lhoestq HF staff commited on
Commit
c44b871
1 Parent(s): 69ce3af

Simplify Common Voice code (#3817)

Browse files

* simplify common_voice

* one iter_archive per split

Commit from https://github.com/huggingface/datasets/commit/f5929980e7d857d491a20c35305ea01a24c25326

Files changed (1) hide show
  1. common_voice.py +59 -90
common_voice.py CHANGED
@@ -659,135 +659,100 @@ class CommonVoice(datasets.GeneratorBasedBuilder):
659
 
660
  def _split_generators(self, dl_manager):
661
  """Returns SplitGenerators."""
662
- streaming = dl_manager.is_streaming
663
  archive_path = dl_manager.download(_DATA_URL.format(self.config.name))
664
- if streaming:
665
- # Here we use iter_archive in streaming mode because dl_manager.download_and_extract
666
- # doesn't work to stream TAR archives (we have to stream the files in the archive one by one).
667
- #
668
- # The iter_archive method returns an iterable of (path_within_archive, file_obj) for every
669
- # file in the TAR archive.
670
- #
671
- archive_iterator = dl_manager.iter_archive(archive_path)
672
- # we locate the data using the path within the archive
673
- path_to_data = "/".join(["cv-corpus-6.1-2020-12-11", self.config.name])
674
- path_to_clips = "/".join([path_to_data, "clips"])
675
- metadata_filepaths = {
676
- split: "/".join([path_to_data, f"{split}.tsv"])
677
- for split in ["train", "test", "dev", "other", "validated", "invalidated"]
678
- }
679
- else:
680
- # In non-streaming we can extract the archive locally as usual
681
- extracted_dir = dl_manager.extract(archive_path)
682
- archive_iterator = None
683
- # we locate the data using the local path
684
- path_to_data = os.path.join(extracted_dir, "cv-corpus-6.1-2020-12-11", self.config.name)
685
- path_to_clips = os.path.join(path_to_data, "clips")
686
- metadata_filepaths = {
687
- split: os.path.join(path_to_data, f"{split}.tsv")
688
- for split in ["train", "test", "dev", "other", "validated", "invalidated"]
689
- }
690
 
691
  return [
692
  datasets.SplitGenerator(
693
  name=datasets.Split.TRAIN,
694
  gen_kwargs={
695
- "streaming": streaming,
696
- "archive_iterator": archive_iterator,
697
- "filepath": metadata_filepaths["train"],
 
 
698
  "path_to_clips": path_to_clips,
699
  },
700
  ),
701
  datasets.SplitGenerator(
702
  name=datasets.Split.TEST,
703
  gen_kwargs={
704
- "streaming": streaming,
705
- "archive_iterator": archive_iterator,
706
- "filepath": metadata_filepaths["test"],
 
 
707
  "path_to_clips": path_to_clips,
708
  },
709
  ),
710
  datasets.SplitGenerator(
711
  name=datasets.Split.VALIDATION,
712
  gen_kwargs={
713
- "streaming": streaming,
714
- "archive_iterator": archive_iterator,
715
- "filepath": metadata_filepaths["dev"],
 
 
716
  "path_to_clips": path_to_clips,
717
  },
718
  ),
719
  datasets.SplitGenerator(
720
  name="other",
721
  gen_kwargs={
722
- "streaming": streaming,
723
- "archive_iterator": archive_iterator,
724
- "filepath": metadata_filepaths["other"],
 
 
725
  "path_to_clips": path_to_clips,
726
  },
727
  ),
728
  datasets.SplitGenerator(
729
  name="validated",
730
  gen_kwargs={
731
- "streaming": streaming,
732
- "archive_iterator": archive_iterator,
733
- "filepath": metadata_filepaths["validated"],
 
 
734
  "path_to_clips": path_to_clips,
735
  },
736
  ),
737
  datasets.SplitGenerator(
738
  name="invalidated",
739
  gen_kwargs={
740
- "streaming": streaming,
741
- "archive_iterator": archive_iterator,
742
- "filepath": metadata_filepaths["invalidated"],
 
