Syoy commited on
Commit
b501913
1 Parent(s): 1d9716c

added embeddings and feature "domain"

Browse files
README.md CHANGED
@@ -70,30 +70,16 @@ pip install renumics-spotlight
70
 
71
  and simply run the following codeblock:
72
  ```jupyterpython
73
- from datasets import load_dataset
74
- from renumics import spotlight
75
-
76
- ds = load_dataset("renumics/dcase23-task2-enriched", "dev")
77
- df = ds.to_pandas()
78
- spotlight.show(df, dtype={'path': spotlight.Audio})
79
- ```
80
-
81
- to load a more advanced view, we can utilize spotlight's layout function and add clear names for the machine_types in the dataset.
82
- ```jupyterpython
83
- import pandas as pd
84
  from datasets import load_dataset, load_dataset_builder
85
  from renumics import spotlight
86
 
87
- train = load_dataset("renumics/dcase23-task2-enriched", "dev", split="train", streaming=False)
88
- test = load_dataset("renumics/dcase23-task2-enriched", "dev", split="test", streaming=False)
89
- db = load_dataset_builder("renumics/dcase23-task2-enriched", "dev")
90
 
91
- df = pd.concat([db.config.to_spotlight(train), db.config.to_spotlight(test)])
92
- spotlight.show(df, dtype={'audio': spotlight.Audio}, layout=db.config.get_layout())
93
  ```
94
 
95
- [//]: # (todo: add embeddings column)
96
-
97
  ## Dataset Structure
98
 
99
  ### Data Instances
@@ -104,9 +90,9 @@ a ClassLabel for the label and a ClassLabel for the class.
104
  ```
105
  {'audio': {'array': array([ 0. , 0.00024414, -0.00024414, ..., -0.00024414,
106
  0. , 0. ], dtype=float32),
107
- 'path': 'train/fan_section_01_source_train_normal_0592_f-n_A.wav',
108
- 'sampling_rate': 16000
109
- }
110
  'path': 'train/fan_section_01_source_train_normal_0592_f-n_A.wav'
111
  'section': 1
112
  'd1p': 'f-n'
@@ -118,6 +104,9 @@ a ClassLabel for the label and a ClassLabel for the class.
118
  'domain': 0 (source)
119
  'label': 0 (normal)
120
  'class': 1 (fan)
 
 
 
121
  }
122
  ```
123
 
@@ -125,16 +114,15 @@ The length of each audio file is 10 seconds.
125
 
126
  ### Data Fields
127
 
128
- [//]: # (todo)
129
-
130
- - `audio`:
131
  - `path`: a string representing the path of the audio file inside the _tar.gz._-archive.
132
- - `section`: an integer
133
- - `d*p`:
134
- - `d*v`:
135
- - `domain`:
136
- - `class`:
137
  - `label`: an integer whose value may be either _0_, indicating that the audio sample is _normal_, _1_, indicating that the audio sample contains an _anomaly_.
 
138
 
139
  ### Data Splits
140
 
@@ -149,6 +137,8 @@ The information for the evaluation dataset will follow after release.
149
 
150
  ## Dataset Creation
151
 
 
 
152
  ### Curation Rationale
153
 
154
  This dataset is the "development dataset" for the [DCASE 2023 Challenge Task 2 "First-Shot Unsupervised Anomalous Sound Detection for Machine Condition Monitoring"](https://dcase.community/challenge2023/task-unsupervised-anomalous-sound-detection-for-machine-condition-monitoring).
@@ -163,7 +153,26 @@ The data consists of the normal/anomalous operating sounds of seven types of rea
163
  - Slide rail
164
  - Valve
165
 
166
- ### Supported Tasks and Leaderboards (original: Overview of the task)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
167
 
168
  Anomalous sound detection (ASD) is the task of identifying whether the sound emitted from a target machine is normal or anomalous. Automatic detection of mechanical failure is an essential technology in the fourth industrial revolution, which involves artificial-intelligence-based factory automation. Prompt detection of machine anomalies by observing sounds is useful for monitoring the condition of machines.
169
 
@@ -185,25 +194,6 @@ For a completely new machine type, hyperparameters of the trained model cannot b
185
 
186
  While sounds from multiple machines of the same machine type can be used to enhance detection performance, it is often the case that sound data from only one machine are available for a machine type. In such a case, the system should be able to train models using only one machine from a machine type.
187
 
