Video Data Loading#

Load video data for curation using NeMo Curator.

How it Works#

NeMo Curator loads videos with a composite stage that discovers files and extracts metadata:

VideoReader is a composite stage that is broken down into a

  1. Partitioning (list files) stage

  • Local paths use FilePartitioningStage to list files

  • Remote URLs (for example, s3://, gcs://)

    • use ClientPartitioningStage backed by fsspec.

    • Optional input_list_json_path allows explicit file lists under a root prefix.

  1. Reader stage (VideoReaderStage)

  • This stage downloads the bytes (local or via FSPath) for each listed file

  • Calls video.populate_metadata() to extract resolution, fps, duration, encoding format, and other fields.

You can set

  • video_limit to limit the number of files to be processed; use None for unlimited.

  • verbose=True to log detailed per-video information.

Local and Cloud#

Use VideoReader to load videos from local paths or remote URLs.

Example#

from nemo_curator.pipeline import Pipeline
from nemo_curator.stages.video.io.video_reader import VideoReader

pipe = Pipeline(name="video_read", description="Read videos and extract metadata")
pipe.add_stage(VideoReader(input_video_path="s3://my-bucket/videos/", video_limit=None, verbose=True))
pipe.run()

Explicit File List (JSON)#

For remote datasets, ClientPartitioningStage can use an explicit file list JSON. Each entry must be an absolute path under the specified root.

JSON Format#

[
  "s3://my-bucket/datasets/videos/video1.mp4",
  "s3://my-bucket/datasets/videos/video2.mkv",
  "s3://my-bucket/datasets/more_videos/video3.webm"
]

If any entry is outside the root, the stage raises an error.

Example#

from nemo_curator.pipeline import Pipeline
from nemo_curator.stages.client_partitioning import ClientPartitioningStage
from nemo_curator.stages.video.io.video_reader import VideoReaderStage

ROOT = "s3://my-bucket/datasets/"
JSON_LIST = "s3://my-bucket/lists/videos.json"

pipe = Pipeline(name="video_read_json_list", description="Read specific videos via JSON list")
pipe.add_stage(
    ClientPartitioningStage(
        file_paths=ROOT,
        input_list_json_path=JSON_LIST,
        files_per_partition=1,
        file_extensions=[".mp4", ".mov", ".avi", ".mkv", ".webm"],
    )
)
pipe.add_stage(VideoReaderStage(verbose=True))
pipe.run()

Supported File Types#

The loader filters these video extensions by default:

  • .mp4

  • .mov

  • .avi

  • .mkv

  • .webm

Metadata on Load#

After a successful read, the loader populates the following metadata fields for each video:

  • size (bytes)

  • width, height

  • framerate

  • num_frames

  • duration (seconds)

  • video_codec, pixel_format, audio_codec

  • bit_rate_k

Note

With verbose=True, the loader logs size, resolution, fps, duration, weight, and bit rate for each processed video.