 
743
  "path_to_clips": path_to_clips,
744
  },
745
  ),
746
  ]
747
 
748
- def _generate_examples(self, streaming, archive_iterator, filepath, path_to_clips):
749
  """Yields examples."""
750
- if streaming:
751
- yield from self._generate_examples_streaming(archive_iterator, filepath, path_to_clips)
752
- else:
753
- yield from self._generate_examples_non_streaming(filepath, path_to_clips)
754
-
755
- def _generate_examples_non_streaming(self, filepath, path_to_clips):
756
-
757
- data_fields = list(self._info().features.keys())
758
-
759
- # audio is not a header of the csv files
760
- data_fields.remove("audio")
761
- path_idx = data_fields.index("path")
762
-
763
- with open(filepath, encoding="utf-8") as f:
764
- lines = f.readlines()
765
- headline = lines[0]
766
-
767
- column_names = headline.strip().split("\t")
768
- assert (
769
- column_names == data_fields
770
- ), f"The file should have {data_fields} as column names, but has {column_names}"
771
-
772
- for id_, line in enumerate(lines[1:]):
773
- field_values = line.strip().split("\t")
774
-
775
- # set absolute path for mp3 audio file
776
- field_values[path_idx] = os.path.join(path_to_clips, field_values[path_idx])
777
-
778
- # if data is incomplete, fill with empty values
779
- if len(field_values) < len(data_fields):
780
- field_values += (len(data_fields) - len(field_values)) * ["''"]
781
-
782
- result = {key: value for key, value in zip(data_fields, field_values)}
783
-
784
- # set audio feature
785
- result["audio"] = field_values[path_idx]
786
-
787
- yield id_, result
788
-
789
- def _generate_examples_streaming(self, archive_iterator, filepath, path_to_clips):
790
- """Yields examples in streaming mode."""
791
  data_fields = list(self._info().features.keys())
792
 
793
  # audio is not a header of the csv files
@@ -796,8 +761,10 @@ class CommonVoice(datasets.GeneratorBasedBuilder):
796
 
797
  all_field_values = {}
798
  metadata_found = False
 
799
  for path, f in archive_iterator:
800
- if path == filepath:
 
801
  metadata_found = True
802
  lines = f.readlines()
803
  headline = lines[0].decode("utf-8")
@@ -811,11 +778,13 @@ class CommonVoice(datasets.GeneratorBasedBuilder):
811
  # set full path for mp3 audio file
812
  audio_path = "/".join([path_to_clips, field_values[path_idx]])
813
  all_field_values[audio_path] = field_values
 
814
  elif path.startswith(path_to_clips):
815
  assert metadata_found, "Found audio clips before the metadata TSV file."
816
  if not all_field_values:
817
  break
818
  if path in all_field_values:
 
819
  field_values = all_field_values[path]
820
 
821
  # if data is incomplete, fill with empty values
@@ -826,7 +795,7 @@ class CommonVoice(datasets.GeneratorBasedBuilder):
826
 
827
  # set audio feature
828
  result["audio"] = {"path": path, "bytes": f.read()}
829
- # set path to None since the path doesn't exist locally in streaming mode
830
- result["path"] = None
831
 
832
  yield path, result
659
 
660
  def _split_generators(self, dl_manager):
661
  """Returns SplitGenerators."""
662
+ # Download the TAR archive that contains the audio files:
663
  archive_path = dl_manager.download(_DATA_URL.format(self.config.name))
664
+
665
+ # First we locate the data using the path within the archive:
666
+ path_to_data = "/".join(["cv-corpus-6.1-2020-12-11", self.config.name])
667
+ path_to_clips = "/".join([path_to_data, "clips"])
668
+ metadata_filepaths = {
669
+ split: "/".join([path_to_data, f"{split}.tsv"])
670
+ for split in ["train", "test", "dev", "other", "validated", "invalidated"]
671
+ }
672
+ # (Optional) In non-streaming mode, we can extract the archive locally to have actual local audio files:
673
+ local_extracted_archive = dl_manager.extract(archive_path) if not dl_manager.is_streaming else None
674
+
675
+ # To access the audio data from the TAR archives using the download manager,
676
+ # we have to use the dl_manager.iter_archive method.
677
+ #
678
+ # This is because dl_manager.download_and_extract
679
+ # doesn't work to stream TAR archives in streaming mode.
680
+ # (we have to stream the files of a TAR archive one by one)
681
+ #
682
+ # The iter_archive method returns an iterable of (path_within_archive, file_obj) for every
683
+ # file in the TAR archive.
 