188
- ### Source Data
189
-
190
- #### Definition
191
-
192
- We first define key terms in this task: "machine type," "section," "source domain," "target domain," and "attributes.".
193
-
194
- - "Machine type" indicates the type of machine, which in the development dataset is one of seven: fan, gearbox, bearing, slide rail, valve, ToyCar, and ToyTrain.
195
- - A section is defined as a subset of the dataset for calculating performance metrics.
196
- - The source domain is the domain under which most of the training data and some of the test data were recorded, and the target domain is a different set of domains under which some of the training data and some of the test data were recorded. There are differences between the source and target domains in terms of operating speed, machine load, viscosity, heating temperature, type of environmental noise, signal-to-noise ratio, etc.
197
- - Attributes are parameters that define states of machines or types of noise.
198
-
199
- #### Description
200
-
201
- This dataset consists of seven machine types. For each machine type, one section is provided, and the section is a complete set of training and test data. For each section, this dataset provides (i) 990 clips of normal sounds in the source domain for training, (ii) ten clips of normal sounds in the target domain for training, and (iii) 100 clips each of normal and anomalous sounds for the test. The source/target domain of each sample is provided. Additionally, the attributes of each sample in the training and test data are provided in the file names and attribute csv files.
202
-
203
- #### Recording procedure
204
-
205
- Normal/anomalous operating sounds of machines and its related equipment are recorded. Anomalous sounds were collected by deliberately damaging target machines. For simplifying the task, we use only the first channel of multi-channel recordings; all recordings are regarded as single-channel recordings of a fixed microphone. We mixed a target machine sound with environmental noise, and only noisy recordings are provided as training/test data. The environmental noise samples were recorded in several real factory environments. We will publish papers on the dataset to explain the details of the recording procedure by the submission deadline.
206
-
207
  ## Considerations for Using the Data
208
 
209
  ### Social Impact of Dataset
@@ -262,4 +252,4 @@ If you use this dataset, please cite all the following papers. We will publish a
262
  doi = {10.5281/zenodo.7687464},
263
  url = {https://doi.org/10.5281/zenodo.7687464}
264
  }
265
- ```
 
70
 
71
  and simply run the following codeblock:
72
  ```jupyterpython
 
 
 
 
 
 
 
 
 
 
 
73
  from datasets import load_dataset, load_dataset_builder
74
  from renumics import spotlight
75
 
76
+ ds = load_dataset("dcase23-task2-enriched.py", "dev", split="all", streaming=False)
77
+ db = load_dataset_builder("dcase23-task2-enriched.py", "dev")
 
78
 
79
+ df = db.config.to_spotlight(ds)
80
+ spotlight.show(df, dtype={'audio': spotlight.Audio, "ast-finetuned-audioset-10-10-0.4593-embeddings": spotlight.Embedding}, layout=db.config.get_layout())
81
  ```
82
 
 
 
83
  ## Dataset Structure
84
 
85
  ### Data Instances
 
90
  ```
91
  {'audio': {'array': array([ 0. , 0.00024414, -0.00024414, ..., -0.00024414,
92
  0. , 0. ], dtype=float32),
93
+ 'path': 'train/fan_section_01_source_train_normal_0592_f-n_A.wav',
94
+ 'sampling_rate': 16000
95
+ }
96
  'path': 'train/fan_section_01_source_train_normal_0592_f-n_A.wav'
97
  'section': 1
98
  'd1p': 'f-n'
 
