Clip Encoding

Convert extracted clip buffers into compressed media files suitable for storage and training workflows using encoders. NeMo Curator provides both CPU and GPU-based encoders.

Use Cases

• Convert raw clip buffers into a standard format (such as H.264 in MP4) for portability.
• Normalize heterogeneous inputs (encoding formats, bit rates, containers) into a consistent output.
• Reduce storage footprint with controlled quality settings.

Before You Start

If you only need embeddings or analysis and do not require saved media files, you can skip encoding. When writing clips, NeMo Curator produces .mp4 by default.

Quickstart

Use the pipeline stage or the example script flags to encode clips with CPU or GPU encoders.

1
from nemo_curator.pipeline import Pipeline
2
from nemo_curator.stages.video.clipping.clip_extraction_stages import FixedStrideExtractorStage, ClipTranscodingStage
3
4
pipe = Pipeline(name="transcode_example")
5
pipe.add_stage(FixedStrideExtractorStage(clip_len_s=10.0, clip_stride_s=10.0))
6
pipe.add_stage(ClipTranscodingStage(encoder="h264_nvenc", encode_batch_size=16, encoder_threads=1, verbose=True))
7
pipe.run()

Encoder Options

$
bash /opt/Curator/docker/common/install_h264_support.sh --with-libopenh264

Verify ffmpeg/NVENC Support

To use h264_nvenc, confirm that your ffmpeg build includes NVENC support and install the GPU drivers:

$
ffmpeg -hide_banner -encoders | grep nvenc
$
ffmpeg -hide_banner -hwaccels | grep -i nv
$
nvidia-smi

Expected output includes entries like V..... h264_nvenc and cuda in the hardware accelerators list. If not present, install an ffmpeg build with NVENC and ensure NVIDIA drivers and CUDA are available.

Configure

Use ClipTranscodingStage to control encoder choice, batching, and acceleration:

1
from nemo_curator.stages.video.clipping.clip_extraction_stages import ClipTranscodingStage
2
3
transcode = ClipTranscodingStage(
4
    encoder="h264_nvenc",
5
    use_hwaccel=True,             # enable NVENC when using h264_nvenc
6
    encoder_threads=1,            # CPU thread count for CPU encoders
7
    encode_batch_size=16,         # number of clips per encode batch
8
    num_clips_per_chunk=32,       # chunking for downstream writing
9
    use_input_bit_rate=False,     # set True to preserve input bit rate when available
10
    num_cpus_per_worker=6.0,
11
    verbose=True,
12
)

Parameters

Troubleshooting

• “Encoder not found”: Your ffmpeg build may lack the encoder; verify with ffmpeg -encoders.
• “No NVENC capable devices found”: Install NVIDIA drivers/CUDA and ensure the GPU is visible in nvidia-smi.
• Output mismatch or low quality: Revisit encoder defaults; set explicit bit rate/quality settings as needed, or enable use_input_bit_rate.