Remote training I/O once more (#3437)

* Redo remote I/O changes once more; this time without messing with taskcluster

* Add bin changes

* Fix merge-induced issue?

* For the interleaved case with multiple collections, unpack audio on the fly

To reproduce the previous failure

rm data/smoke_test/ldc93s1.csv
rm data/smoke_test/ldc93s1.sdb
rm -rf /tmp/ldc93s1_cache_sdb_csv
rm -rf /tmp/ckpt_sdb_csv
rm -rf /tmp/train_sdb_csv

./bin/run-tc-ldc93s1_new_sdb_csv.sh 109 16000
python -u DeepSpeech.py --noshow_progressbar --noearly_stop --train_files ./data/smoke_test/ldc93s1.sdb,./data/smoke_test/ldc93s1.csv --train_batch_size 1 --feature_cache /tmp/ldc93s1_cache_sdb_csv --dev_files ./data/smoke_test/ldc93s1.sdb,./data/smoke_test/ldc93s1.csv --dev_batch_size 1 --test_files ./data/smoke_test/ldc93s1.sdb,./data/smoke_test/ldc93s1.csv --test_batch_size 1 --n_hidden 100 --epochs 109 --max_to_keep 1 --checkpoint_dir /tmp/ckpt_sdb_csv --learning_rate 0.001 --dropout_rate 0.05 --export_dir /tmp/train_sdb_csv --scorer_path data/smoke_test/pruned_lm.scorer --audio_sample_rate 16000

* Attempt to preserve length information with a wrapper around `map()`… this gets pretty python-y

* Call the right `__next__()`

* Properly implement the rest of the map wrappers here……

* Fix trailing whitespace situation and other linter complaints

* Remove data accidentally checked in

* Fix overlay augmentations

* Wavs must be open in rb mode if we're passing in an external file pointer -- this confused me

* Lint whitespace

* Revert "Fix trailing whitespace situation and other linter complaints"

This reverts commit c3c45397a2f98e9b00d00c18c4ced4fc52475032.

* Fix linter issue but without such an aggressive diff

* Move unpack_maybe into sample_collections

* Use unpack_maybe in place of duplicate lambda

* Fix confusing comment

* Add clarifying comment for on-the-fly unpacking

bin/compare_samples.py

@@ -15,8 +15,8 @@ def fail(message):
 
 
 def compare_samples():
-    sample1 = load_sample(CLI_ARGS.sample1)
+    sample1 = load_sample(CLI_ARGS.sample1).unpack()
-    sample2 = load_sample(CLI_ARGS.sample2)
+    sample2 = load_sample(CLI_ARGS.sample2).unpack()
     if sample1.audio_format != sample2.audio_format:
         fail('Samples differ on: audio-format ({} and {})'.format(sample1.audio_format, sample2.audio_format))
     if sample1.duration != sample2.duration:

training/deepspeech_training/train.py

@@ -35,6 +35,7 @@ from .util.feeding import create_dataset, audio_to_features, audiofile_to_featur
 from .util.flags import create_flags, FLAGS
 from .util.helpers import check_ctcdecoder_version, ExceptionBox
 from .util.logging import create_progressbar, log_debug, log_error, log_info, log_progress, log_warn
+from .util.io import open_remote, remove_remote, listdir_remote, is_remote_path, isdir_remote
 
 check_ctcdecoder_version()
 
@@ -512,9 +513,10 @@ def train():
     best_dev_path = os.path.join(FLAGS.save_checkpoint_dir, 'best_dev')
 
     # Save flags next to checkpoints
+    if not is_remote_path(FLAGS.save_checkpoint_dir):
         os.makedirs(FLAGS.save_checkpoint_dir, exist_ok=True)
     flags_file = os.path.join(FLAGS.save_checkpoint_dir, 'flags.txt')
-    with open(flags_file, 'w') as fout:
+    with open_remote(flags_file, 'w') as fout:
         fout.write(FLAGS.flags_into_string())
 
     with tfv1.Session(config=Config.session_config) as session:
@@ -541,7 +543,7 @@ def train():
                 feature_cache_index = FLAGS.feature_cache + '.index'
                 if epoch % FLAGS.cache_for_epochs == 0 and os.path.isfile(feature_cache_index):
                     log_info('Invalidating feature cache')
-                    os.remove(feature_cache_index)  # this will let TF also overwrite the related cache data files
+                    remove_remote(feature_cache_index)  # this will let TF also overwrite the related cache data files
 
