Add unbalanced config. Fix memory leak.

AudioSet.py

@@ -51,6 +51,10 @@ _LICENSE = "cc-by-4.0"
 
 _URL_PREFIX = "https://huggingface.co/datasets/agkphysics/AudioSet/resolve/main"
 
+_N_BAL_TRAIN_TARS = 10
+_N_UNBAL_TRAIN_TARS = 870
+_N_EVAL_TARS = 9
+
 
 def _iter_tar(path):
     """Iterate through the tar archive, but without skipping some files, which the HF
@@ -68,6 +72,20 @@ def _iter_tar(path):
 class AudioSetDataset(datasets.GeneratorBasedBuilder):
     VERSION = datasets.Version("1.0.0")
 
+    BUILDER_CONFIGS = [
+        datasets.BuilderConfig(
+            name="balanced",
+            version=VERSION,
+            description="Balanced training and balanced evaluation set.",
+        ),
+        datasets.BuilderConfig(
+            name="unbalanced",
+            version=VERSION,
+            description="Full unbalanced training set and balanced evaluation set.",
+        ),
+    ]
+    DEFAULT_CONFIG_NAME = "balanced"
+
     def _info(self) -> datasets.DatasetInfo:
         return datasets.DatasetInfo(
             description=_DESCRIPTION,
@@ -92,14 +110,24 @@ class AudioSetDataset(datasets.GeneratorBasedBuilder):
         prefix = prefix + "/data"
 
         _LABEL_URLS = {
-            "bal_train": f"{prefix}/balanced_train_segments.csv",
+            "bal_train": (
+                f"{prefix}/balanced_train_segments.csv"
+                if self.config.name == "balanced"
+                else f"{prefix}/unbalanced_train_segments.csv"
+            ),
             "eval": f"{prefix}/eval_segments.csv",
             "ontology": f"{prefix}/ontology.json",
         }
-
         _DATA_URLS = {
-            "bal_train": [f"{prefix}/bal_train0{i}.tar" for i in range(10)],
-            "eval": [f"{prefix}/eval0{i}.tar" for i in range(9)],
+            "bal_train": (
+                [f"{prefix}/bal_train0{i}.tar" for i in range(_N_BAL_TRAIN_TARS)]
+                if self.config.name == "balanced"
+                else [
+                    f"{prefix}/unbal_train{i:03d}.tar"
+                    for i in range(_N_UNBAL_TRAIN_TARS)
+                ]
+            ),
+            "eval": [f"{prefix}/eval0{i}.tar" for i in range(_N_EVAL_TARS)],
         }
 
         tar_files = dl_manager.download(_DATA_URLS)
@@ -129,30 +157,27 @@ class AudioSetDataset(datasets.GeneratorBasedBuilder):
         ]
 
     def _generate_examples(self, labels, ontology, audio_files):
+        with open(ontology) as fid:
+            ontology_data = json.load(fid)
+        id_to_name = {x["id"]: x["name"] for x in ontology_data}
+
         labels_df = pd.read_csv(
             labels,
             skiprows=3,
             header=None,
             skipinitialspace=True,
             names=["vid_id", "start", "end", "labels"],
+            index_col="vid_id",
         )
-        with open(ontology) as fid:
-            ontology_data = json.load(fid)
-        id_to_name = {x["id"]: x["name"] for x in ontology_data}
 
-        examples = {}
-        for _, row in labels_df.iterrows():
-            label_ids = row["labels"].split(",")
+        for path, fid in audio_files:
+            vid_id = os.path.splitext(os.path.basename(path))[0]
+            label_ids = labels_df.loc[vid_id, "labels"].split(",")
             human_labels = [id_to_name[x] for x in label_ids]
-            examples[row["vid_id"]] = {
-                "video_id": row["vid_id"],
+            example = {
+                "video_id": vid_id,
                 "labels": label_ids,
                 "human_labels": human_labels,
+                "audio": {"path": path, "bytes": fid.read()},
             }
-
-        for path, fid in audio_files:
-            vid_id = os.path.splitext(os.path.basename(path))[0]
-            if vid_id in examples:
-                audio = {"path": path, "bytes": fid.read()}
-                examples[vid_id]["audio"] = audio
-                yield vid_id, examples[vid_id]
+            yield vid_id, example

