• TorchCodec API without PyTorch. CPU video/audio decoder interfaces follow TorchCodec and return NumPy arrays.
• Validated playback semantics. Frame selection, ordering, timestamps and audio ranges are checked against TorchCodec 0.17.0 and independently generated media.
• Efficient batch decoding. Rust/PyO3 bindings to FFmpeg process frame batches in a single native call, avoiding per-frame Python calls. Closing a decoder releases its FFmpeg resources without waiting for Python's cyclic GC.
• Lightweight installation. Linux wheels are 10.3–10.5 MiB (v0.1.4), including FFmpeg shared libraries. NumPy is the only Python dependency.

Quick start

uv pip install tensorcodec

Use an existing virtual environment, or create one with uv venv first. No separate FFmpeg installation is needed for the published Linux wheels.

from tensorcodec.decoders import VideoDecoder, AudioDecoder

with VideoDecoder("video.mp4") as video:
    frame = video[0]                                  # RGB array: (C, H, W)
    batch = video.get_frames_at([4, 0, 4])              # requested order, including duplicates
    clip = video.get_frames_played_in_range(0, 1, fps=8)

with AudioDecoder("audio.wav", sample_rate=16000, num_channels=1) as audio:
    samples = audio.get_samples_played_in_range(0, 1)
    waveform = samples.data                           # float32: (channels, samples)

Arrays keep their storage after the decoder closes. Paths, URLs, encoded bytes, 1-D uint8 arrays and seekable file objects are supported.

For time-based windows without an initial full packet scan (since v0.1.4):

with VideoDecoder("video.mkv", seek_mode="timestamp") as decoder:
    frames = decoder.get_frames_played_at([10.0, 10.1, 10.2])

Decoder transforms (since v0.1.5) resize and crop inside the decoder, as TorchCodec's do:

from tensorcodec.transforms import CenterCrop, Resize

with VideoDecoder("video.mp4", transforms=[Resize((256, 340)), CenterCrop((224, 224))]) as decoder:
    frames = decoder.get_frames_at([0, 10])  # (2, 3, 224, 224)

This TensorCodec extension selects by actual PTS and retries seeks that overshoot. It supports time queries, including ranges with explicit fps, but not frame indices, len(decoder), or get_all_frames(). See the contract.

Features

TensorCodec 0.1.5 relative to TorchCodec 0.17.0. ✓ supported · △ partial support · — not implemented.

FPS-based frame queries are supported; clip samplers are a separate API.

Decoder compatibility

For high-bit-depth video, use VideoDecoder(path, output_dtype="auto") to select float32 above 8 bits, or output_dtype="uint16" for full-range 16-bit RGB. HDR output retains PQ/HLG encoding without SDR tone mapping. Rotation is applied automatically, and metadata dimensions match the output.

For unmodified samples, use VideoDecoder(path, output_format="native"). Supported formats: gray, gray12le, gray16le/be, rgb24, rgba. Native output preserves channel count, integer values and pixel coordinates; expected_pixel_format optionally asserts the source format.

Package size

Linux CPU wheels, Python 3.12. Download / unpacked size in MiB.

TensorCodec and PyAV bundle FFmpeg; TorchCodec needs it separately. Other dependencies are excluded. Measurements.

Scope and compatibility

The supported CPU API is checked for frame selection, ordering, timestamps, durations, stream selection and metadata, both against TorchCodec 0.17.0 and independently generated media.

• Pixel comparisons allow color-conversion rounding of at most 1 uint8 unit or 1/65535 for float32 in the tested cases.
• Empty index lists are supported, including the case affected by the reference's empty-list dtype inference bug.
• NumPy output preserves the decoder API structure; callers expecting torch.Tensor must adapt their array handling.

See the compatibility contract and playback rules for the tested behavior.

Current limits

• Wheels: Linux x86_64 and ARM64 (aarch64), glibc 2.17+, CPython 3.10+. NumPy must also provide a compatible wheel; newer Python versions may require a newer glibc. macOS 14+ wheels support ARM64 and x86_64. Windows, musl/Alpine and free-threaded Python wheels are not provided.
• Exact seeking: scans packet timestamps when opening the decoder. Incorrect container keyframe flags can produce corrupt frames; repaired input or corrected frame mappings are needed in that case.
• Audio ranges: decode from the beginning, so late ranges can be expensive.
• Video conversion: no HDR-to-SDR tone mapping or native YUV-plane output. Reflected and non-right-angle display matrices are unsupported.

See container behavior for seek limitations and benchmark tools for workload measurements.

Development and verification

uv sync --group dev --group oracle

uv run --group oracle pytest tests/test_video_contract.py tests/test_audio_contract.py --backend torchcodec
uv run --group oracle pytest --compare

# Rebuild after changing Rust code.
uv run --group oracle maturin develop --locked --uv

TensorCodec's own code is licensed under Apache-2.0. Versions through v0.1.4 were released under MIT.