             # Setup progress bar
             class LossWidget(progressbar.widgets.FormatLabel):
@@ -810,13 +812,13 @@ def export():
 
         output_filename = FLAGS.export_file_name + '.pb'
         if FLAGS.remove_export:
-            if os.path.isdir(FLAGS.export_dir):
+            if isdir_remote(FLAGS.export_dir):
                 log_info('Removing old export')
-                shutil.rmtree(FLAGS.export_dir)
+                remove_remote(FLAGS.export_dir)
 
         output_graph_path = os.path.join(FLAGS.export_dir, output_filename)
 
-        if not os.path.isdir(FLAGS.export_dir):
+        if not is_remote_path(FLAGS.export_dir) and not os.path.isdir(FLAGS.export_dir):
             os.makedirs(FLAGS.export_dir)
 
         frozen_graph = tfv1.graph_util.convert_variables_to_constants(
@@ -829,7 +831,7 @@ def export():
             dest_nodes=output_names)
 
         if not FLAGS.export_tflite:
-            with open(output_graph_path, 'wb') as fout:
+            with open_remote(output_graph_path, 'wb') as fout:
                 fout.write(frozen_graph.SerializeToString())
         else:
             output_tflite_path = os.path.join(FLAGS.export_dir, output_filename.replace('.pb', '.tflite'))
@@ -840,7 +842,7 @@ def export():
             converter.allow_custom_ops = True
             tflite_model = converter.convert()
 
-            with open(output_tflite_path, 'wb') as fout:
+            with open_remote(output_tflite_path, 'wb') as fout:
                 fout.write(tflite_model)
 
         log_info('Models exported at %s' % (FLAGS.export_dir))
@@ -851,7 +853,7 @@ def export():
         FLAGS.export_model_version))
 
     model_runtime = 'tflite' if FLAGS.export_tflite else 'tensorflow'
-    with open(metadata_fname, 'w') as f:
+    with open_remote(metadata_fname, 'w') as f:
         f.write('---\n')
         f.write('author: {}\n'.format(FLAGS.export_author_id))
         f.write('model_name: {}\n'.format(FLAGS.export_model_name))
@@ -873,6 +875,10 @@ def export():
 def package_zip():
     # --export_dir path/to/export/LANG_CODE/ => path/to/export/LANG_CODE.zip
     export_dir = os.path.join(os.path.abspath(FLAGS.export_dir), '') # Force ending '/'
+    if is_remote_path(export_dir):
+        log_error("Cannot package remote path zip %s. Please do this manually." % export_dir)
+        return
+
     zip_filename = os.path.dirname(export_dir)
     
     shutil.copy(FLAGS.scorer_path, export_dir)
@@ -959,7 +965,7 @@ def main(_):
         tfv1.reset_default_graph()
         FLAGS.export_tflite = True
 
-        if os.listdir(FLAGS.export_dir):
+        if listdir_remote(FLAGS.export_dir):
             log_error('Directory {} is not empty, please fix this.'.format(FLAGS.export_dir))
             sys.exit(1)
 

training/deepspeech_training/util/audio.py

@@ -8,6 +8,7 @@ import numpy as np
 
 from .helpers import LimitingPool
 from collections import namedtuple
+from .io import open_remote, remove_remote, copy_remote, is_remote_path
 
 AudioFormat = namedtuple('AudioFormat', 'rate channels width')
 
@@ -117,15 +118,19 @@ class Sample:
         self.audio_type = new_audio_type
 
 
-def _change_audio_type(sample_and_audio_type):
+def _unpack_and_change_audio_type(sample_and_audio_type):
-    sample, audio_type, bitrate = sample_and_audio_type
+    packed_sample, audio_type, bitrate = sample_and_audio_type
+    if hasattr(packed_sample, 'unpack'):
+        sample = packed_sample.unpack()
+    else:
+        sample = packed_sample
     sample.change_audio_type(audio_type, bitrate=bitrate)
     return sample
 
