AVENC

Video compression and lazy access for meteorological GRIB grids.

Interactive examples & public datasets · GRIBView · Zenodo archive

Install with pip

pip install avenc

AVENC 1.0.0 is available on PyPI and as standalone native installers. Use a virtual environment. Supported binary-wheel targets: Python 3.11–3.14, Linux x86_64/aarch64 (glibc >=2.28), macOS Intel/Apple Silicon (>=15). Install prebuilt FFmpeg 5.1+ separately with brew install ffmpeg-full or sudo apt install ffmpeg; no FFmpeg compilation required.

Check the installation

grib2avenc --help
avenc2grib --help
python -c "import avenc; print(avenc.__version__)"

Convert GRIB to AVENC and back

grib2avenc input.grib output.mp4
avenc2grib output.mp4 reconstructed.grib

The default is video-lossless with finite quantization; it does not promise byte-identical scientific values.

Open using xarray

import xarray as xr

with xr.open_dataset("output.mp4", engine="avenc", cache=False) as ds:
    print(ds)                         # inspect variable/dimension names
    field = ds[next(iter(ds.data_vars))]
    print(field.isel({field.dims[0]: 0}).load())

The wheel installs the native tools automatically and preserves avenc and avenc_xarray. See installation for pipx, Homebrew, standalone archives, source builds, dependencies and troubleshooting, and releases for maintainers.

---

AVENC design and detailed usage: meteorological grids as video streams

AVENC packs GRIB fields into MP4 video streams and embeds the metadata needed to reconstruct scientific grids. It provides command-line encoding, GRIB/Float32/Zarr reconstruction, lazy Python/xarray reads and a browser viewer/extractor.

Repository: https://github.com/filippi/avenc

Build the command-line tools

Native builds target Linux and macOS. Windows users should use a Linux/WSL environment; the native implementation uses POSIX facilities.

Install a C++17 compiler, CMake 3.18+, ecCodes, zlib and FFmpeg/ffprobe. Available encoders and pixel formats depend on the FFmpeg build.

# Debian/Ubuntu prerequisites:
sudo apt install build-essential cmake libeccodes-dev zlib1g-dev ffmpeg

# From this repository:
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel 4
bin/grib2avenc --help

On macOS, install the corresponding dependencies with Homebrew, including cmake, eccodes and ffmpeg. If ecCodes is not found, set ECCODES_HOME or pass -DCMAKE_PREFIX_PATH=/path/to/dependencies. ./build.sh is a convenience wrapper for the same build commands.

Optional installation:

cmake --install build --prefix "$HOME/.local"
export PATH="$HOME/.local/bin:$PATH"

The subset library is installed beside the tools and the HTTP helper under share/avenc/. HTTP CLI input additionally requires python3 on PATH. Builds are portable by default; -DENABLE_NATIVE_ARCH=ON opts into optimization for the build CPU. Keep ENABLE_FAST_MATH=OFF for missing-value correctness.

Encode GRIB fields

bin/grib2avenc input.grib output.mp4
bin/grib2avenc --compact source/*.grib output.compact.mp4
bin/grib2avenc --high source/*.grib output.high.mp4
bin/grib2avenc --list-options

Without an option, AVENC selects Video-lossless. One input with no output name writes <input>.mp4; multiple inputs require an explicit MP4 output path.

Option	Resolved profile	Purpose
default / --lossless	av1_lossless12_auto_400	Lossless 12-bit AV1 video; quantization error remains
--high	hevc_vgood12	Low additional codec error
--medium	vp9_hq10_auto_420	Balanced trade-off
--compact / --low	av1_good10_auto_400	Compression priority
--fast	avc_xcr8	Speed priority
--apple	hevc_xcr8	Software HEVC alias; does not select VideoToolbox

Very small grids may not fit a profile's tile arrangement (the default AV1 profile currently fails on a 64 × 64 synthetic field). Use --fast or a custom encoder configuration with suitable tiling for such inputs.

All shortcuts use automatic quantization-range selection. auto selects the range policy per grouped stream; it can mix frameMinMax and groupMinMax in one MP4. --enc selects a named configuration; --enc_config_table profiles.csv loads custom configurations. See grib2avenc --help and avpack --help for advanced codec, normalization, grouping and metadata options.

The original GRIB bitmap is authoritative for missing cells. Stable shared bitmaps are stored in metadata; the default taper3 fill reduces discontinuities in masked video regions before the original mask is restored. --mask-fill min selects minimum fill instead. Fill is not a change to the scientific valid-cell mask and does not guarantee an error bound on neighboring valid values.

Video losslessness is not scientific losslessness. Mapping physical values to finite integer codes, precision conversion and subsequent GRIB repacking can introduce error even with a lossless video codec. Evaluate each dataset/profile.

