TorchCodec

TorchCodec is a Python library for decoding video and audio data into PyTorch
tensors, on CPU and CUDA GPU. It also supports audio encoding, and video
encoding will come soon! It aims to be fast, easy to use, and well integrated
into the PyTorch ecosystem. If you want to use PyTorch to train ML models on
videos and audio, TorchCodec is how you turn these into data.

We achieve these capabilities through:

Pythonic APIs that mirror Python and PyTorch conventions.
Relying on FFmpeg to do the decoding and encoding.
TorchCodec uses the version of FFmpeg you already have installed. FFmpeg is a
mature library with broad coverage available on most systems. It is, however,
not easy to use. TorchCodec abstracts FFmpeg's complexity to ensure it is used
correctly and efficiently.
Returning data as PyTorch tensors, ready to be fed into PyTorch transforms
or used directly to train models.

Using TorchCodec

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

Decoding

from torchcodec.decoders import VideoDecoder

device = "cpu"  # or e.g. "cuda" !
decoder = VideoDecoder("path/to/video.mp4", device=device)

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

# 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]

# Indexing, with PTS and duration info:
decoder.get_frames_at(indices=[2, 100])
# FrameBatch:
#   data (shape): torch.Size([2, 3, 270, 480])
#   pts_seconds: tensor([0.0667, 3.3367], dtype=torch.float64)
#   duration_seconds: tensor([0.0334, 0.0334], dtype=torch.float64)

# Time-based indexing with PTS and duration info
decoder.get_frames_played_at(seconds=[0.5, 10.4])
# FrameBatch:
#   data (shape): torch.Size([2, 3, 270, 480])
#   pts_seconds: tensor([ 0.4671, 10.3770], dtype=torch.float64)
#   duration_seconds: tensor([0.0334, 0.0334], dtype=torch.float64)

Clip sampling


from torchcodec.samplers import clips_at_regular_timestamps

clips_at_regular_timestamps(
    decoder,
    seconds_between_clip_starts=1.5,
    num_frames_per_clip=4,
    seconds_between_frames=0.1
)
# FrameBatch:
#   data (shape): torch.Size([9, 4, 3, 270, 480])
#   pts_seconds: tensor([[ 0.0000,  0.0667,  0.1668,  0.2669],
#         [ 1.4681,  1.5682,  1.6683,  1.7684],
#         [ 2.9696,  3.0697,  3.1698,  3.2699],
#         ... (truncated), dtype=torch.float64)
#   duration_seconds: tensor([[0.0334, 0.0334, 0.0334, 0.0334],
#         [0.0334, 0.0334, 0.0334, 0.0334],
#         [0.0334, 0.0334, 0.0334, 0.0334],
#         ... (truncated), dtype=torch.float64)

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

Installing CPU-only TorchCodec

Install the latest stable version of PyTorch following the
official instructions. For other
versions, refer to the table below for compatibility between versions of
torch and torchcodec.
Install FFmpeg, if it's not already installed. Linux distributions usually
come with FFmpeg pre-installed. TorchCodec supports all major FFmpeg versions
in [4, 7].

If FFmpeg is not already installed, or you need a more recent version, an
easy way to install it is to use conda:
```
conda install ffmpeg
# or
conda install ffmpeg -c conda-forge
```
Install TorchCodec:
```
pip install torchcodec
```

The following table indicates the compatibility between versions of
torchcodec, torch and Python.

`torchcodec`	`torch`	Python
`main` / `nightly`	`main` / `nightly`	`>=3.9`, `<=3.13`
`0.4`	`2.7`	`>=3.9`, `<=3.13`
`0.3`	`2.7`	`>=3.9`, `<=3.13`
`0.2`	`2.6`	`>=3.9`, `<=3.13`
`0.1`	`2.5`	`>=3.9`, `<=3.12`
`0.0.3`	`2.4`	`>=3.8`, `<=3.12`

Installing CUDA-enabled TorchCodec

First, make sure you have a GPU that has NVDEC hardware that can decode the
format you want. Refer to Nvidia's GPU support matrix for more details
here.

Install Pytorch corresponding to your CUDA Toolkit using the
official instructions. You'll
need the libnpp and libnvrtc CUDA libraries, which are usually part of
the CUDA Toolkit.
Install or compile FFmpeg with NVDEC support.
TorchCodec with CUDA should work with FFmpeg versions in [4, 7].

If FFmpeg is not already installed, or you need a more recent version, an
easy way to install it is to use conda:
```
conda install ffmpeg
# or
conda install ffmpeg -c conda-forge
```
If you are building FFmpeg from source you can follow Nvidia's guide to
configuring and installing FFmpeg with NVDEC support
here.

After installing FFmpeg make sure it has NVDEC support when you list the supported
decoders:
```
ffmpeg -decoders | grep -i nvidia
# This should show a line like this:
# V..... h264_cuvid           Nvidia CUVID H264 decoder (codec h264)
```
To check that FFmpeg libraries work with NVDEC correctly you can decode a sample video:
```
ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i test/resources/nasa_13013.mp4 -f null -
```

Install TorchCodec by passing in an --index-url parameter that corresponds
to your CUDA Toolkit version, example:

# This corresponds to CUDA Toolkit version 12.6. It should be the same one
# you used when you installed PyTorch (If you installed PyTorch with pip).
pip install torchcodec --index-url=https://download.pytorch.org/whl/cu126

Note that without passing in the --index-url parameter, pip installs
the CPU-only version of TorchCodec.

Benchmark Results

The following was generated by running our benchmark script on a lightly loaded 22-core machine with an Nvidia A100 with
5 NVDEC decoders.

benchmark_results

The top row is a Mandelbrot video
generated from FFmpeg that has a resolution of 1280x720 at 60 fps and is 120 seconds long.
The bottom row is promotional video from NASA
that has a resolution of 960x540 at 29.7 fps and is 206 seconds long. Both videos were
encoded with libx264 and yuv420p pixel format. All decoders, except for TorchVision, used FFmpeg 6.1.2. TorchVision used FFmpeg 4.2.2.

For TorchCodec, the "approx" label means that it was using approximate mode
for seeking.

Contributing

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

License

TorchCodec is released under the BSD 3 license.

However, TorchCodec may be used with code not written by Meta which may be
distributed under different licenses.

For example, if you build TorchCodec with ENABLE_CUDA=1 or use the CUDA-enabled
release of torchcodec, please review CUDA's license here:
Nvidia licenses.

Dependencies of torchcodec (3).
Dependency	Optional
numpy	true
pillow	true
pytest	true

Package profile

torchcodec

Releases

Release table

PyPI Downloads

Dependencies

Details

Project URLs

Dependencies

Package description