Datasets:

Languages:
English
Multilinguality:
monolingual
Size Categories:
1T<n
Source Datasets:
original
ArXiv:
License:
polinaeterna HF staff commited on
Commit
3333d98
1 Parent(s): b1faecb

add some comments and todos

Browse files
Files changed (1) hide show
  1. peoples_speech.py +18 -29
peoples_speech.py CHANGED
@@ -61,35 +61,8 @@ _LICENSE = [
61
  "cc-by-sa-3.0", "cc-by-sa-4.0"
62
  ]
63
 
64
- # TODO: Add link to the official dataset URLs here
65
- # The HuggingFace Datasets library doesn't host the datasets but only points to the original files.
66
- # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
67
- _URLS = {
68
- "clean-cc-by": {
69
- "audio_tar": "",
70
- "manifest": "",
71
- },
72
- "dirty-cc-by": {
73
- "audio_tar": "",
74
- "manifest": "",
75
- },
76
- "clean-cc-by-sa": {
77
- "audio_tar": "",
78
- "manifest": "",
79
- },
80
- "dirty-cc-by-sa": {
81
- "audio_tar": "",
82
- "manifest": "",
83
- },
84
- "microset": {
85
- "audio_tar": "",
86
- "manifest": "",
87
- },
88
- }
89
-
90
- # _BASE_URL = "https://huggingface.co/datasets/MLCommons/peoples_speech/resolve/main/"
91
-
92
  # relative path to data inside dataset's repo
 
93
  _DATA_URL = "{config}/{config}_00000{archive_id}.tar"
94
 
95
  # relative path to metadata inside dataset's repo
@@ -101,6 +74,7 @@ class PeoplesSpeech(datasets.GeneratorBasedBuilder):
101
 
102
  VERSION = datasets.Version("1.1.0")
103
  BUILDER_CONFIGS = [
 
104
  datasets.BuilderConfig(name="clean", version=VERSION, description="Clean, CC-BY licensed subset."),
105
  datasets.BuilderConfig(name="dirty", version=VERSION, description="Dirty, CC-BY licensed subset."),
106
  datasets.BuilderConfig(name="clean_sa", version=VERSION, description="Clean, CC-BY-SA licensed subset."),
@@ -132,17 +106,30 @@ class PeoplesSpeech(datasets.GeneratorBasedBuilder):
132
  # TODO: this should be changed to the actual number of archives further
133
  urls = [_DATA_URL.format(config=self.config.name, archive_id=i) for i in range(5)]
134
  archive_paths = [dl_manager.download(url) for url in urls]
 
 
135
  local_extracted_archive_paths = [dl_manager.extract(path) for path in archive_paths] \
136
  if not dl_manager.is_streaming else [None] * len(archive_paths)
137
 
138
  manifest_url = _MANIFEST_URL.format(config=self.config.name)
139
  manifest_path = dl_manager.download_and_extract(manifest_url)
140
 
 
 
 
 
 
 
 
 
 
 
141
  return [
142
  datasets.SplitGenerator(
143
  name=datasets.Split.TRAIN,
144
  gen_kwargs={
145
  "local_extracted_archive_paths": local_extracted_archive_paths,
 
146
  "archives": [dl_manager.iter_archive(path) for path in archive_paths],
147
  "manifest_path": manifest_path
148
  },
@@ -165,9 +152,11 @@ class PeoplesSpeech(datasets.GeneratorBasedBuilder):
165
  "duration_ms": duration
166
  }
167
 
168
- print("generating examples")
169
  for local_extracted_archive_path, archive in zip(local_extracted_archive_paths, archives):
 
170
  for audio_filename, audio_file in archive:
 
 
171
  path = os.path.join(local_extracted_archive_path, audio_filename) if local_extracted_archive_path \
172
  else audio_filename
173
  yield audio_filename, {
61
  "cc-by-sa-3.0", "cc-by-sa-4.0"
62
  ]
63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64
  # relative path to data inside dataset's repo
65
+ # TODO: change according to the scheme of generating urls to the audio archives
66
  _DATA_URL = "{config}/{config}_00000{archive_id}.tar"
67
 
68
  # relative path to metadata inside dataset's repo
74
 
75
  VERSION = datasets.Version("1.1.0")
76
  BUILDER_CONFIGS = [
77
+ # TODO: add "subset" config
78
  datasets.BuilderConfig(name="clean", version=VERSION, description="Clean, CC-BY licensed subset."),
79
  datasets.BuilderConfig(name="dirty", version=VERSION, description="Dirty, CC-BY licensed subset."),
80
  datasets.BuilderConfig(name="clean_sa", version=VERSION, description="Clean, CC-BY-SA licensed subset."),
106
  # TODO: this should be changed to the actual number of archives further
107
  urls = [_DATA_URL.format(config=self.config.name, archive_id=i) for i in range(5)]
108
  archive_paths = [dl_manager.download(url) for url in urls]
109
+
110
+ # In non-streaming mode, we extract the archives to have the data locally:
111
  local_extracted_archive_paths = [dl_manager.extract(path) for path in archive_paths] \
112
  if not dl_manager.is_streaming else [None] * len(archive_paths)
113
 
114
  manifest_url = _MANIFEST_URL.format(config=self.config.name)
115
  manifest_path = dl_manager.download_and_extract(manifest_url)
116
 
117
+ # To access the audio data from the TAR archives using the download manager,
118
+ # we have to use the dl_manager.iter_archive method
119
+ #
120
+ # This is because dl_manager.download_and_extract
121
+ # doesn't work to stream TAR archives in streaming mode.
122
+ # (we have to stream the files of a TAR archive one by one)
123
+ #
124
+ # The iter_archive method returns an iterable of (path_within_archive, file_obj) for every
125
+ # file in a TAR archive.
126
+
127
  return [
128
  datasets.SplitGenerator(
129
  name=datasets.Split.TRAIN,
130
  gen_kwargs={
131
  "local_extracted_archive_paths": local_extracted_archive_paths,
132
+ # use iter_archive here to access the files in the TAR archives:
133
  "archives": [dl_manager.iter_archive(path) for path in archive_paths],
134
  "manifest_path": manifest_path
135
  },
152
  "duration_ms": duration
153
  }
154
 
 
155
  for local_extracted_archive_path, archive in zip(local_extracted_archive_paths, archives):
156
+ # Here we iterate over all the files within the TAR archive:
157
  for audio_filename, audio_file in archive:
158
+ # if an audio file exists locally (i.e. in default, non-streaming mode) set the full path to it
159
+ # joining path to directory that the archive was extracted to and audio filename.
160
  path = os.path.join(local_extracted_archive_path, audio_filename) if local_extracted_archive_path \
161
  else audio_filename
162
  yield audio_filename, {