104
  'domain': 0 (source)
105
  'label': 0 (normal)
106
  'class': 1 (fan)
107
+ 'ast-finetuned-audioset-10-10-0.4593-embeddings': [[0.8152204155921936,
108
+ 1.5862374305725098, ...,
109
+ 1.7154160737991333]]
110
  }
111
  ```
112
 
 
114
 
115
  ### Data Fields
116
 
117
+ - `audio`: an `datasets.Audio`
 
 
118
  - `path`: a string representing the path of the audio file inside the _tar.gz._-archive.
119
+ - `section`: an integer representing the section, see [Definition](#Description)
120
+ - `d*p`: a string representing the name of the d*-parameter
121
+ - `d*v`: a string representing the value of the corresponding d*-parameter
122
+ - `domain`: an integer whose value may be either _0_, indicating that the audio sample is from the _source_ domain, _1_, indicating that the audio sample is from the _target_.
123
+ - `class`: an integer as class label.
124
  - `label`: an integer whose value may be either _0_, indicating that the audio sample is _normal_, _1_, indicating that the audio sample contains an _anomaly_.
125
+ - `ast-finetuned-audioset-10-10-0.4593-embeddings`: an `datasets.Array2D` representing audio embeddings that are generated with the [Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer#transformers.ASTFeatureExtractor).
126
 
127
  ### Data Splits
128
 
 
137
 
138
  ## Dataset Creation
139
 
140
+ The following information is copied from the original [dataset upload on zenodo.org](https://zenodo.org/record/7690148#.ZAXsSdLMLmE)
141
+
142
  ### Curation Rationale
143
 
144
  This dataset is the "development dataset" for the [DCASE 2023 Challenge Task 2 "First-Shot Unsupervised Anomalous Sound Detection for Machine Condition Monitoring"](https://dcase.community/challenge2023/task-unsupervised-anomalous-sound-detection-for-machine-condition-monitoring).
 
153
  - Slide rail
154
  - Valve
155
 
156
+ ### Source Data
157
+
158
+ #### Definition
159
+
160
+ We first define key terms in this task: "machine type," "section," "source domain," "target domain," and "attributes.".
161
+
162
+ - "Machine type" indicates the type of machine, which in the development dataset is one of seven: fan, gearbox, bearing, slide rail, valve, ToyCar, and ToyTrain.
163
+ - A section is defined as a subset of the dataset for calculating performance metrics.
164
+ - The source domain is the domain under which most of the training data and some of the test data were recorded, and the target domain is a different set of domains under which some of the training data and some of the test data were recorded. There are differences between the source and target domains in terms of operating speed, machine load, viscosity, heating temperature, type of environmental noise, signal-to-noise ratio, etc.
165
+ - Attributes are parameters that define states of machines or types of noise.
166
+
167
+ #### Description
168
+
169
+ This dataset consists of seven machine types. For each machine type, one section is provided, and the section is a complete set of training and test data. For each section, this dataset provides (i) 990 clips of normal sounds in the source domain for training, (ii) ten clips of normal sounds in the target domain for training, and (iii) 100 clips each of normal and anomalous sounds for the test. The source/target domain of each sample is provided. Additionally, the attributes of each sample in the training and test data are provided in the file names and attribute csv files.
170
+
171
+ #### Recording procedure
172
+
173
+ Normal/anomalous operating sounds of machines and its related equipment are recorded. Anomalous sounds were collected by deliberately damaging target machines. For simplifying the task, we use only the first channel of multi-channel recordings; all recordings are regarded as single-channel recordings of a fixed microphone. We mixed a target machine sound with environmental noise, and only noisy recordings are provided as training/test data. The environmental noise samples were recorded in several real factory environments. We will publish papers on the dataset to explain the details of the recording procedure by the submission deadline.
174
+
175
+ ### Supported Tasks and Leaderboards
176
 
177
  Anomalous sound detection (ASD) is the task of identifying whether the sound emitted from a target machine is normal or anomalous. Automatic detection of mechanical failure is an essential technology in the fourth industrial revolution, which involves artificial-intelligence-based factory automation. Prompt detection of machine anomalies by observing sounds is useful for monitoring the condition of machines.
178
 
 
194
 
195
  While sounds from multiple machines of the same machine type can be used to enhance detection performance, it is often the case that sound data from only one machine are available for a machine type. In such a case, the system should be able to train models using only one machine from a machine type.
196
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
197
  ## Considerations for Using the Data
198
 
199
  ### Social Impact of Dataset
 
252
  doi = {10.5281/zenodo.7687464},
253
  url = {https://doi.org/10.5281/zenodo.7687464}
254
  }
255
+ ```
data/dev_metadata.csv CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fd4c496569294420a1b5e4a5b9a8b130b740da5bd9d9d4714746627429b24535
3
- size 825327
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:013c511f2e63467d9c40a73241ae6d235204a8d70a600115a78894f8ce872528
3
+ size 884134
data/dev_test.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c3163c1b0fc19edc4478bd27e4e2065e6a6f69b9a97a6e08a55d0e6b0b56e2d5
3
- size 364523414
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fbe71bfc8cf7d48a6909ec890df02217dcd05361b209a05cf3b36e74386de9ab
3
+ size 364525257
data/dev_train.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:973e811e8911a2d264ec26d7ca0bb0d3f06d2bf1394cabcda4ef902d934b765f
3
- size 1830350107
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91ea5403dae9a1ff6c936fa6b55ee240897cfa591dfba9e099cbde2862b1a05c
3
+ size 1830365732
dcase23-task2-enriched.py CHANGED
@@ -2,6 +2,7 @@ import os
2
  import datasets
3
  import datasets.info
4
  import pandas as pd
 