 
-def change_audio_types(samples, audio_type=AUDIO_TYPE_PCM, bitrate=None, processes=None, process_ahead=None):
+def change_audio_types(packed_samples, audio_type=AUDIO_TYPE_PCM, bitrate=None, processes=None, process_ahead=None):
     with LimitingPool(processes=processes, process_ahead=process_ahead) as pool:
-        yield from pool.imap(_change_audio_type, map(lambda s: (s, audio_type, bitrate), samples))
+        yield from pool.imap(_unpack_and_change_audio_type, map(lambda s: (s, audio_type, bitrate), packed_samples))
 
 
 def get_audio_type_from_extension(ext):
@@ -168,29 +173,45 @@ class AudioFile:
         self.audio_format = audio_format
         self.as_path = as_path
         self.open_file = None
+        self.open_wav = None
         self.tmp_file_path = None
+        self.tmp_src_file_path = None
 
     def __enter__(self):
         if self.audio_path.endswith('.wav'):
-            self.open_file = wave.open(self.audio_path, 'r')
+            self.open_file = open_remote(self.audio_path, 'rb')
-            if read_audio_format_from_wav_file(self.open_file) == self.audio_format:
+            self.open_wav = wave.open(self.open_file)
+            if read_audio_format_from_wav_file(self.open_wav) == self.audio_format:
                 if self.as_path:
+                    self.open_wav.close()
                     self.open_file.close()
                     return self.audio_path
-                return self.open_file
+                return self.open_wav
+            self.open_wav.close()
             self.open_file.close()
+
+        # If the format isn't right, copy the file to local tmp dir and do the conversion on disk
+        if is_remote_path(self.audio_path):
+            _, self.tmp_src_file_path = tempfile.mkstemp(suffix='.wav')
+            copy_remote(self.audio_path, self.tmp_src_file_path)
+            self.audio_path = self.tmp_file_path
+
         _, self.tmp_file_path = tempfile.mkstemp(suffix='.wav')
         convert_audio(self.audio_path, self.tmp_file_path, file_type='wav', audio_format=self.audio_format)
         if self.as_path:
             return self.tmp_file_path
-        self.open_file = wave.open(self.tmp_file_path, 'r')
+        self.open_wav = wave.open(self.tmp_file_path, 'rb')
-        return self.open_file
+        return self.open_wav
 
     def __exit__(self, *args):
         if not self.as_path:
+            self.open_wav.close()
+            if self.open_file:
                 self.open_file.close()
         if self.tmp_file_path is not None:
             os.remove(self.tmp_file_path)
+        if self.tmp_src_file_path is not None:
+            os.remove(self.tmp_src_file_path)
 
 
 def read_frames(wav_file, frame_duration_ms=30, yield_remainder=False):
@@ -320,6 +341,7 @@ def read_opus(opus_file):
 
 
 def write_wav(wav_file, pcm_data, audio_format=DEFAULT_FORMAT):
+    # wav_file is already a file-pointer here
     with wave.open(wav_file, 'wb') as wav_file_writer:
         wav_file_writer.setframerate(audio_format.rate)
         wav_file_writer.setnchannels(audio_format.channels)

training/deepspeech_training/util/augmentations.py

@@ -8,7 +8,7 @@ import numpy as np
 from multiprocessing import Queue, Process
 from .audio import gain_db_to_ratio, max_dbfs, normalize_audio, AUDIO_TYPE_NP, AUDIO_TYPE_PCM, AUDIO_TYPE_OPUS
 from .helpers import LimitingPool, int_range, float_range, pick_value_from_range, tf_pick_value_from_range, MEGABYTE
-from .sample_collections import samples_from_source
+from .sample_collections import samples_from_source, unpack_maybe
 
 BUFFER_SIZE = 1 * MEGABYTE
 SPEC_PARSER = re.compile(r'^(?P<cls>[a-z_]+)(\[(?P<params>.*)\])?$')
@@ -150,6 +150,12 @@ def _init_augmentation_worker(preparation_context):
     AUGMENTATION_CONTEXT = preparation_context
 
 
+def _load_and_augment_sample(timed_sample, context=None):
+    sample, clock = timed_sample
+    realized_sample = unpack_maybe(sample)
+    return _augment_sample((realized_sample, clock), context)
+
+
 def _augment_sample(timed_sample, context=None):
     context = AUGMENTATION_CONTEXT if context is None else context
     sample, clock = timed_sample
@@ -213,12 +219,12 @@ def apply_sample_augmentations(samples,
         context = AugmentationContext(audio_type, augmentations)
         if process_ahead == 0:
             for timed_sample in timed_samples():
-                yield _augment_sample(timed_sample, context=context)
+                yield _load_and_augment_sample(timed_sample, context=context)
         else:
             with LimitingPool(process_ahead=process_ahead,
                               initializer=_init_augmentation_worker,
                               initargs=(context,)) as pool:
-                yield from pool.imap(_augment_sample, timed_samples())
+                yield from pool.imap(_load_and_augment_sample, timed_samples())
     finally:
         for augmentation in augmentations:
             augmentation.stop()
@@ -256,6 +262,7 @@ class Overlay(SampleAugmentation):
         self.enqueue_process.start()
 
