Using Google Cloud Storage (gs://) paths fails

[activatedgeek]

Description

Using Google Cloud Storage for --data_dir fails. Using gs paths like gs://my-bucket/path/to/folder fails.

NOTE: This is most likely a Tensorflow bug than Tensor2Tensor bug but keeping it here for the record. ...

Environment information

OS: macOS 10.12

$ pip freeze | grep tensor
tensor2tensor==1.6.3
tensorboard==1.8.0
tensorflow==1.8.0

$ python -V
Python 3.6.5

For bugs: reproduction and error logs

# Steps to reproduce:
t2t-datagen --tmp_dir=/tmp --problem=translate_ende_wmt32k --data_dir=gs://my_bucket/path

# Error logs:
Traceback (most recent call last):
  File "/.virtualenvs/env3/bin/t2t-datagen", line 27, in <module>
    tf.app.run()
  File "/.virtualenvs/env3/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 126, in run
    _sys.exit(main(argv))
  File "/.virtualenvs/env3/bin/t2t-datagen", line 23, in main
    t2t_datagen.main(argv)
  File "/.virtualenvs/env3/lib/python3.6/site-packages/tensor2tensor/bin/t2t_datagen.py", line 171, in main
    tf.gfile.MakeDirs(FLAGS.data_dir)
  File "/.virtualenvs/env3/lib/python3.6/site-packages/tensorflow/python/lib/io/file_io.py", line 374, in recursive_create_dir
    pywrap_tensorflow.RecursivelyCreateDir(compat.as_bytes(dirname), status)
  File "/.virtualenvs/env3/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 519, in __exit__
    c_api.TF_GetCode(self.status.status))
tensorflow.python.framework.errors_impl.InvalidArgumentError: 'object' must be a non-empty string. (File: gs://my_bucket/)

Fixes/Hacks

Not recommended but a good enough to isolate the cause.

• This issue can be fixed by manually creating the bucket folder in GCS and commenting out tf.gfile.MakeDirs(FLAGS.data_dir).

• Instead of using multi-level bucket paths, just use the root path - gs://my_bucket.

The error comes from the pywrap_tensorflow binary and cannot be debugged further without going into Tensorflow source I believe.

[activatedgeek]

/cc @lukaszkaiser

[activatedgeek]

This also doesn't work with TensorFlow 1.7.0.