5
  from pathlib import Path
6
  from datasets import load_dataset
7
  from typing import Iterable, Dict, Optional, Union, List
@@ -49,6 +50,38 @@ DATA_URLS = {
49
  "test": "data/dev_test.tar.gz",
50
  "metadata": "data/dev_metadata.csv",
51
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
52
  }
53
 
54
  STATS = {
@@ -60,11 +93,25 @@ STATS = {
60
  'homepage': "https://zenodo.org/record/7687464#.ZABmANLMLmH",
61
  "splits": ["train", "test"],
62
  },
 
 
 
 
 
 
 
 
 
 
 
 
63
  }
64
  }
65
 
66
  DATASET = {
67
  'dev': 'DCASE 2023 Challenge Task 2 Development Dataset',
 
 
68
  }
69
 
70
  _SPOTLIGHT_LAYOUT = "data/config-spotlight-layout.json"
@@ -82,6 +129,7 @@ class DCASE2023Task2DatasetConfig(datasets.BuilderConfig):
82
  self.release_date = kwargs.pop("release_date", None)
83
  self.homepage = kwargs.pop("homepage", None)
84
  self.data_urls = kwargs.pop("data_urls", None)
 
85
  self.splits = kwargs.pop("splits", None)
86
  self.rename = kwargs.pop("rename", None)
87
  self.layout = kwargs.pop("layout", None)
@@ -98,13 +146,38 @@ class DCASE2023Task2DatasetConfig(datasets.BuilderConfig):
98
  )
99
 
100
  def to_spotlight(self, data: Union[pd.DataFrame, datasets.Dataset]) -> pd.DataFrame:
 
 
 
 
 
 
 
 
 
 
101
  if type(data) == datasets.Dataset:
 
 
 
 
 
 
 
 
 
 
102
  df = data.to_pandas()
103
- df["split"] = data.split
104
  df["config"] = data.config_name
105
 
 
106
  class_names = data.features["class"].names
107
  df["class_name"] = df["class"].apply(lambda x: class_names[x])
 
 
 
 
108
  elif type(data) == pd.DataFrame:
109
  df = data
110
  else:
@@ -123,7 +196,7 @@ class DCASE2023Task2Dataset(datasets.GeneratorBasedBuilder):
123
  """Dataset for the DCASE 2023 Challenge Task 2 "First-Shot Unsupervised Anomalous Sound Detection
124
  for Machine Condition Monitoring"."""
125
 
126
- VERSION = datasets.Version("0.0.2")
127
 
128
  DEFAULT_CONFIG_NAME = "dev"
129
 
@@ -143,8 +216,7 @@ class DCASE2023Task2Dataset(datasets.GeneratorBasedBuilder):
143
  ]
144
 
145
  def _info(self):
146
- features = datasets.Features(
147
- {
148
  "audio": datasets.Audio(sampling_rate=16_000),
149
  "path": datasets.Value("string"),
150
  "section": datasets.Value("int64"),
@@ -154,10 +226,14 @@ class DCASE2023Task2Dataset(datasets.GeneratorBasedBuilder):
154
  "d2v": datasets.Value("string"),
155
  "d3p": datasets.Value("string"),
156
  "d3v": datasets.Value("string"),
 
157
  "label": datasets.ClassLabel(num_classes=_NUM_TARGETS, names=_TARGET_NAMES),
158
  "class": datasets.ClassLabel(num_classes=_NUM_CLASSES, names=_CLASS_NAMES),
159
  }
160
- )
 
 
 
161
 
162
  return datasets.DatasetInfo(
163
  # This is the description that will appear on the datasets page.
@@ -178,11 +254,15 @@ class DCASE2023Task2Dataset(datasets.GeneratorBasedBuilder):
178
  audio_path = {}
179
  local_extracted_archive = {}
180
  split_type = {"train": datasets.Split.TRAIN, "test": datasets.Split.TEST}
 
181
 
182
  for split in split_type:
183
- audio_path[split] = dl_manager.download(self.config.data_urls[split])
184
- local_extracted_archive[split] = dl_manager.extract(
185
- audio_path[split]) if not dl_manager.is_streaming else None
 
 
 
186
 
187
  return [
188
  datasets.SplitGenerator(
@@ -191,9 +271,10 @@ class DCASE2023Task2Dataset(datasets.GeneratorBasedBuilder):
191
  "split": split,
192
  "local_extracted_archive": local_extracted_archive[split],
193
  "audio_files": dl_manager.iter_archive(audio_path[split]),
 
