(x)

print(gradients.jacobian(y, x))

On the current master it fails with

Traceback (most recent call last):
  File "test.py", line 7, in <module>
    print(gradients.jacobian(y, x))
  File "/Users/asobolev/dev/tensorflow/pfor_test/venv/lib/python3.6/site-packages/tensorflow/python/ops/parallel_for/gradients.py", line 58, in jacobian
    pfor_outputs = control_flow_ops.pfor(loop_fn, output_size)
  File "/Users/asobolev/dev/tensorflow/pfor_test/venv/lib/python3.6/site-packages/tensorflow/python/ops/parallel_for/control_flow_ops.py", line 122, in pfor
    outputs.append(converter.convert(loop_fn_output))
  File "/Users/asobolev/dev/tensorflow/pfor_test/venv/lib/python3.6/site-packages/tensorflow/python/ops/parallel_for/pfor.py", line 1075, in convert
    output = self._convert_helper(y)
  File "/Users/asobolev/dev/tensorflow/pfor_test/venv/lib/python3.6/site-packages/tensorflow/python/ops/parallel_for/pfor.py", line 1214, in _convert_helper
    if flags.FLAGS.op_conversion_fallback_to_while_loop:
  File "/Users/asobolev/dev/tensorflow/pfor_test/venv/lib/python3.6/site-packages/absl/flags/_flagvalues.py", line 488, in __getattr__
    raise _exceptions.UnparsedFlagAccessError(error_message)
absl.flags._exceptions.UnparsedFlagAccessError: Trying to access flag --op_conversion_fallback_to_while_loop before flags were parsed.

This is caused by the fact that no converter is registered for the SoftplusGrad op, and when parallel_for sees no converter defined, it tries to fall back to a while loop, if this is allowed with a special CLI flag. The problem, though, is that it uses raw absl call which requires one to first explicitly parse command line arguments. To mitigate that, one should use a flags wrapper from the tensorflow.platform, which is what I did.

Also, this pull introduces a converter for SoftplusGrad in the same way as, say, SigmoidGrad.

[artsobolev]

@agarwal-ashish, any news?