Media Stream API - Pt. 1

This tutorial shows how to use torchaudio’s I/O stream API to fetch and decode audio/video data and apply preprocessings that libavfilter provides.

Note

This tutorial requires Streaming API and FFmpeg libraries (>=4.1, <5).

The Streaming API is available in nightly builds. Please refer to https://pytorch.org/get-started/locally/ for instructions.

There are multiple ways to install FFmpeg libraries. If you are using Anaconda Python distribution, conda install -c anaconda 'ffmpeg<5' will install the required libraries.

Overview

Streaming API leverages the powerful I/O features of ffmpeg.

It can

• Load audio/video in variety of formats

• Load audio/video from local/remote source

• Load audio/video from file-like object

• Load audio/video from microphone, camera and screen

• Generate synthetic audio/video signals.

• Load audio/video chunk by chunk

• Change the sample rate / frame rate, image size, on-the-fly

• Apply filters and preprocessings

The streaming API works in three steps.

1. Open media source (file, device, synthetic pattern generator)

2. Configure output stream

3. Stream the media

At this moment, the features that the ffmpeg integration provides are limited to the form of

<some media source> -> <optional processing> -> <tensor>

If you have other forms that can be useful to your usecases, (such as integration with torch.Tensor type) please file a feature request.

Preparation

import torch
import torchaudio

print(torch.__version__)
print(torchaudio.__version__)

Out:

1.12.0
0.12.0

try:
    from torchaudio.io import StreamReader
except ModuleNotFoundError:
    try:
        import google.colab

        print(
            """
            To enable running this notebook in Google Colab, install nightly
            torch and torchaudio builds and the requisite third party libraries by
            adding the following code block to the top of the notebook before running it:

            !pip3 uninstall -y torch torchvision torchaudio
            !pip3 install --pre torch torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cpu
            !add-apt-repository -y ppa:savoury1/ffmpeg4
            !apt-get -qq install -y ffmpeg
            """
        )
    except ModuleNotFoundError:
        pass
    raise

import IPython
import matplotlib.pyplot as plt

base_url = "https://download.pytorch.org/torchaudio/tutorial-assets"
AUDIO_URL = f"{base_url}/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav"
VIDEO_URL = f"{base_url}/stream-api/NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4"

Opening the source

There are mainly three different sources that streaming API can handle. Whichever source is used, the remaining processes (configuring the output, applying preprocessing) are same.

1. Common media formats (resource indicator of string type or file-like object)

2. Audio / Video devices

3. Synthetic audio / video sources

The following section covers how to open common media formats. For the other streams, please refer to the Media Stream API - Pt.2.

Note

The coverage of the supported media (such as containers, codecs and protocols) depend on the FFmpeg libraries found in the system.

If StreamReader raises an error opening a source, please check that ffmpeg command can handle it.

Local files

To open a media file, you can simply pass the path of the file to the constructor of StreamReader.

StreamReader(src="audio.wav")

StreamReader(src="audio.mp3")

This works for image file, video file and video streams.

# Still image
StreamReader(src="image.jpeg")

# Video file
StreamReader(src="video.mpeg")

Network protocols

You can directly pass a URL as well.

# Video on remote server
StreamReader(src="https://example.com/video.mp4")

# Playlist format
StreamReader(src="https://example.com/playlist.m3u")

# RTMP
StreamReader(src="rtmp://example.com:1935/live/app")

File-like objects

You can also pass a file-like object. A file-like object must implement read method conforming to io.RawIOBase.read.

If the given file-like object has seek method, StreamReader uses it as well. In this case the seek method is expected to conform to io.IOBase.seek.

# Open as fileobj with seek support
with open("input.mp4", "rb") as src:
    StreamReader(src=src)

In case where third-party libraries implement seek so that it raises an error, you can write a wrapper class to mask the seek method.

class UnseekableWrapper:
    def __init__(self, obj):
        self.obj = obj

    def read(self, n):
        return self.obj.read(n)

import requests

response = requests.get("https://example.com/video.mp4", stream=True)
s = StreamReader(UnseekableWrapper(response.raw))

import boto3

response = boto3.client("s3").get_object(Bucket="my_bucket", Key="key")
s = StreamReader(UnseekableWrapper(response["Body"]))

Note

When using an unseekable file-like object, the source media has to be streamable. For example, a valid MP4-formatted object can have its metadata either at the beginning or at the end of the media data. Those with metadata at the beginning can be opened without method seek, but those with metadata at the end cannot be opened without seek.

