CODEWITHSUNDEEP
pythontensorflow
Richard_wth
Richard_wth

Reputation: 702

TensorFlow 1.5.0-rc0: error using `tf.app.flags`

The following flags were defined in a misc_fun.py file to include machine and directories info:

import tensorflow as tf
flags = tf.app.flags
FLAGS = flags.FLAGS
# definitions
flags.DEFINE_string(
    'DEFAULT_IN',
    '~/PycharmProjects/myNN/Data/',
    """Default input folder.""")
...

It worked fine in TensorFlow 1.0 - 1.4 versions (with Pycharm). After updating to TensorFlow 1.5.-rc0, the following error occurred:

Usage:

from misc_fun import FLAGS
FLAGS.DEFAULT_IN = FLAGS.DEFAULT_DOWNLOAD  # change default input folder

Error:

UnparsedFlagAccessError: Trying to access flag --DEFAULT_DOWNLOAD before flags were parsed.

However print(FLAGS) worked fine, which gives: 

misc_fun:
  --DEFAULT_DOWNLOAD: default download folder for large datasets.
    (default: '/home/username/Downloads/Data/')
  --DEFAULT_IN: default input folder.
    (default: '~/PycharmProjects/myNN/Data/')
...

I tried FLAGS = flags.FLAGS(sys.argv), resulting in the following error:

UnrecognizedFlagError: Unknown command line flag 'f'

Although there is a workaround using the class object, I wonder what could be the problem here.

Upvotes: 4

Views: 7019

Answers (3)

aysljc
aysljc

Reputation: 1018

I have tried adding the following line below.

tf.app.flags.DEFINE_string('f', '', 'kernel')

This solution is different from others in that it is simple and easy to try. You just need to add this into your code, and it doesn't change your system. Please let me know if this solution helps solve other people's problems.

The reference for this solution is from a Chinese website: https://blog.csdn.net/qq_39956625/article/details/80500291

Upvotes: 8

PouyaB
PouyaB

Reputation: 849

Alternatively you can use FLAGS(sys.argv, known_only=True) to parse all related flags (the ones defined using tf.app.flags.DEFINE_xxx). This will release any other args that are not known. Useful if you have some command line arguments that are not related to TF.

Upvotes: 0

rerx
rerx

Reputation: 1183

With 1.5.0-rc0 the Tensorflow maintainers have switched tf.app.flags to the flags module from abseil. Unfortunately, it is not 100% API compatible to the previous implementation. I worked around your problem with something like

remaining_args = FLAGS([sys.argv[0]] + [flag for flag in sys.argv if flag.startswith("--")])
assert(remaining_args == [sys.argv[0]])

before accessing the FLAGS object the first time.

Upvotes: 6

Related Questions