Support file-like object in load function

[mthrok]

This PR adds file-like object support to load function.

• From stdin

  import sys
  import torchaudio
  data, sr = torchaudio.load(sys.stdin.buffer)

• From an archive without extracting it in the local filesystem

  import tarfile
  import torchaudio
  
  with tarfile.TarFile(archive_path, 'r') as tarobj:
      data, sample_rate = torchaudio.load(tarobj.extractfile(member_name))

• Streaming over the network
  With requests

  import requests
  import torchaudio
  
  with requests.get(url, stream=True) as resp:
      data, sample_rate = torchaudio.load(resp.raw)

  With boto3

  import boto3
  
  s3 = boto3.client('s3')
  obj = s3.get_object(Bucket=bucket, Key=key)
  data, sample_rate = torchaudio.load(obj['Body'])

For sox_io backend, the new loading function load_audio_fileobj, which works on any Python object with read method that returns bytes. It is added and bound via PyBind11.

For soundfile backend, pysoundfile natively supports file-like object

closes #754 #800
See als #1115

[rbracco]

Thank you, this is an awesome and much needed feature. Your work is appreciated!

[mthrok]

Thank you, this is an awesome and much needed feature. Your work is appreciated!

Hi @rbracco

Thanks :) . Please try it out and give feedback.
Please consider this as prototype feature.
I am observing sporadic test failure in our CI.
I am trying to figure it out what are the conditions that are causing the failure.

https://app.circleci.com/pipelines/github/pytorch/audio/4658/workflows/c810196e-4e45-4d27-8089-9844eb19adf6/jobs/148952

[rbracco]

I have tried it out by sending audio from a React-based frontend to a fastapi backend that receives a file-like object as an upload and then reads it into torchaudio. It worked great with wav files, but when I tried to simplify by moving to mp3 (smaller, more efficient) I started getting RuntimeError: Error loading audio file: failed to open file. in load.

All of the following work:
-Loading a normal MP3 in torchaudio (on the backend server)
-Downloading the MP3 generated by my frontend and opening/playing in an audio player.
-Sending the blob to backend server, writing it to a .mp3 file, and then loading it in torchaudio as a normal file (code below)

For this reason I think it is likely an issue with the new file-like object loading, but unfortunately I have very shallow knowledge of audio formats, C++, and how this works under the hood. I'm happy to keep helping with testing though, please let me know if you have any ideas for a fix, or if this is worthy of a separate issue. Thanks again, I'm really excited for this feature and I'm in an audio ML chat on telegram with many others who feel the same.

with open("audios/example.mp3", "wb") as f:
        f.write(file.file.read())
    y, sr = torchaudio.load("audios/example.mp3")

[mthrok]

Hi @rbracco

Thanks for trying out and the feedback. I am happy to hear that this addition is helpful for you.

For your issue,

• I merged Fix load from file object for small files and shorter bytes #1181 for some tweak. Not sure which nightly you tried but if you tested with yesterday's nightly, can you try again with today's nightly?
• Can you try with format="mp3"? For MP3, libsox does not detect it from the header, so it needs to be told.

If you still see an issue, please open an issue. I am more than happy to look at it.
I also want to polish this feature but I am seeing test failure that happens sporadically (like this) and trying to figure out the cause.