194
  "metadata_file": dl_manager.download_and_extract(self.config.data_urls["metadata"]),
195
  },
196
- ) for split in split_type
197
  ]
198
 
199
  def _generate_examples(
@@ -201,6 +282,7 @@ class DCASE2023Task2Dataset(datasets.GeneratorBasedBuilder):
201
  split: str,
202
  local_extracted_archive: Union[Dict, List],
203
  audio_files: Optional[Iterable],
 
204
  metadata_file: Optional[str],
205
  ):
206
  """Yields examples."""
@@ -215,6 +297,8 @@ class DCASE2023Task2Dataset(datasets.GeneratorBasedBuilder):
215
  audio = {"path": path, "bytes": f.read()}
216
  result = {field: None for field in data_fields}
217
  result.update(metadata[metadata["path"] == lookup].T.squeeze().to_dict())
 
 
218
  result["path"] = path
219
  yield id_, {**result, "audio": audio}
220
  id_ += 1
 
2
  import datasets
3
  import datasets.info
4
  import pandas as pd
5
+ import numpy as np
6
  from pathlib import Path
7
  from datasets import load_dataset
8
  from typing import Iterable, Dict, Optional, Union, List
 
50
  "test": "data/dev_test.tar.gz",
51
  "metadata": "data/dev_metadata.csv",
52
  },
53
+ "add": {
54
+ "train": "data/add_train.tar.gz",
55
+ "test": "data/add_test.tar.gz",
56
+ "metadata": "data/add_metadata.csv",
57
+ },
58
+ "eval": {
59
+ "test": "data/eval_test.tar.gz",
60
+ "metadata": "data/eval_metadata.csv",
61
+ },
62
+ }
63
+
64
+ EMBEDDING_URLS = {
65
+ "dev": {
66
+ "ast-finetuned-audioset-10-10-0.4593-embeddings": {
67
+ "train": "data/MIT_ast-finetuned-audioset-10-10-0.4593-embeddings_dev_train.npz",
68
+ "test": "data/MIT_ast-finetuned-audioset-10-10-0.4593-embeddings_dev_test.npz",
69
+ "size": (1, 768),
70
+ "dtype": "float32",
71
+ },
72
+ },
73
+ "add": {
74
+ "ast-finetuned-audioset-10-10-0.4593-embeddings": {
75
+ "train": "",
76
+ "test": "",
77
+ },
78
+ },
79
+ "eval": {
80
+ "ast-finetuned-audioset-10-10-0.4593-embeddings": {
81
+ "train": "",
82
+ "test": "",
83
+ },
84
+ },
85
  }
86
 
87
  STATS = {
 
93
  'homepage': "https://zenodo.org/record/7687464#.ZABmANLMLmH",
94
  "splits": ["train", "test"],
95
  },
96
+ # 'add': {
97
+ # 'date': None,
98
+ # 'version': "0.0.0",
99
+ # 'homepage': None,
100
+ # "splits": ["train", "test"],
101
+ # },
102
+ # 'eval': {
103
+ # 'date': None,
104
+ # 'version': "0.0.0",
105
+ # 'homepage': None,
106
+ # "splits": ["test"],
107
+ # },
108
  }
109
  }
110
 
111
  DATASET = {
112
  'dev': 'DCASE 2023 Challenge Task 2 Development Dataset',
113
+ 'add': 'DCASE 2023 Challenge Task 2 Additional Train Dataset',
114
+ 'eval': 'DCASE 2023 Challenge Task 2 Evaluation Dataset',
115
  }
116
 
117
  _SPOTLIGHT_LAYOUT = "data/config-spotlight-layout.json"
 
129
  self.release_date = kwargs.pop("release_date", None)
130
  self.homepage = kwargs.pop("homepage", None)
131
  self.data_urls = kwargs.pop("data_urls", None)
132
+ self.embeddings_urls = kwargs.pop("embeddings_urls", None)
133
  self.splits = kwargs.pop("splits", None)
134
  self.rename = kwargs.pop("rename", None)
135
  self.layout = kwargs.pop("layout", None)
 
