Reduce usage of tf.compat.v1 APIs

[stephanwlee]

TensorBoard have migrated to TF 2.0 APIs but there are some usages of V1 APIs due to not-so-compatible changes. Since we are interested in moving off of tf.compat.v1, we will either have to find a replacement or introduce some changes in behavior.

Do note that this list exclude v1 API usages in tests (they are mainly written in graph style and is tracked in #1705) and TB v1 summary.

• Converted tf.app.run to absl.app.run #1725: tf.compat.v1.app -> absl.app
• Use absl.flags instead of tf.flags #1726: tf.compat.v1.flags -> absl.flags
• Use v2 API for GFile #1733: tf.compat.v1.gfile.GFile -> io.gfile.GFile
• Use v2 API for GFile #1733: tf.compat.v1.gfile.Open -> io.gfile.GFile
• Use the renamed image.resize TF API #1734: tf.compat.v1.image.resize_images -> tf.image.resize
• Use the renamed image.resize TF API #1734: tf.compat.v1.image.resize_nearest_neighbor -> tf.image.resize
• Change TF graph proto to TB graph proto #1730: tf.compat.v1.GraphDef -> TB proto
• Change TF graph proto to TB graph proto #1730: tf.compat.v1.MetaGraphDef -> TB proto
• [CLEANUP] tf.RunMetadata to tb.RunMetadata #1721: tf.compat.v1.RunMetadata -> TB proto
• [CLEANUP] Change tf.SessionLog.* to TB.SessionLog.* #1720: tf.compat.v1.SessionLog -> TB proto
• Update TensorBoard to import pywrap_tensorflow from tensorflow.python #2096: tf.compat.v1.pywrap_tensorflow.NewCheckpointReader -> tf.train.load_checkpoint
• replace deprecated tf.compat.v1.train.checkpoint_exists() #2157: tf.compat.v1.checkpoint_exists -> tf.io.gfile.glob() for the prefix
• tf.compat.v1.Session -> Use eager mode
• tf.contrib.ffmpeg -> tf.audio.encode_wav or python stdlib's wave module
• tf.contrib.summary.create_db_writer -> replace with python implementation (obsolete)
• tf.compat.v1.pywrap_tensorflow.EventsWriter -> migrated to tensorflow.python._pywrap_events_writer private API via Update TensorBoard to import pywrap_tensorflow from tensorflow.python #2096 and then Use _pywrap_events_writer #2683.
• tf.compat.v1.io.tf_record_iterator -> tracked in Replace tf_record_iterator with a non-deprecated way to read TFRecords #1711
• tf.compat.v1.train.summary_iterator - should become our own wrapper of tf_record_iterator for the places we need it (only tests)
• Change EventFileLoader to use tf_record_iterator when possible #3185: tf.compat.v1.pywrap_tensorflow.PyRecordReader_New -> converted to be only a legacy fallback for when we can't use tf.compat.v1.io.tf_record_iterator

[nfelt]

Edited to note that I think tf.compat.v1.train.summary_iterator should also be tracked in #1711. tf.train.summary_iterator should go away in TF 2.0, and we should replace it with whatever we resolve #1711 with.

[thealphacod3r]

@stephanwlee @nfelt can I work on the remaining ones, the two mentioned
tf.compat.v1.Session -> Use eager mode
tf.contrib.ffmpeg -> Have to find the replacement ??

[stephanwlee]

I recently learned about https://github.com/tensorflow/io which now includes ffmpeg.

[nfelt]

I think tensorflow/io's ffmpeg support is only for VideoDataset, i.e. it's only for reading videos.

For the v2 audio summary op we can just use tf.audio.encode_wav() now that it exists after tensorflow/tensorflow@b30d3f5.

For the v1 audio summary op and the v1 audio_pb function (which uses contrib ffmpeg via PersistentOpEvaluator), we could either backport the use of tf.audio.encode_wav(), or just leave them as only working with TF 1.x.

For the v2 equivalent of audio_pb, we should just use the python standard library's wave module: https://docs.python.org/2/library/wave.html

Lastly, the Beholder plugin uses ffmpeg to encode videos, but it literally just shells out to ffmpeg so there's no usage of tf.contrib.ffmpeg there.

@amanp592 if you'd like to work on any of this, this could also be a good place to start.