     def apply(self, sample, clock=0.0):
+        sample = unpack_maybe(sample)
         sample.change_audio_type(new_audio_type=AUDIO_TYPE_NP)
         n_layers = pick_value_from_range(self.layers, clock=clock)
         audio = sample.audio
@@ -265,6 +272,7 @@ class Overlay(SampleAugmentation):
             while overlay_offset < len(audio):
                 if self.current_sample is None:
                     next_overlay_sample = self.queue.get()
+                    next_overlay_sample = unpack_maybe(next_overlay_sample)
                     next_overlay_sample.change_audio_type(new_audio_type=AUDIO_TYPE_NP)
                     self.current_sample = next_overlay_sample.audio
                 n_required = len(audio) - overlay_offset

training/deepspeech_training/util/check_characters.py

@@ -19,6 +19,7 @@ import csv
 import os
 import sys
 import unicodedata
+from .io import open_remote
 
 def main():
     parser = argparse.ArgumentParser()
@@ -27,14 +28,14 @@ def main():
     parser.add_argument("-alpha", "--alphabet-format", help="Bool. Print in format for alphabet.txt", action="store_true")
     parser.add_argument("-unicode", "--disable-unicode-variants", help="Bool. DISABLE check for unicode consistency (use with --alphabet-format)", action="store_true")
     args = parser.parse_args()
-    in_files = [os.path.abspath(i) for i in args.csv_files.split(",")]
+    in_files = args.csv_files.split(",")
 
     print("### Reading in the following transcript files: ###")
     print("### {} ###".format(in_files))
 
     all_text = set()
     for in_file in in_files:
-        with open(in_file, "r") as csv_file:
+        with open_remote(in_file, "r") as csv_file:
             reader = csv.reader(csv_file)
             try:
                 next(reader, None)  # skip the file header (i.e. "transcript")

training/deepspeech_training/util/config.py

@@ -13,7 +13,7 @@ from .gpu import get_available_gpus
 from .logging import log_error, log_warn
 from .helpers import parse_file_size
 from .augmentations import parse_augmentations
-
+from .io import path_exists_remote
 
 class ConfigSingleton:
     _config = None
@@ -139,7 +139,7 @@ def initialize_globals():
     c.audio_step_samples = FLAGS.audio_sample_rate * (FLAGS.feature_win_step / 1000)
 
     if FLAGS.one_shot_infer:
-        if not os.path.exists(FLAGS.one_shot_infer):
+        if not path_exists_remote(FLAGS.one_shot_infer):
             log_error('Path specified in --one_shot_infer is not a valid file.')
             sys.exit(1)
 

training/deepspeech_training/util/downloader.py

@@ -2,6 +2,7 @@ import requests
 import progressbar
 
 from os import path, makedirs
+from .io import open_remote, path_exists_remote, is_remote_path
 
 SIMPLE_BAR = ['Progress ', progressbar.Bar(), ' ', progressbar.Percentage(), ' completed']
 
@@ -9,17 +10,18 @@ def maybe_download(archive_name, target_dir, archive_url):
     # If archive file does not exist, download it...
     archive_path = path.join(target_dir, archive_name)
 
-    if not path.exists(target_dir):
+    if not is_remote_path(target_dir) and not path.exists(target_dir):
         print('No path "%s" - creating ...' % target_dir)
         makedirs(target_dir)
 
-    if not path.exists(archive_path):
+    if not path_exists_remote(archive_path):
         print('No archive "%s" - downloading...' % archive_path)
         req = requests.get(archive_url, stream=True)
         total_size = int(req.headers.get('content-length', 0))
         done = 0
-        with open(archive_path, 'wb') as f:
+        with open_remote(archive_path, 'wb') as f:
             bar = progressbar.ProgressBar(max_value=total_size if total_size > 0 else progressbar.UnknownLength, widgets=SIMPLE_BAR)
+
             for data in req.iter_content(1024*1024):
                 done += len(data)
                 f.write(data)