146
  )
147
 
148
  def to_spotlight(self, data: Union[pd.DataFrame, datasets.Dataset]) -> pd.DataFrame:
149
+
150
+ def get_split(path: str) -> str:
151
+ fn = os.path.basename(path)
152
+ if "train" in fn:
153
+ return "train"
154
+ elif "test" in fn:
155
+ return "test"
156
+ else:
157
+ raise NotImplementedError
158
+
159
  if type(data) == datasets.Dataset:
160
+ # remove embedding columns first -> throws error in .to_pandas()
161
+ embeddings = {}
162
+ emb_features = [key for key, val in data.features.items() if type(val) == datasets.Array2D]
163
+ if len(emb_features) > 0:
164
+ embeddings = {
165
+ key: [np.asarray(emb).reshape(-1,) for emb in data[key].copy()] for key in emb_features
166
+ }
167
+ data = data.remove_columns(emb_features)
168
+
169
+ # retrieve split
170
  df = data.to_pandas()
171
+ df["split"] = data.split._name if "+" not in data.split._name else df["path"].map(get_split)
172
  df["config"] = data.config_name
173
 
174
+ # get clearnames for classes
175
  class_names = data.features["class"].names
176
  df["class_name"] = df["class"].apply(lambda x: class_names[x])
177
+
178
+ # append embeddings
179
+ for emb_name, emb_list in embeddings.items():
180
+ df[emb_name] = emb_list
181
  elif type(data) == pd.DataFrame:
182
  df = data
183
  else:
 
196
  """Dataset for the DCASE 2023 Challenge Task 2 "First-Shot Unsupervised Anomalous Sound Detection
197
  for Machine Condition Monitoring"."""
198
 
199
+ VERSION = datasets.Version("0.0.3")
200
 
201
  DEFAULT_CONFIG_NAME = "dev"
202
 
 
216
  ]
217
 
218
  def _info(self):
219
+ features = {
 
220
  "audio": datasets.Audio(sampling_rate=16_000),
221
  "path": datasets.Value("string"),
222
  "section": datasets.Value("int64"),
 
226
  "d2v": datasets.Value("string"),
227
  "d3p": datasets.Value("string"),
228
  "d3v": datasets.Value("string"),
229
+ "domain": datasets.ClassLabel(num_classes=2, names=["source", "target"]),
230
  "label": datasets.ClassLabel(num_classes=_NUM_TARGETS, names=_TARGET_NAMES),
231
  "class": datasets.ClassLabel(num_classes=_NUM_CLASSES, names=_CLASS_NAMES),
232
  }
233
+ features.update({
234
+ emb_name: datasets.Array2D(shape=emb["size"], dtype=emb["dtype"]) for emb_name, emb in self.config.embeddings_urls.items()
235
+ })
236
+ features = datasets.Features(features)
237
 
238
  return datasets.DatasetInfo(
239
  # This is the description that will appear on the datasets page.
 
254
  audio_path = {}
255
  local_extracted_archive = {}
256
  split_type = {"train": datasets.Split.TRAIN, "test": datasets.Split.TEST}
257
+ embeddings = {split: dict() for split in split_type}
258
 
259
  for split in split_type:
260
+ if split in self.config.splits:
261
+ audio_path[split] = dl_manager.download(self.config.data_urls[split])
262
+ local_extracted_archive[split] = dl_manager.extract(
263
+ audio_path[split]) if not dl_manager.is_streaming else None
264
+ for emb_name, emb_data in self.config.embeddings_urls.items():
265
+ embeddings[split][emb_name] = np.load(dl_manager.download_and_extract(emb_data[split]) + "/arr_0.npy", allow_pickle=True).item()
266
 
267
  return [
268
  datasets.SplitGenerator(
 
271
  "split": split,
272
  "local_extracted_archive": local_extracted_archive[split],
273
  "audio_files": dl_manager.iter_archive(audio_path[split]),
274
+ "embeddings": embeddings[split],
275
  "metadata_file": dl_manager.download_and_extract(self.config.data_urls["metadata"]),
276
  },
277
+ ) for split in split_type if split in self.config.splits
278
  ]
279
 
280
  def _generate_examples(
 
282
  split: str,
283
  local_extracted_archive: Union[Dict, List],
284
  audio_files: Optional[Iterable],
285
+ embeddings: Optional[Dict],
286
  metadata_file: Optional[str],
287
  ):
288
  """Yields examples."""
 
297
  audio = {"path": path, "bytes": f.read()}
298
  result = {field: None for field in data_fields}
299
  result.update(metadata[metadata["path"] == lookup].T.squeeze().to_dict())
300
+ for emb_key in embeddings.keys():
301
+ result[emb_key] = embeddings[emb_key][lookup]
302
  result["path"] = path
303
  yield id_, {**result, "audio": audio}
304
  id_ += 1