script to create AudioSet numpy files.

Author: cgnorthcutt

dataset-scripts/audioset_preprocessing.py

@@ -0,0 +1,122 @@
+# coding: utf-8
+
+# Copyright (c) 2021-2060 Curtis G. Northcutt
+# This file is part of cgnorthcutt/label-errors.
+#
+# cleanlab is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# cgnorthcutt/label-errors is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+
+# This agreement applies to this version and all previous versions of
+# cgnorthcutt/label-errors.
+
+"""
+Preprocess the tfrecord AudioSet feature embeddings into numpy data files.
+
+Resources used:
+1. https://github.com/tensorflow/models/tree/master/research/audioset
+2. https://research.google.com/audioset/download.html
+3. https://github.com/audioset/ontology
+"""
+
+
+import argparse
+import os
+import numpy as np
+import tensorflow as tf  # version 1.15.4
+import multiprocessing
+import tqdm
+import pickle
+from keras.preprocessing.sequence import pad_sequences
+
+parser = argparse.ArgumentParser(description='PyTorch ImageNet Training')
+parser.add_argument(
+    '--audioset-dir', metavar='AUDIOSET_DIR',
+    help='Specify path to ../audioset/audioset_v1_embeddings/',
+)
+
+
+def read_data(path, include_times=False):
+    result = [[], [], []]
+    if include_times:
+        result += [[], []]
+    for example in tf.python_io.tf_record_iterator(path):
+        tf_example = tf.train.Example.FromString(example)
+        vid_id = tf_example.features.feature['video_id'].bytes_list.value[
+            0].decode(encoding='UTF-8')
+        label = tf_example.features.feature['labels'].int64_list.value
+        if include_times:
+            result[3].append(tf_example.features.feature[
+                                 'start_time_seconds'].float_list.value)
+            result[4].append(tf_example.features.feature[
+                                 'end_time_seconds'].float_list.value)
+        tf_seq_example = tf.train.SequenceExample.FromString(example)
+        tf_feature = tf_seq_example.feature_lists.feature_list[
+            'audio_embedding'].feature
+        n_frames = len(tf_feature)
+        audio_frames = []
+        # Iterate through frames.
+        for i in range(n_frames):
+            hexembed = tf_feature[i].bytes_list.value[0].hex()
+            arrayembed = [int(hexembed[i:i + 2], 16) for i in
+                          range(0, len(hexembed), 2)]
+            audio_frames.append(arrayembed)
+        result[0].append(vid_id)
+        result[1].append(list(label))
+        result[2].append(np.stack(audio_frames).astype(np.uint8))
+    return result
+
+
+def pad(feature_matrix, maxlen=10):
+    return pad_sequences(feature_matrix.T, maxlen=maxlen).T.astype(np.uint8)
+
+
+def preprocess_data(path, prefix='bal_train'):
+    fns = [path + fn for fn in os.listdir(path)]
+    with multiprocessing.Pool(multiprocessing.cpu_count()) as p:
+        results = list(tqdm.tqdm(p.imap(read_data, fns), total=len(fns)))
+
+    print('\nAll files read in. Now post-processing.')
+    video_ids = [v for r in results for v in r[0]]
+    labels = [l for r in results for l in r[1]]
+    features = [f for r in results for f in r[2]]
+    del results  # Free memory
+    # Make all inputs exactly the same shape.
+    print("Padding with 0 to make all features shape (10,128) of type uint8.")
+    with multiprocessing.Pool(multiprocessing.cpu_count()) as p:
+        features = list(tqdm.tqdm(p.imap(pad, features), total=len(features)))
+
+    print('Saving pickled results.')
+    with open(prefix + '_features.p', 'wb') as wf:
+        pickle.dump(features, wf, pickle.HIGHEST_PROTOCOL)
+    with open(prefix + '_video_ids.p', 'wb') as wf:
+        pickle.dump(video_ids, wf, pickle.HIGHEST_PROTOCOL)
+    with open(prefix + '_labels.p', 'wb') as wf:
+        pickle.dump(labels, wf, pickle.HIGHEST_PROTOCOL)
+
+    print('Preprocessing complete.')
+
+
+def main(audioset_dir):
+    for kind in ["eval", "bal_train", "unbal_train"]:
+        preprocess_data(audioset_dir + kind + "/", prefix=kind)
+
+
+if __name__ == '__main__':
+    arg_parser = parser.parse_args()
+    if arg_parser.audioset_dir is None:
+        parser.error("Specify the path to the audioset embeddings "
+                     "directory.\nFor example, if the data is stored in "
+                     "'/datasets/audioset/audioset_v1_embeddings/' "
+                     "you should call this script like this:\npython "
+                     "audioset_preprocessing.py --audioset-dir "
+                     "'/datasets/audioset/audioset_v1_embeddings/''")
+    main(audioset_dir=arg_parser.audioset_dir)