 
 
 
 
 
684
 
685
  return [
686
  datasets.SplitGenerator(
687
  name=datasets.Split.TRAIN,
688
  gen_kwargs={
689
+ "local_extracted_archive": local_extracted_archive,
690
+ "archive_iterator": dl_manager.iter_archive(
691
+ archive_path
692
+ ), # use iter_archive here to access the files in the TAR archives
693
+ "metadata_filepath": metadata_filepaths["train"],
694
  "path_to_clips": path_to_clips,
695
  },
696
  ),
697
  datasets.SplitGenerator(
698
  name=datasets.Split.TEST,
699
  gen_kwargs={
700
+ "local_extracted_archive": local_extracted_archive,
701
+ "archive_iterator": dl_manager.iter_archive(
702
+ archive_path
703
+ ), # use iter_archive here to access the files in the TAR archives
704
+ "metadata_filepath": metadata_filepaths["test"],
705
  "path_to_clips": path_to_clips,
706
  },
707
  ),
708
  datasets.SplitGenerator(
709
  name=datasets.Split.VALIDATION,
710
  gen_kwargs={
711
+ "local_extracted_archive": local_extracted_archive,
712
+ "archive_iterator": dl_manager.iter_archive(
713
+ archive_path
714
+ ), # use iter_archive here to access the files in the TAR archives
715
+ "metadata_filepath": metadata_filepaths["dev"],
716
  "path_to_clips": path_to_clips,
717
  },
718
  ),
719
  datasets.SplitGenerator(
720
  name="other",
721
  gen_kwargs={
722
+ "local_extracted_archive": local_extracted_archive,
723
+ "archive_iterator": dl_manager.iter_archive(
724
+ archive_path
725
+ ), # use iter_archive here to access the files in the TAR archives
726
+ "metadata_filepath": metadata_filepaths["other"],
727
  "path_to_clips": path_to_clips,
728
  },
729
  ),
730
  datasets.SplitGenerator(
731
  name="validated",
732
  gen_kwargs={
733
+ "local_extracted_archive": local_extracted_archive,
734
+ "archive_iterator": dl_manager.iter_archive(
735
+ archive_path
736
+ ), # use iter_archive here to access the files in the TAR archives
737
+ "metadata_filepath": metadata_filepaths["validated"],
738
  "path_to_clips": path_to_clips,
739
  },
740
  ),
741
  datasets.SplitGenerator(
742
  name="invalidated",
743
  gen_kwargs={
744
+ "local_extracted_archive": local_extracted_archive,
745
+ "archive_iterator": dl_manager.iter_archive(
746
+ archive_path
747
+ ), # use iter_archive here to access the files in the TAR archives
748
+ "metadata_filepath": metadata_filepaths["invalidated"],
749
  "path_to_clips": path_to_clips,
750
  },
751
  ),
752
  ]
753
 
754
+ def _generate_examples(self, local_extracted_archive, archive_iterator, metadata_filepath, path_to_clips):
755
  """Yields examples."""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
756
  data_fields = list(self._info().features.keys())
757
 
758
  # audio is not a header of the csv files
761
 
762
  all_field_values = {}
763
  metadata_found = False
764
+ # Here we iterate over all the files within the TAR archive:
765
  for path, f in archive_iterator:
766
+ # Parse the metadata CSV file
767
+ if path == metadata_filepath:
768
  metadata_found = True
769
  lines = f.readlines()
770
  headline = lines[0].decode("utf-8")
778
  # set full path for mp3 audio file
779
  audio_path = "/".join([path_to_clips, field_values[path_idx]])
780
  all_field_values[audio_path] = field_values
781
+ # Else, read the audio file and yield an example
782
  elif path.startswith(path_to_clips):
783
  assert metadata_found, "Found audio clips before the metadata TSV file."
784
  if not all_field_values:
785
  break
786
  if path in all_field_values:
787
+ # retrieve the metadata corresponding to this audio file
788
  field_values = all_field_values[path]
789
 
790
  # if data is incomplete, fill with empty values
795
 
796
  # set audio feature
797
  result["audio"] = {"path": path, "bytes": f.read()}
798
+ # set path to None if the audio file doesn't exist locally (i.e. in streaming mode)
799
+ result["path"] = os.path.join(local_extracted_archive, path) if local_extracted_archive else None
800
 
801
  yield path, result