[wchargin]

fwiw, I have noted extremely poor performance of the wave module in
the past. The following blog post describes a way to improve the default
performance from 10m24s to 51s (for 5 minutes of audio output):

https://soledadpenades.com/posts/2009/fastest-way-to-generate-wav-files-in-python-using-the-wave-module/

By contrast, tf.audio.encode_audio completes in 0.271s (hundreds of
times faster than even the “fast” version above):

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os
import tempfile
import time


class Timer(object):
  def __init__(self):
    self._last = time.time()

  def log(self, msg):
    now = time.time()
    print("[%0.3f] %s" % (now - self._last, msg))
    self._last = now


SAMPLE_LEN = 13230000


def main():
  t = Timer()
  import tensorflow as tf
  t.log("Loaded TensorFlow.")
  audio = tf.random.uniform(shape=(SAMPLE_LEN, 2)) * 2 - 1
  t.log("Generated audio (%d samples)." % SAMPLE_LEN)
  encoded = tf.audio.encode_wav(audio, sample_rate=44100).numpy()
  t.log("Encoded audio.")
  filename = os.path.join(tempfile.mkdtemp(), "noise.wav")
  with open(filename, "w") as outfile:
    outfile.write(encoded)
  t.log("Wrote to: %s" % filename)


if __name__ == "__main__":
  main()

[0.729] Loaded TensorFlow.
[0.102] Generated audio (13230000 samples).
[0.148] Encoded audio.
[0.021] Wrote to: /tmp/tmp0Jqp0u/noise.wav

All this to say: we should benchmark, at least roughly, before and after
any such change.

[nfelt]

A belated response to the comment about WAV encoding - I agree that we should benchmark.

That said, I am not sure the above article really establishes that the wave module itself is a bottleneck. They aren't precise about where the timing was done, but it doesn't look like they're differentiating between the loop that generates audio values and the actual call to write out the data. I'm not terribly surprised that python is much slower than C++ if you ask it to generate a random value and preprocess it in a python for loop 13230000 times.

Looking at the wave source, if you ignore big-endian data conversion, writeframesraw(data) essentially just boils down to self._file.write(data) - there's actually no real processing happening. I think the wave module mostly just takes care of writing out the header and doing bookkeeping.
https://github.com/python/cpython/blob/7c2c01f02a1821298a62dd16ecc3a12da663e14b/Lib/wave.py#L421-L440

So I don't see any particular reason why wave would have to be slower than writing to a regular file in python, if the logic that we write to transform the input into bytes can be done efficiently. And this transformation is pretty simple - the article above just uses int16s and then struct.pack('h', value), and equivalently in numpy we can do np.array(values, dtype=np.int16).tobytes().

[HebaGamalElDin]

what for this line of code : 'tf.compat.v1.summary.image'
gets this error : 'AttributeError: 'module' object has no attribute 'v1' .

[wchargin]

@HebaGamalElDin: The tf.compat.v1 module was introduced in TensorFlow
version 1.13, and the tf.compat.v2 module will be in TensorFlow 1.14
(RC currently available).

If you’re seeing that error when running TensorBoard, please open a
separate bug report and be sure to include the requested environment
information.

[HebaGamalElDin]

@wchargin : what for python 2.7 , as almost Tensorrflow version 1.13 , 1.14 isn't applicable for python 2.7 .

[wchargin]

@HebaGamalElDin: TensorFlow versions 1.13 and 1.14 work just fine on
Python 2.7:

>>> import tensorflow, sys
>>> sys.version
'2.7.16 (default, Apr  6 2019, 01:42:57) \n[GCC 7.3.0]'
>>> tensorflow.__version__
'1.13.1'

>>> import tensorflow, sys
>>> sys.version
'2.7.16 (default, Apr  6 2019, 01:42:57) \n[GCC 7.3.0]'
>>> tensorflow.__version__
'1.14.0-rc0'

TensorFlow will support Python 2.7 at least until its end of life,
which is still seven months out.

[HebaGamalElDin]

@wchargin : I have searched more for that version of tensorflow for windows but I just found the applicable versions for Linux and mac OS only .

[wchargin]

@HebaGamalElDin: I was able to install TensorFlow 1.13 on Windows
without difficulty:

If you continue to have problems installing TensorFlow, please consult
Stack Overflow or open an issue on the TensorFlow repository:
https://github.com/tensorflow/tensorflow/issues/new?template=10-build-installation-issue.md