TorchCodec 0.0.2

A video decoder for PyTorch

Installation | Simple Example | Detailed Example | Documentation | Contributing | License

TorchCodec

TorchCodec is a Python package that provides easy-to-use and fast APIs to decode video frames to PyTorch Tensors. These tensors can then be transformed and fed to ML models. TorchCodec needs a working FFMPEG installation (that the user provides) and uses it as a library to do the decoding. We use FFMPEG because of the diverse video formats it supports as well as its popularity.

Under the hood, TorchCodec uses Pytorch's C++ custom ops to wrap C++ code that does the memory management of packets, frames and codecs. TorchCodec hides that complexity under its API (see example usage below).

[!NOTE] ������ TorchCodec is still in early development stage and some APIs may be updated in future versions without a deprecation cycle, depending on user feedback. If you have any suggestions or issues, please let us know by opening an issue!

Using TorchCodec

Here's a condensed summary of what you can do with TorchCodec. For a more detailed example, check out our documentation!

from torchcodec.decoders import SimpleVideoDecoder

decoder = SimpleVideoDecoder("path/to/video.mp4")

decoder.metadata
# VideoStreamMetadata:
#   num_frames: 250
#   duration_seconds: 10.0
#   bit_rate: 31315.0
#   codec: h264
#   average_fps: 25.0
#   ... (truncated output)

len(decoder)  # == decoder.metadata.num_frames!
# 250
decoder.metadata.average_fps  # Note: instantaneous fps can be higher or lower
# 25.0

# Simple Indexing API
decoder[0]  # uint8 tensor of shape [C, H, W]
decoder[0 : -1 : 20]  # uint8 stacked tensor of shape [N, C, H, W]


# Iterate over frames:
for frame in decoder:
    pass

# Indexing, with PTS and duration info
decoder.get_frame_at(len(decoder) - 1)
# Frame:
#   data (shape): torch.Size([3, 400, 640])
#   pts_seconds: 9.960000038146973
#   duration_seconds: 0.03999999910593033

decoder.get_frames_at(start=10, stop=30, step=5)
# FrameBatch:
#   data (shape): torch.Size([4, 3, 400, 640])
#   pts_seconds: tensor([0.4000, 0.6000, 0.8000, 1.0000])
#   duration_seconds: tensor([0.0400, 0.0400, 0.0400, 0.0400])

# Time-based indexing with PTS and duration info
decoder.get_frame_displayed_at(pts_seconds=2)
# Frame:
#   data (shape): torch.Size([3, 400, 640])
#   pts_seconds: 2.0
#   duration_seconds: 0.03999999910593033

You can use the following snippet to generate a video with FFmpeg and tryout TorchCodec:

fontfile=/usr/share/fonts/dejavu-sans-mono-fonts/DejaVuSansMono-Bold.ttf
output_video_file=/tmp/output_video.mp4

ffmpeg -f lavfi -i \
    color=size=640x400:duration=10:rate=25:color=blue \
    -vf "drawtext=fontfile=${fontfile}:fontsize=30:fontcolor=white:x=(w-text_w)/2:y=(h-text_h)/2:text='Frame %{frame_num}'" \
    ${output_video_file}

Installing TorchCodec

Note: if you're on MacOS, you'll need to build from source. Instructions below assume you're on Linux.

First install the latest stable version of PyTorch following the official instructions.

Then:

pip install torchcodec

You will also need FFmpeg installed on your system, and TorchCodec decoding capabilities are determined by your underlying FFmpeg installation. There are different options to install FFmpeg e.g.:

conda install ffmpeg
# or
conda install ffmpeg -c conda-forge

Your Linux distribution probably comes with FFmpeg pre-installed as well. TorchCodec supports all major FFmpeg version in [4, 7].

Planned future work

We are actively working on the following features:

• Ship wheels for MacOS, so that MacOS users can pip install torchcodec. For now this is only supported on Linux, but MacOS users can build from source.
• GPU decoding
• Audio decoding

Let us know if you have any feature requests by opening an issue!

Contributing

We welcome contributions to TorchCodec! Please see our contributing guide for more details.

License

TorchCodec is released under the BSD 3 license.

If you are building with ENABLE_CUDA and/or ENABLE_NVTX please review Nvidia licenses.