Headerless media

If attempting to load headerless raw data, you can use format and option to specify the format of the data.

Say, you converted an audio file into faw format with sox command as follow;

# Headerless, 16-bit signed integer PCM, resampled at 16k Hz.
$ sox original.wav -r 16000 raw.s2

Such audio can be opened like following.

StreamReader(src="raw.s2", format="s16le", option={"sample_rate": "16000"})

Checking the source streams

Once the media is opened, we can inspect the streams and configure the output streams.

You can check the number of source streams with num_src_streams.

Note

The number of streams is NOT the number of channels. Each audio stream can contain an arbitrary number of channels.

To check the metadata of source stream you can use get_src_stream_info() method and provide the index of the source stream.

This method returns SourceStream. If a source stream is audio type, then the return type is SourceAudioStream, which is a subclass of SourceStream, with additional audio-specific attributes. Similarly, if a source stream is video type, then the return type is SourceVideoStream.

For regular audio formats and still image formats, such as WAV and JPEG, the number of souorce streams is 1.

streamer = StreamReader(AUDIO_URL)
print("The number of source streams:", streamer.num_src_streams)
print(streamer.get_src_stream_info(0))

Out:

The number of source streams: 1
StreamReaderSourceAudioStream(media_type='audio', codec='pcm_s16le', codec_long_name='PCM signed 16-bit little-endian', format='s16', bit_rate=256000, num_frames=0, bits_per_sample=0, sample_rate=16000.0, num_channels=1)

Container formats and playlist formats may contain multiple streams of different media type.

src = "https://devstreaming-cdn.apple.com/videos/streaming/examples/img_bipbop_adv_example_fmp4/master.m3u8"
streamer = StreamReader(src)
print("The number of source streams:", streamer.num_src_streams)
for i in range(streamer.num_src_streams):
    print(streamer.get_src_stream_info(i))

Out:

The number of source streams: 27
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=350041, num_frames=0, bits_per_sample=8, width=960, height=540, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=1471810, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=1076143, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=783546, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=535363, num_frames=0, bits_per_sample=8, width=1280, height=720, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=193825, num_frames=0, bits_per_sample=8, width=768, height=432, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=116805, num_frames=0, bits_per_sample=8, width=640, height=360, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=62544, num_frames=0, bits_per_sample=8, width=480, height=270, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=350041, num_frames=0, bits_per_sample=8, width=960, height=540, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=1471810, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=1076143, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=783546, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=535363, num_frames=0, bits_per_sample=8, width=1280, height=720, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=193825, num_frames=0, bits_per_sample=8, width=768, height=432, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=116805, num_frames=0, bits_per_sample=8, width=640, height=360, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=62544, num_frames=0, bits_per_sample=8, width=480, height=270, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=350041, num_frames=0, bits_per_sample=8, width=960, height=540, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=1471810, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=1076143, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=783546, num_frames=0, bits_per_sample=8, width=1920, height=1080, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=535363, num_frames=0, bits_per_sample=8, width=1280, height=720, frame_rate=60.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=193825, num_frames=0, bits_per_sample=8, width=768, height=432, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=116805, num_frames=0, bits_per_sample=8, width=640, height=360, frame_rate=30.0)
StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=62544, num_frames=0, bits_per_sample=8, width=480, height=270, frame_rate=30.0)
StreamReaderSourceAudioStream(media_type='audio', codec='aac', codec_long_name='AAC (Advanced Audio Coding)', format='fltp', bit_rate=83695, num_frames=0, bits_per_sample=0, sample_rate=48000.0, num_channels=2)
StreamReaderSourceAudioStream(media_type='audio', codec='ac3', codec_long_name='ATSC A/52A (AC-3)', format='fltp', bit_rate=384000, num_frames=0, bits_per_sample=0, sample_rate=48000.0, num_channels=6)
StreamReaderSourceAudioStream(media_type='audio', codec='eac3', codec_long_name='ATSC A/52B (AC-3, E-AC-3)', format='fltp', bit_rate=100150, num_frames=0, bits_per_sample=0, sample_rate=48000.0, num_channels=6)

Configuring output streams

The stream API lets you stream data from an arbitrary combination of the input streams. If your application does not need audio or video, you can omit them. Or if you want to apply different preprocessing to the same source stream, you can duplicate the source stream.