README.md

@@ -1,13 +1,23 @@
 ---
+language:
+- en
 license: cc-by-4.0
+size_categories:
+- 10K<n<100K
+- 1M<n<10M
+source_datasets:
+- original
 task_categories:
 - audio-classification
+paperswithcode_id: audioset
+pretty_name: AudioSet
+config_names:
+- balanced
+- unbalanced
 tags:
 - audio
-size_categories:
-- 10K<n<100K
-paperswithcode_id: audioset
 dataset_info:
+- config_name: balanced
   features:
   - name: video_id
     dtype: string
@@ -26,29 +36,52 @@ dataset_info:
     num_examples: 17142
   download_size: 49805654900
   dataset_size: 49779893265
+- config_name: unbalanced
+  features:
+  - name: video_id
+    dtype: string
+  - name: audio
+    dtype: audio
+  - name: labels
+    sequence: string
+  - name: human_labels
+    sequence: string
+  splits:
+  - name: train
+    num_bytes: 2408656417541
+    num_examples: 1738788
+  - name: test
+    num_bytes: 23763682278
+    num_examples: 17142
+  download_size: 2433673104977
+  dataset_size: 2432420099819
 ---
 
 # Dataset Card for AudioSet
 
-## Dataset Details
-This repository contains the balanced training set and evaluation set of
-the [AudioSet data](
-https://research.google.com/audioset/dataset/index.html). The YouTube
-videos were downloaded in March 2023, so not all of the original audios
-are available.
-
-The distribuion of audio clips is as follows. In parentheses is the dict
-key used for HugginFace `datasets`:
-- `bal_train` (`train`): 18685 audio clips out of 22160 originally.
-- `eval` (`test`): 17142 audio clips out of 20371 originally.
-
-You can use the `datasets` library to load this dataset, in which case
-the raw audio will be returned along with a sequence of one or more
-labels. Note that the raw audio is returned without further processing,
-so you will need to decode and possibly downsample the audio for model
-training.
-
-Example instance from the `train` subset:
+## Dataset Description
+- **Homepage**: https://research.google.com/audioset/index.html
+- **Paper**: https://storage.googleapis.com/gweb-research2023-media/pubtools/pdf/45857.pdf
+- **Leaderboard**: https://paperswithcode.com/sota/audio-classification-on-audioset
+
+### Dataset Summary
+[AudioSet](https://research.google.com/audioset/dataset/index.html) is a
+dataset of 10-second clips from YouTube, annotated into one or more
+sound categories, following the AudioSet ontology.
+
+### Supported Tasks and Leaderboards
+- `audio-classification`: Classify audio clips into categories. The
+  leaderboard is available
+  [here](https://paperswithcode.com/sota/audio-classification-on-audioset)
+
+### Languages
+The class labels in the dataset are in English.
+
+
+## Dataset Structure
+
+### Data Instances
+Example instance from the dataset:
 ```python
 {
  'video_id': '--PJHxphWEs',
@@ -63,8 +96,78 @@ Example instance from the `train` subset:
 }
 ```
 
-Most audio is sampled at 48 kHz 24 bit, but about 10% is sampled at
-44.1 kHz 24 bit. Audio files are stored in the FLAC format.
+### Data Fields
+Instances have the following fields:
+- `video_id`: a `string` feature containing the original YouTube ID.
+- `audio`: an `Audio` feature containing the audio data and sample rate.
+- `labels`: a sequence of `string` features containing the labels
+  associated with the audio clip.
+- `human_labels`: a sequence of `string` features containing the
+  human-readable forms of the same labels as in `labels`.
+
+### Data Splits
+The distribuion of audio clips is as follows:
+
+#### `balanced` configuration
+|           |train|test |
+|-----------|----:|----:|
+|# instances|18685|17142|
+
+#### `unbalanced` configuration
+|           |train  |test |
+|-----------|------:|----:|
+|# instances|1738788|17142|
+
+
+## Dataset Creation
+
+### Curation Rationale
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+### Source Data
+
+#### Initial Data Collection and Normalization
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+#### Who are the source language producers?
+The labels are from the AudioSet ontology. Audio clips are from YouTube.
+
+### Annotations
+
+#### Annotation process
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+#### Who are the annotators?
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+### Personal and Sensitive Information
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+## Considerations for Using the Data
+
+### Social Impact of Dataset
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+### Discussion of Biases
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+### Other Known Limitations
+1. The YouTube videos in this copy of AudioSet were downloaded in March
+   2023, so not all of the original audios are available. The number of
+   clips able to be downloaded is as follows:
+   - Balanced train: 18685 audio clips out of 22160 originally.
+   - Unbalanced train: 1738788 clips out of 2041789 originally.
+   - Evaluation: 17142 audio clips out of 20371 originally.
+2. Most audio is sampled at 48 kHz 24 bit, but about 10% is sampled at
+   44.1 kHz 24 bit. Audio files are stored in the FLAC format.
+
+## Additional Information
+
+### Dataset Curators
+[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+
+### Licensing Information
+The AudioSet data is licensed under CC-BY-4.0
 
 ## Citation
 ```bibtex