training/deepspeech_training/util/evaluate_tools.py

@@ -10,7 +10,7 @@ from attrdict import AttrDict
 
 from .flags import FLAGS
 from .text import levenshtein
-
+from .io import open_remote
 
 def pmap(fun, iterable):
     pool = Pool()
@@ -124,5 +124,5 @@ def save_samples_json(samples, output_path):
         We set ensure_ascii=True to prevent json from escaping non-ASCII chars
         in the texts.
     '''
-    with open(output_path, 'w') as fout:
+    with open_remote(output_path, 'w') as fout:
         json.dump(samples, fout, default=float, ensure_ascii=False, indent=2)

training/deepspeech_training/util/helpers.py

@@ -78,6 +78,32 @@ class Interleaved:
         return self.len
 
 
+class LenMap:
+    """
+    Wrapper around python map() output object that preserves the original collection length
+    by implementing __len__.
+    """
+    def __init__(self, fn, iterable):
+        try:
+            self.length = len(iterable)
+        except TypeError:
+            self.length = None
+        self.mapobj = map(fn, iterable)
+
+    def __iter__(self):
+        self.mapobj = self.mapobj.__iter__()
+        return self
+
+    def __next__(self):
+        return self.mapobj.__next__()
+
+    def __getitem__(self, key):
+        return self.mapobj.__getitem__(key)
+
+    def __len__(self):
+        return self.length
+
+
 class LimitingPool:
     """Limits unbound ahead-processing of multiprocessing.Pool's imap method
     before items get consumed by the iteration caller.

training/deepspeech_training/util/io.py

@@ -0,0 +1,81 @@
+"""
+A set of I/O utils that allow us to open files on remote storage as if they were present locally and access
+into HDFS storage using Tensorflow's C++ FileStream API.
+Currently only includes wrappers for Google's GCS, but this can easily be expanded for AWS S3 buckets.
+"""
+import os
+from tensorflow.io import gfile
+
+
+def is_remote_path(path):
+    """
+    Returns True iff the path is one of the remote formats that this
+    module supports
+    """
+    return path.startswith('gs://') or path.startswith('hdfs://')
+
+
+def path_exists_remote(path):
+    """
+    Wrapper that allows existance check of local and remote paths like
+    `gs://...`
+    """
+    if is_remote_path(path):
+        return gfile.exists(path)
+    return os.path.exists(path)
+
+
+def copy_remote(src, dst, overwrite=False):
+    """
+    Allows us to copy a file from local to remote or vice versa
+    """
+    return gfile.copy(src, dst, overwrite)
+
+
+def open_remote(path, mode='r', buffering=-1, encoding=None, newline=None, closefd=True, opener=None):
+    """
+    Wrapper around open() method that can handle remote paths like `gs://...`
+    off Google Cloud using Tensorflow's IO helpers.
+
+    buffering, encoding, newline, closefd, and opener are ignored for remote files
+
+    This enables us to do:
+    with open_remote('gs://.....', mode='w+') as f:
+        do something with the file f, whether or not we have local access to it
+    """
+    if is_remote_path(path):
+        return gfile.GFile(path, mode=mode)
+    return open(path, mode, buffering=buffering, encoding=encoding, newline=newline, closefd=closefd, opener=opener)
+
+
+def isdir_remote(path):
+    """
+    Wrapper to check if remote and local paths are directories
+    """
+    if is_remote_path(path):
+        return gfile.isdir(path)
+    return os.path.isdir(path)
+
+
+def listdir_remote(path):
+    """
+    Wrapper to list paths in local dirs (alternative to using a glob, I suppose)
+    """
+    if is_remote_path(path):
+        return gfile.listdir(path)
+    return os.listdir(path)
+
+
+def glob_remote(filename):
+    """
+    Wrapper that provides globs on local and remote paths like `gs://...`
+    """
+    return gfile.glob(filename)
+
+
+def remove_remote(filename):
+    """
+    Wrapper that can remove local and remote files like `gs://...`
+    """
+    # Conditional import
+    return gfile.remove_remote(filename)

training/deepspeech_training/util/sample_collections.py

@@ -8,7 +8,7 @@ import tarfile
 from pathlib import Path
 from functools import partial
 
-from .helpers import KILOBYTE, MEGABYTE, GIGABYTE, Interleaved
+from .helpers import KILOBYTE, MEGABYTE, GIGABYTE, Interleaved, LenMap
 from .audio import (
     Sample,
     DEFAULT_FORMAT,
@@ -18,6 +18,7 @@ from .audio import (
     get_audio_type_from_extension,
     write_wav
 )
+from .io import open_remote, is_remote_path
 