Default streams

When there are multiple streams in the source, it is not immediately clear which stream should be used.

FFmpeg implements some heuristics to determine the default stream. The resulting stream index is exposed via

default_audio_stream and default_video_stream.

Configuring output streams

Once you know which source stream you want to use, then you can configure output streams with add_basic_audio_stream() and add_basic_video_stream().

These methods provide a simple way to change the basic property of media to match the application’s requirements.

The arguments common to both methods are;

• frames_per_chunk: How many frames at maximum should be returned at each iteration. For audio, the resulting tensor will be the shape of (frames_per_chunk, num_channels). For video, it will be (frames_per_chunk, num_channels, height, width).

• buffer_chunk_size: The maximum number of chunks to be buffered internally. When the StreamReader buffered this number of chunks and is asked to pull more frames, StreamReader drops the old frames/chunks.

• stream_index: The index of the source stream.

• decoder: If provided, override the decoder. Useful if it fails to detect the codec.

• decoder_option: The option for the decoder.

For audio output stream, you can provide the following additional parameters to change the audio properties.

• format: By default the StreamReader returns tensor of float32 dtype, with sample values ranging [-1, 1]. By providing format argument the resulting dtype and value range is changed.

• sample_rate: When provided, StreamReader resamples the audio on-the-fly.

For video output stream, the following parameters are available.

• format: Image frame format. By default StreamReader returns frames in 8-bit 3 channel, in RGB order.

• frame_rate: Change the frame rate by dropping or duplicating frames. No interpolation is performed.

• width, height: Change the image size.

streamer = StreamReader(...)

# Stream audio from default audio source stream
# 256 frames at a time, keeping the original sampling rate.
streamer.add_basic_audio_stream(
    frames_per_chunk=256,
)

# Stream audio from source stream `i`.
# Resample audio to 8k Hz, stream 256 frames at each
streamer.add_basic_audio_stream(
    frames_per_chunk=256,
    stream_index=i,
    sample_rate=8000,
)

# Stream video from default video source stream.
# 10 frames at a time, at 30 FPS
# RGB color channels.
streamer.add_basic_video_stream(
    frames_per_chunk=10,
    frame_rate=30,
    format="rgb24"
)

# Stream video from source stream `j`,
# 10 frames at a time, at 30 FPS
# BGR color channels with rescaling to 128x128
streamer.add_basic_video_stream(
    frames_per_chunk=10,
    stream_index=j,
    frame_rate=30,
    width=128,
    height=128,
    format="bgr24"
)

You can check the resulting output streams in a similar manner as checking the source streams. num_out_streams reports the number of configured output streams, and get_out_stream_info() fetches the information about the output streams.

for i in range(streamer.num_out_streams):
    print(streamer.get_out_stream_info(i))

If you want to remove an output stream, you can do so with remove_stream() method.

# Removes the first output stream.
streamer.remove_stream(0)

Streaming

To stream media data, the streamer alternates the process of fetching and decoding the source data, and passing the resulting audio / video data to client code.

There are low-level methods that performs these operations. is_buffer_ready(), process_packet() and pop_chunks().

In this tutorial, we will use the high-level API, iterator protocol. It is as simple as a for loop.

streamer = StreamReader(...)
streamer.add_basic_audio_stream(...)
streamer.add_basic_video_stream(...)

for chunks in streamer.stream():
    audio_chunk, video_chunk = chunks
    ...

Example

Let’s take an example video to configure the output streams. We will use the following video.

Source: https://svs.gsfc.nasa.gov/13013 (This video is in public domain)

Credit: NASA’s Goddard Space Flight Center.

NASA’s Media Usage Guidelines: https://www.nasa.gov/multimedia/guidelines/index.html

Opening the source media

Firstly, let’s list the available streams and its properties.

streamer = StreamReader(VIDEO_URL)
for i in range(streamer.num_src_streams):
    print(streamer.get_src_stream_info(i))

Out:

StreamReaderSourceVideoStream(media_type='video', codec='h264', codec_long_name='H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10', format='yuv420p', bit_rate=9958354, num_frames=6175, bits_per_sample=8, width=1920, height=1080, frame_rate=29.97002997002997)
StreamReaderSourceAudioStream(media_type='audio', codec='aac', codec_long_name='AAC (Advanced Audio Coding)', format='fltp', bit_rate=317375, num_frames=9658, bits_per_sample=0, sample_rate=48000.0, num_channels=2)

Now we configure the output stream.