Inspect, reconstruct and compare

bin/avenc2grib --list output.mp4
bin/avenc2grib --stats output.mp4
bin/avenc2grib output.mp4 reconstructed.grib
bin/avenc2grib --group stream:0 output.mp4 one_stream.grib
bin/avenc-compare --full input.grib output.mp4 --json errors.json
bin/avenc-compare --fast input.grib output.mp4 --json sampled-errors.json

Use an identically concatenated reference for multi-file inputs: cat source/*.grib > reference.grib. Comparison matches scientific identity, not positional message order. Full and sampled comparison are distinct scopes. Detailed metrics, normalization and output options: avenc-compare.md. The legacy ComputeError utility remains for compatibility.

Subset output is little-endian Float32, ordered (frame, row, column):

bin/avenc2grib --stream-id 0 --frame-start 0 --frame-count 4 \
  --ni-start 20 --nj-start 30 --width 16 --height 12 output.mp4 > subset.f32

Python/xarray

Install the unified package from the repository root; it includes the C++ CLI tools.

python3 -m venv .venv
. .venv/bin/activate
python -m pip install .

import xarray as xr

with xr.open_dataset("output.mp4", engine="avenc", cache=False) as ds:
    print(ds)
    field = next(iter(ds.data_vars))
    # Inspect the variable's dimensions before selecting a point or window.
    first_frame = ds[field].isel({ds[field].dims[0]: 0}).load()

GRIB-packed MP4 files are decoded through avenc2grib. Native Python array packages use a separate pybind11 codec. The Python compress_dataarray API can encode a three-dimensional DataArray with a time dimension; its options, metadata and masking policy are not interchangeable with grib2avenc profiles. It materializes a contiguous Float32 input and is not a streaming Dask encoder. See the Python guide and the runnable examples there.

HTTP access

from avenc import open_avenc

with open_avenc("https://your-server.example/output.mp4") as ds:
    name = next(iter(ds.data_vars))
    first_frame = ds[name].isel({ds[name].dims[0]: 0}).load()
    print(ds.avenc.http_stats)

The CLI also accepts HTTP(S) input URLs. A shared seekable source provides a bounded byte-range cache and transfer statistics. Servers ignoring Range can use a bounded temporary-file fallback. Lazy opening does not prove minimal transfer or spatially selective decoding: preceding GOP frames/full spatial frames may still need decoding. Measure the actual request and bytes transferred.

Experimental Zarr export

bin/avenc2grib --zarr-compress gzip output.mp4 decoded.zarr
bin/avenc2grib --stream-id 0 --frame-start 0 --frame-count 4 \
  --ni-start 20 --nj-start 30 --width 16 --height 12 \
  --zarr-out subset.zarr output.mp4

The exporter writes Zarr v2 arrays per stream, plus quantization/time metadata. The layout is experimental, not a universal training-ready schema. This is an export from reconstructed values, not a new accuracy guarantee.

Browser tools

www/viewer.html and www/extractor.html use WebCodecs and MP4Box. Serve the folder over HTTP on localhost; put your files under www/datasets/ (not tracked).

python3 -m http.server 8000 --directory www
# Open http://localhost:8000/viewer.html?dataset=./datasets/output.mp4

www/index.php is an optional dataset listing and requires PHP rather than the Python server. The metadata inspector uses Range, but the shared viewer/extractor model currently loads the whole MP4. Browser codec support, secure-context requirements, CORS and CDN availability apply. See www/README.md.

Tests

ctest --test-dir build --output-on-failure
python -m pip install '.[test]'
python -m pytest -q python/avenc/tests/test_metadata_v2.py
python -m unittest discover -s tests -p 'test_avenc_compare.py'
python -m unittest discover -s python/avenc/tests -p 'test_http_input.py'
node www/tests/metadata.mjs
node www/tests/decoder.mjs

The comparison and local HTTP tests generate small synthetic fields. External network tests are opt-in. Dataset-dependent xarray/Dask checks require AVENC_TEST_MP4, optionally AVENC_TRUTH_GRIB and AVENC2GRIB_BIN, and skip if inputs are unavailable. AVENC_BIN_DIR selects an out-of-tree native test build.

Research data and licence

The manuscript and large experiments are maintained separately from the software. Dataset identifier: https://doi.org/10.5281/zenodo.23502851 (reserved for the accompanying archive; availability depends on publication).

AVENC is licensed under Apache-2.0. FFmpeg, ecCodes, codecs and browser libraries retain their own licences. Hardware acceleration depends on the encoder, FFmpeg build and device; it is not enabled merely by choosing a profile name. Agent-oriented usage and limitations: skills.md.