 BIG_ENDIAN = 'big'
 INT_SIZE = 4
@@ -59,6 +60,37 @@ class LabeledSample(Sample):
         self.transcript = transcript
 
 
+class PackedSample:
+    """
+    A wrapper that we can carry around in an iterator and pass to a child process in order to
+    have the child process do the loading/unpacking of the sample, allowing for parallel file
+    I/O.
+    """
+    def __init__(self, filename, audio_type, label):
+        self.filename = filename
+        self.audio_type = audio_type
+        self.label = label
+
+    def unpack(self):
+        with open_remote(self.filename, 'rb') as audio_file:
+            data = audio_file.read()
+        if self.label is None:
+            s = Sample(self.audio_type, data, sample_id=self.filename)
+        s = LabeledSample(self.audio_type, data, self.label, sample_id=self.filename)
+        return s
+
+
+def unpack_maybe(sample):
+    """
+    Loads the supplied sample from disk (or the network) if the audio isn't loaded in to memory already.
+    """
+    if hasattr(sample, 'unpack'):
+        realized_sample = sample.unpack()
+    else:
+        realized_sample = sample
+    return realized_sample
+
+
 def load_sample(filename, label=None):
     """
     Loads audio-file as a (labeled or unlabeled) sample
@@ -69,21 +101,19 @@ def load_sample(filename, label=None):
         Filename of the audio-file to load as sample
     label : str
         Label (transcript) of the sample.
-        If None: return util.audio.Sample instance
+        If None: returned result.unpack() will return util.audio.Sample instance
-        Otherwise: return util.sample_collections.LabeledSample instance
+        Otherwise: returned result.unpack()  util.sample_collections.LabeledSample instance
 
     Returns
     -------
+    util.sample_collections.PackedSample, a wrapper object, on which calling unpack() will return
         util.audio.Sample instance if label is None, else util.sample_collections.LabeledSample instance
     """
     ext = os.path.splitext(filename)[1].lower()
     audio_type = get_audio_type_from_extension(ext)
     if audio_type is None:
         raise ValueError('Unknown audio type extension "{}"'.format(ext))
-    with open(filename, 'rb') as audio_file:
+    return PackedSample(filename, audio_type, label)
-        if label is None:
-            return Sample(audio_type, audio_file.read(), sample_id=filename)
-        return LabeledSample(audio_type, audio_file.read(), label, sample_id=filename)
 
 
 class DirectSDBWriter:
@@ -119,7 +149,7 @@ class DirectSDBWriter:
             raise ValueError('Audio type "{}" not supported'.format(audio_type))
         self.audio_type = audio_type
         self.bitrate = bitrate
-        self.sdb_file = open(sdb_filename, 'wb', buffering=buffering)
+        self.sdb_file = open_remote(sdb_filename, 'wb', buffering=buffering)
         self.offsets = []
         self.num_samples = 0
 
@@ -215,7 +245,7 @@ class SDB:  # pylint: disable=too-many-instance-attributes
         """
         self.sdb_filename = sdb_filename
         self.id_prefix = sdb_filename if id_prefix is None else id_prefix
-        self.sdb_file = open(sdb_filename, 'rb', buffering=REVERSE_BUFFER_SIZE if reverse else buffering)
+        self.sdb_file = open_remote(sdb_filename, 'rb', buffering=REVERSE_BUFFER_SIZE if reverse else buffering)
         self.offsets = []
         if self.sdb_file.read(len(MAGIC)) != MAGIC:
             raise RuntimeError('No Sample Database')
@@ -332,6 +362,8 @@ class CSVWriter:  # pylint: disable=too-many-instance-attributes
         labeled : bool or None
             If True: Writes labeled samples (util.sample_collections.LabeledSample) only.
             If False: Ignores transcripts (if available) and writes (unlabeled) util.audio.Sample instances.
+        
+        Currently only works with local files (not gs:// or hdfs://...)
         """
         self.csv_filename = Path(csv_filename)
         self.csv_base_dir = self.csv_filename.parent.resolve().absolute()
@@ -345,7 +377,7 @@ class CSVWriter:  # pylint: disable=too-many-instance-attributes
         self.labeled = labeled
         if labeled:
             fieldnames.append('transcript')
-        self.csv_file = open(csv_filename, 'w', encoding='utf-8', newline='')
+        self.csv_file = open_remote(csv_filename, 'w', encoding='utf-8', newline='')
         self.csv_writer = csv.DictWriter(self.csv_file, fieldnames=fieldnames)
         self.csv_writer.writeheader()
         self.counter = 0
@@ -380,7 +412,7 @@ class CSVWriter:  # pylint: disable=too-many-instance-attributes
 