Configuring ouptut streams

# fmt: off
# Audio stream with 8k Hz
streamer.add_basic_audio_stream(
    frames_per_chunk=8000,
    sample_rate=8000,
)

# Audio stream with 16k Hz
streamer.add_basic_audio_stream(
    frames_per_chunk=16000,
    sample_rate=16000,
)

# Video stream with 960x540 at 1 FPS.
streamer.add_basic_video_stream(
    frames_per_chunk=1,
    frame_rate=1,
    width=960,
    height=540,
    format="rgb24",
)

# Video stream with 320x320 (stretched) at 3 FPS, grayscale
streamer.add_basic_video_stream(
    frames_per_chunk=3,
    frame_rate=3,
    width=320,
    height=320,
    format="gray",
)
# fmt: on

Note

When configuring multiple output streams, in order to keep all streams synced, set parameters so that the ratio between frames_per_chunk and sample_rate or frame_rate is consistent across output streams.

Checking the output streams.

for i in range(streamer.num_out_streams):
    print(streamer.get_out_stream_info(i))

Out:

StreamReaderOutputStream(source_index=1, filter_description='aresample=8000,aformat=sample_fmts=fltp')
StreamReaderOutputStream(source_index=1, filter_description='aresample=16000,aformat=sample_fmts=fltp')
StreamReaderOutputStream(source_index=0, filter_description='fps=1,scale=width=960:height=540,format=pix_fmts=rgb24')
StreamReaderOutputStream(source_index=0, filter_description='fps=3,scale=width=320:height=320,format=pix_fmts=gray')

Remove the second audio stream.

streamer.remove_stream(1)
for i in range(streamer.num_out_streams):
    print(streamer.get_out_stream_info(i))

Out:

StreamReaderOutputStream(source_index=1, filter_description='aresample=8000,aformat=sample_fmts=fltp')
StreamReaderOutputStream(source_index=0, filter_description='fps=1,scale=width=960:height=540,format=pix_fmts=rgb24')
StreamReaderOutputStream(source_index=0, filter_description='fps=3,scale=width=320:height=320,format=pix_fmts=gray')

Streaming

Jump to the 10 second point.

streamer.seek(10.0)

Now, let’s finally iterate over the output streams.

n_ite = 3
waveforms, vids1, vids2 = [], [], []
for i, (waveform, vid1, vid2) in enumerate(streamer.stream()):
    waveforms.append(waveform)
    vids1.append(vid1)
    vids2.append(vid2)
    if i + 1 == n_ite:
        break

For audio stream, the chunk Tensor will be the shape of (frames_per_chunk, num_channels), and for video stream, it is (frames_per_chunk, num_color_channels, height, width).

print(waveforms[0].shape)
print(vids1[0].shape)
print(vids2[0].shape)

Out:

torch.Size([8000, 2])
torch.Size([1, 3, 540, 960])
torch.Size([3, 1, 320, 320])

Let’s visualize what we received.

k = 3
fig = plt.figure()
gs = fig.add_gridspec(3, k * n_ite)
for i, waveform in enumerate(waveforms):
    ax = fig.add_subplot(gs[0, k * i : k * (i + 1)])
    ax.specgram(waveform[:, 0], Fs=8000)
    ax.set_yticks([])
    ax.set_xticks([])
    ax.set_title(f"Iteration {i}")
    if i == 0:
        ax.set_ylabel("Stream 0")
for i, vid in enumerate(vids1):
    ax = fig.add_subplot(gs[1, k * i : k * (i + 1)])
    ax.imshow(vid[0].permute(1, 2, 0))  # NCHW->HWC
    ax.set_yticks([])
    ax.set_xticks([])
    if i == 0:
        ax.set_ylabel("Stream 1")
for i, vid in enumerate(vids2):
    for j in range(3):
        ax = fig.add_subplot(gs[2, k * i + j : k * i + j + 1])
        ax.imshow(vid[j].permute(1, 2, 0), cmap="gray")
        ax.set_yticks([])
        ax.set_xticks([])
        if i == 0 and j == 0:
            ax.set_ylabel("Stream 2")
plt.tight_layout()
plt.show(block=False)