问题
I am trying to use the 128 byte embeddings produced by the pre-trained base of the VGGish model for transfer learning on audio data. Using python vggish_inference_demo.py --wav_file ... to encode my training data to a tfrecord worked fine, but now I want to use this as an input to another model (e.g. a neural network I create with keras or something else). Using some similar questions and the documentation, I go this far with the first embedding record of one file:
tfrecords_filename = 'example1.tfrecord'
record_iterator = tf.python_io.tf_record_iterator(path=tfrecords_filename)
string_record = next(record_iterator)
example = tf.train.SequenceExample()
example.ParseFromString(string_record)
print(example.feature_lists.feature_list['audio_embedding'].feature[0].bytes_list.value)
This produces
[b'\x99\x07\xaa>\xd2_R_\x9f\xbbqN\x99\xa18V\xad\x7f\x93\xf0)\xdd4\x80~\xb0\xa4d\x8e\x85\xb6\x88\xa3?U\xa6Q[\x9b\x038\xff\x00EE>OJ\xa5\xb8\x828)\x97^\x8a\xaa\x12h\xff\xff\xc39\xce\x9b\x13\x80\x00j\xcaZ\xac\xff\xff\x0f\xac\x1c\x90&\xd2.b\xe2{\xc1\x15\xe9\xba\xed\xd4\xa9\xff\xdc\xb5\x99]!\x04\xca\xff\xa6;b\xe0\x19\xbfW\xebP!\xff\xc5\xff\x82\xff\x1a\xbe\xec-h\xff\x8d\xff\r\x96\x00\x00\xff']
I am not even sure what this b'...' is (there's more than 64 and fewer than 128 xs - so not sure how this lines up with anything).
Maybe I am missing some basic Python knowledge here, but how do I turn this into a numeric array of numbers that I can use as an input to some other model?
回答1:
It turns out that these are bytes that can be converted to hex, which can be converted to an array of integers between 0 to 255.
import tensorflow as tf
import numpy as np
tfrecords_filename = 'example1.tfrecord'
record_iterator = tf.python_io.tf_record_iterator(path=tfrecords_filename)
string_record = next(record_iterator)
example = tf.train.SequenceExample()
example.ParseFromString(string_record)
hexembed = example.feature_lists.feature_list['audio_embedding'].feature[0].bytes_list.value[0].hex()
arrayembed = [int(hexembed[i:i+2],16) for i in range(0,len(hexembed),2)]
print(arrayembed)
This produces output in the format that I desired:
[153, 7, 170, 62, 210, 95, 82, 95, 159, 187, 113, 78, 153, 161, 56, 86, 173, 127, 147, 240, 41, 221, 52, 128, 126, 176, 164, 100, 142, 133, 182, 136, 163, 63, 85, 166, 81, 91, 155, 3, 56, 255, 0, 69, 69, 62, 79, 74, 165, 184, 130, 56, 41, 151, 94, 138, 170, 18, 104, 255, 255, 195, 57, 206, 155, 19, 128, 0, 106, 202, 90, 172, 255, 255, 15, 172, 28, 144, 38, 210, 46, 98, 226, 123, 193, 21, 233, 186, 237, 212, 169, 255, 220, 181, 153, 93, 33, 4, 202, 255, 166, 59, 98, 224, 25, 191, 87, 235, 80, 33, 255, 197, 255, 130, 255, 26, 190, 236, 45, 104, 255, 141, 255, 13, 150, 0, 0, 255]
来源：https://stackoverflow.com/questions/53340036/how-to-decode-vggish-audioset-embeddings-from-tfrecord