 
 class TarWriter:  # pylint: disable=too-many-instance-attributes
-    """Sample collection writer for writing a CSV data-set and all its referenced WAV samples to a tar file"""
+    """Sample collection writer for writing a CSV data-set and all its referenced WAV samples to a tar file."""
     def __init__(self,
                  tar_filename,
                  gz=False,
@@ -398,6 +430,8 @@ class TarWriter:  # pylint: disable=too-many-instance-attributes
             If False: Ignores transcripts (if available) and writes (unlabeled) util.audio.Sample instances.
         include : str[]
             List of files to include into tar root.
+
+        Currently only works with local files (not gs:// or hdfs://...)
         """
         self.tar = tarfile.open(tar_filename, 'w:gz' if gz else 'w')
         samples_dir = tarfile.TarInfo('samples')
@@ -498,8 +532,7 @@ class CSV(SampleList):
             If the order of the samples should be reversed
         """
         rows = []
-        csv_dir = Path(csv_filename).parent
+        with open_remote(csv_filename, 'r', encoding='utf8') as csv_file:
-        with open(csv_filename, 'r', encoding='utf8') as csv_file:
             reader = csv.DictReader(csv_file)
             if 'transcript' in reader.fieldnames:
                 if labeled is None:
@@ -508,9 +541,12 @@ class CSV(SampleList):
                 raise RuntimeError('No transcript data (missing CSV column)')
             for row in reader:
                 wav_filename = Path(row['wav_filename'])
-                if not wav_filename.is_absolute():
+                if not wav_filename.is_absolute() and not is_remote_path(row['wav_filename']):
-                    wav_filename = csv_dir / wav_filename
+                    wav_filename = Path(csv_filename).parent / wav_filename
                     wav_filename = str(wav_filename)
+                else:
+                    # Pathlib otherwise removes a / from filenames like hdfs://
+                    wav_filename = row['wav_filename']
                 wav_filesize = int(row['wav_filesize']) if 'wav_filesize' in row else 0
                 if labeled:
                     rows.append((wav_filename, wav_filesize, row['transcript']))
@@ -554,6 +590,11 @@ def samples_from_sources(sample_sources, buffering=BUFFER_SIZE, labeled=None, re
     Loads and combines samples from a list of source files. Sources are combined in an interleaving way to
     keep default sample order from shortest to longest.
 
+    Note that when using distributed training, it is much faster to call this function with single pre-
+    sorted sample source, because this allows for parallelization of the file I/O. (If this function is
+    called with multiple sources, the samples have to be unpacked on a single parent process to allow
+    for reading their durations.)
+
     Parameters
     ----------
     sample_sources : list of str
@@ -570,13 +611,20 @@ def samples_from_sources(sample_sources, buffering=BUFFER_SIZE, labeled=None, re
 
     Returns
     -------
-    iterable of util.sample_collections.LabeledSample (labeled=True) or util.audio.Sample (labeled=False) supporting len
+    iterable of util.sample_collections.PackedSample if a single collection is provided, wrapping
+        LabeledSample (labeled=True) or util.audio.Sample (labeled=False) supporting len
+    or LabeledSample / util.audio.Sample directly, if multiple collections are provided
     """
     sample_sources = list(sample_sources)
     if len(sample_sources) == 0:
         raise ValueError('No files')
     if len(sample_sources) == 1:
         return samples_from_source(sample_sources[0], buffering=buffering, labeled=labeled, reverse=reverse)
-    cols = [samples_from_source(source, buffering=buffering, labeled=labeled, reverse=reverse)
+
+    # If we wish to interleave based on duration, we have to unpack the audio. Note that this unpacking should
+    # be done lazily onn the fly so that it respects the LimitingPool logic used in the feeding code.
+    cols = [LenMap(
+        unpack_maybe, samples_from_source(source, buffering=buffering, labeled=labeled, reverse=reverse))
         for source in sample_sources]
+
     return Interleaved(*cols, key=lambda s: s.duration, reverse=reverse)